2026-09-23

Modern information retrieval increasingly employs both embedding and generative models to handle complex queries. However, current serving systems suffer from low throughput and poor GPU utilization because they execute these models in isolation. Coarse-grained partitioning, such as dedicating GPUs to specific tasks, fails to adapt to dynamic workloads and creates computational "bubbles". To address these, we present Orthrus, a serving system that performs heterogeneous batching within a unified inference loop. The primary challenge lies in unifying embedding and generation workloads with conflicting computational patterns while optimizing batch composition for high performance. Orthrus addresses these challenges through chunked embedding with incremental pooling and by adjusting batch composition in a workload-aware manner. Evaluation on four A100 GPUs shows that, relative to baseline deployments, Orthrus achieves 1.28$\times$--4.52$\times$ higher throughput on controlled workloads and up to 55.8% lower end-to-end p99 latency on an iterative-RAG benchmark. We release our code at https://github.com/illinoisdata/Orthrus .

Retrieval-augmented generation (RAG) gives large language models (LLMs) access to external knowledge, but its conventional retrieve-concatenate-generate pipeline makes retrieval decisions on behalf of the model. As tool use and agent loops become more reliable, an agent can decide whether to retrieve, what to inspect, and when to stop. This shift exposes a new bottleneck: the agent may not know what a knowledge base contains. Traditional knowledge bases expose documents as anonymous text chunks with limited information about scope, purpose, provenance, or relations. We propose Knowledge-as-Skill, an organization scheme that makes a knowledge base discoverable, navigable, and self-descriptive. It has three layers: a discovery layer centered on SKILL.md; a navigation layer with one index.md per directory; and a knowledge layer containing documents with YAML frontmatter for topic, type, provenance, and lifecycle. The design follows the Open Knowledge Format (OKF) and the Skill protocol without modifying the agent framework. We also provide knowledge-as-skill, a pipeline for converting heterogeneous collections of PDFs, Word files, web exports, and notes into this structure. In a preliminary evaluation on the WixQA enterprise customer-support benchmark, our setup obtains 0.889 Factuality and 0.816 Context Recall, compared with reported Corpus2Skill values of 0.767 and 0.708. It obtains slightly lower Faithfulness, lower Context Precision, and more interaction turns. Because the models, prompts, and knowledge-package construction differ, these results are directional cross-work evidence rather than a controlled comparison.

Online A/B experiments are the decision standard for user engagement, but traffic and readout time limit how many conversational-AI changes can be tested. We ask whether an offline signal designed to be computable without treatment-arm user exposure agrees with the outcomes of those experiments. We contribute a reusable construction and diagnosis checklist that treats an offline proxy as a chain of three alignments: behavioral label to product outcome, learned classifier to candidate-assistant behavior, and aggregated offline signal to experiment effect. A companion evaluation protocol audits the whole composite by interval-aware decision agreement, which compares offline and online confidence intervals instead of point estimates, and by within-experiment ranking. The instantiation we evaluate comprises a fixed evaluation suite on which candidate behavior is scored, an engagement classifier trained to predict session/prompt level engagements, and a calibration layer mapping sample-level score differences to online model-level engagement deltas. We then report the audit: 489 paired offline-online contrasts (one candidate arm against its control) from 27 experiments on a deployed multi-turn assistant, spanning model checkpoints to system-prompt tuning. Our primary test uses the 113 contrasts from eight experiments that ran after the map was frozen: on these the composite reaches 81.1% F1, against 34.3% for the raw classifier score it is built on, and makes no wrong-direction calls where that raw score makes 31. Every offline prediction was computed before its experiment ran to prevent overfitting. The evidence supports using the composite to prioritize candidates before scarce experiment traffic is allocated---in our deployment of the experiment, selecting among training checkpoints and tuning system prompts.

Language-model agents often reach a working solution and then fail to consistently deliver it. We study runtime policies: targeted natural-language instructions and action denials applied by the agent harness at states that preceded observed failures, without changing model weights or the user prompt. With this, keeping capability constant, we observe a meaningful unlock in delivered reliability. Across the complete 87-task Terminal-Bench 2.1 suite, with two attempts per task, policies increase repeated success (pass^2) in all three GPT-5.6 tiers: 50.6% to 54.0% for Luna, 55.2% to 60.9% for Terra, and 64.4% to 73.6% for Sol. Sol's best-of-two success changes by 1.2 points while repeated success rises by 9.2, showing that policies chiefly convert reachable solutions into dependable delivery. We further cover 14 tasks under Terra's frozen portfolio. Policy-guided Terra reaches 71.4%, compared with 64.3% for unassisted Sol, at about half the cost, demonstrating how engineering around models could unlock dependability for a use case. To isolate the mechanism we run a randomized five-arm experiment: real policies reach 61% on eligible tasks, versus 39% without a policy, 36% with a timing-matched sham, and 39 to 43% with generic verification or reconsideration. The intended corrective behavior appears in 22 of 24 coded policy attempts, against at most 14 in any other arm. Runtime policies are therefore a practical reliability layer: they make capabilities an agent already possesses substantially more repeatable.

As agent capabilities have grown, practical limitations increasingly stem from constrained context windows rather than model capacity. Common strategies, such as truncation, heuristic aging, and lossy summarization, may discard useful information or introduce hallucination risk. To address these challenges, we propose Dynamic Tool Output Compression (DTOC), a framework for scalable context management in LLM-based agents that models context updates as explicit and reversible operations within the agent reasoning loop. DTOC retains full tool outputs in external memory while inserting compact placeholders into the active context, enabling selective reconstruction when needed. We formalize the DTOC mechanism, integrate it into a ReAct-style agent architecture, and provide a production-oriented implementation supporting on-demand restoration of compressed outputs. Experiments on DeepSWE reveal model-dependent effects: for responsive models (Sonnet 4.6, GPT-5.4), DTOC reduces input tokens (10.3 and 12.7%) and agent steps (2.4 and 32.3%), while increasing solve rates (2.5 and 1.5 times higher) and lowering cost per solved task (3 and 3.5 times lower cost per solved task). For the other models results are more mixed, with GPT-5.5 doubling solve rate and halving cost, but no impact on solve rate and negative impact on cost for the other models. Ablation results show reversibility is critical: disable-only compression variants degraded performance, while full DTOC recovered baseline accuracy at substantially lower context cost. These findings indicate that explicit, reversible context management can improve the efficiency of long-horizon agent reasoning without degrading task performance.

Learned world models are becoming essential to general-purpose agents: by predicting action consequences, they support planning and decision-making while reducing reliance on costly trial and error. This reliance creates a fundamental ambiguity: when a world-model-guided decision fails, the trajectory alone may not reveal whether the agent's decision rule or the world model caused the loss. We formalize this failure-attribution problem as a counterfactual decomposition of return loss and prove that its components are not identifiable from passive interaction, even for finite-horizon planners. This obstruction motivates Dual-Frontier, a learning principle that admits a world-model-guided decision only when its predicted advantage exceeds a certified bound on decision-relevant world-model error; otherwise, evidence is allocated to world-model verification. Action-conditioned value bounds and a closed-loop extension guarantee non-decreasing return for admitted decisions. Calibrated gates and simultaneous confidence sequences support adaptive evidence reuse, with sufficient and necessary verification bounds. Controlled learned-model experiments validate the predicted failure modes and certification behavior, while cross-backbone tool-use benchmarks instantiate the same verify-then-promote rule in realistic agent world-model pipelines, consistently improving decision quality and reliability.

Large language models can process increasingly long prompts, yet their ability to locate and use decisive evidence may degrade as irrelevant or confusable context is added. We formulate this phenomenon, which we call context poisoning, as extreme-value interference in attention: the decisive-evidence score is upper-bounded, while the maximum score among effective distractors grows with their number. Under a softmax retrieval abstraction, we derive a finite-sample upper bound showing that maintaining a fixed accuracy target above base rate requires the evidence margin to scale as $\Omega(\sqrt{\log N})$, where N denotes the effective distractor count rather than necessarily the raw context length. The analysis connects long-context degradation to score aliasing, positional aliasing, and softmax dilution. Controlled experiments show that retrieval accuracy decreases as total context grows in the presence of embedded hard negatives, that the same-format condition produces the largest observed accuracy drop among the tested distractor constructions at fixed context length, and that retrieval gating can improve evidence use while its net benefit depends on preserving evidence recall. These results motivate evidence bottlenecks, alias-resistant representations, retrieve-then-reason architectures, verifier-mediated memory, and contrastive anti-poison training.

Organizations fine-tune small language models on private data and then compress them to 4 bits for resource-efficient deployment. We show that the compression method also affects privacy. What separates the methods is not the bit width but whether they tune their rounding on a small sample of text, the calibration corpus. On our primary model, when each planted record's own opening text is used as the prompt, the two calibration-based methods we test, Activation-aware Weight Quantization (AWQ) and Gradient-based Post-Training Quantization (GPTQ), each reproduce none of the planted records, while the calibration-corpus-free GGUF Q4_K_M format reproduces 5.3% of them. Tracked across five open models with 0.5-7 billion parameters, AWQ leaks least at every size and in both families, with little accuracy loss at 3-7 billion. Controlled experiments associate the difference with calibration-induced rounding error in channels involved in rare-token prediction. Choosing the 4-bit method is therefore a deployment-time privacy decision, not only a question of speed and quality.

Personalised search must satisfy query intent while incorporating user context and historical interactions. LLM-based cross-encoders provide a single reranking interface, but injecting predictive behavioural statistics into their prompts can encourage shortcut learning: reliance on historical signals at the expense of semantic and user-context patterns that generalise to sparse or unseen searches. We study this problem in the personalised search system of a large-scale audio streaming platform using Query Slice Stats (QSS), an interaction-derived behavioural feature summarising historical success for query-candidate pairs. Naive QSS injection improves ranking when the feature is available but reduces robustness when it is removed. We address this with deterministic dual-sample feature-dropout training, which presents each example once with QSS included and once with QSS removed. Offline, QSS injection improves ranking quality by 13.3% when available. Dual-sample training preserves these gains while improving performance under QSS-removed evaluation by 4.0% relative to naive QSS training. In a live online test, both QSS-aware variants improve search success by roughly 2%. The aggregate test does not distinguish dual-sample from features-only training; the cold-start comparison is directionally consistent with the offline results. Paired feature-present and feature-removed training can therefore reduce the tension between exploiting strong behavioural statistics and remaining robust when they are unavailable.

LLM agents increasingly work on long-horizon tasks, and the decisions they make along the way, such as which hypothesis to test or which implementation to build on, determine the outcome of the whole run. Making these decisions well is becoming a key capability for both engineering and research agents. We refer to the ability to make good long-horizon decisions as the taste of an agent. While existing benchmarks measure the end-to-end success of agents on long-horizon tasks, none of them measures the taste of an agent. To address this problem, we build Taste-Bench, a benchmark of taste questions constructed automatically from trajectories that agents produced in engineering and research tasks. Each question presents a decision fork, a point in a trajectory where multiple directions are available and one of them leads to a better outcome, and the evaluated model chooses among these directions without seeing what happens after the fork. We mine these forks automatically from parallel attempts at the same task and from detours inside a single trajectory, without needing human annotation. We evaluate frontier models on Taste-Bench and find that the best model answers only 59.7% of the questions correctly. We further find that forks whose deciding evidence appears later in the trajectory are much harder for every model, and that a larger reasoning budget does not improve the accuracy. Finally, we show that taste can be trained. We distill the judgment of a teacher that has seen the outcome into a student model, and the student makes better decisions on unseen tasks and improves end-to-end success on held-out SWE-bench Pro tasks.

Agents using the Model Context Protocol (MCP) rely on semantic matching to select tools from third-party servers, exposing a semantic supply-chain risk through attacker-controlled metadata and outputs. We introduce A2M (Attraction-to-Manipulation), a two-stage black-box framework for hijacking MCP agents. The Attraction phase optimizes tool metadata to increase invocation probability; the Manipulation phase uses execution traces to refine adversarial tool returns that steer agents toward attacker-desired outcomes. On LiveMCPBench, direct attacks optimized and evaluated on GLM-4.6 achieve a macro-average malicious tool invocation rate of 93.6% across four scenarios, increase weighted token costs to 32.4$\times$ the benign baseline under Cognitive Denial of Service, and attain a mean attack success rate of 74.4% across Information Exfiltration, Environment Integrity Compromise, and Reasoning Derailment. Transfer to four other models without re-optimization yields corresponding macro-averages of 63.6%, 2.7$\times$, and 24.5%. These findings motivate stronger tool vetting and runtime isolation in MCP ecosystems. Code is publicly available at https://github.com/Lilaizhen/A2M.

Large language models are increasingly deployed as course-specific tutors, but their usefulness depends on grounding in vetted instructional materials that are often revised mid-semester. Our prior work built a multimodal retrieval-augmented generation (RAG) system over an authentic machine learning course corpus (Foundations of Machine Learning) and found that retrieval improved contextual grounding, but that fixed retrieval strategies were suboptimal. That motivates a different question: whether how a corpus is structured at ingest time matters more than how much is retrieved at query time. We present a controlled head-to-head comparison of two knowledge representations over an identical classroom corpus: (A) vector RAG, replicating the best-performing configuration from our prior study, and (B) an LLM-compiled wiki (Karpathy framework), in which the corpus is synthesized at ingest into linked concept pages with explicit cross-references and citations back to source materials. We evaluate 59 questions spanning single-fact recall, cross-unit concept linking, synthesis and explanation, and currency after a syllabus revision, scored by an LLM judge against a human-authored rubric. Both representations answered single-fact questions about equally well (9.33 vs. 9.96 of 10), but diverged sharply on questions requiring links across course units. The compiled wiki remained accurate and grounded (9.93; 100% grounded in cited sources), while retrieval scored lower and was markedly less grounded (8.14; 64%). The wiki's citations let students and instructors trace any claim back to the lecture that introduced it, adding a layer of dynamic retrieval that machine learning courses require. While further testing is needed, instructors using AI to support learning in ML courses should consider wiki-based structure for its potential to support foundational elements of best practice.

Multi-source knowledge graphs (KGs) need query mechanisms that expose reliability and exploit domain structure. This paper presents TCMaster, a property-graph query substrate for confidence-aware traversal and workload-guided physical design over Traditional Chinese Medicine KGs. TCMaster integrates pharmacopoeias, prescriptions, molecular databases, and LLM-extracted micro-semantics into a KG with approximately 221K entities and 723K base edges. It annotates edges with provenance-level confidence, rewrites Cypher queries with confidence predicates, ranks multi-hop paths under PRODUCT, MIN, or weighted-average policies, and uses ontology skew through direction selection, herb-attribute bitmaps, and materialized shortcut edges. On Neo4j, direction selection improves attribute lookup by a factor of 1.47, shortcuts accelerate high-fanout target counting by a factor of 4.42, confidence filtering removes 39.3 percent of low-quality heterogeneous paths, and KG retrieval improves TCMbench QA accuracy by 20.0 percentage points.

Large Language Models (LLMs) are progressively used for advanced engineering tasks, includes Computer-Aided Design (CAD) documentation, standards compliance verification, and knowledge retrieval. Still, they are prone to produce hallucinations, outputs that seem convincing but aren't based on context that limit their trustworthiness in high-end engineering applications where precision and compliance are crucial. The paper introduces CoVe-RAG+, a unified framework that integrates Chain-of-Verification (CoVe) with Retrieval-Augmented Generation (RAG) to mitigate hallucinations in the results generated by large language models (LLMs). CoVe-RAG+ supports LLM verification in external sources of authority, such as engineering standards, CAD information, and simulation reports, while applying an iterative self-verification process to validate important claims. CoVe-RAG+ is assessed on engineering activities such as CAD model documentation, standards compliance verification, and the reutilization of historical design data. Experimental findings indicate a 28% improvement in factual accuracy relative to baseline CoVe and RAG methodologies. Moreover, CoVe-RAG+ strengthens user confidence by providing elucidative verification reports and source traceability. The findings indicate that CoVe-RAG+ provides a scalable and reliable option for implementing LLMs in engineering design processes where factual accuracy is critical.

Post-training is becoming a service (PTaaS): a customer hands an operator data and a goal, and a forward-deployed engineer (FDE) returns a fine-tuned, evaluated, and deployed model under a budget, a human-approval gate, and reproducibility requirements. Seating an LLM agent in the FDE seat raises a question existing benchmarks cannot answer: not whether an agent can raise a metric, but whether it can be trusted to deliver. We answer it on a governed delivery plane, where an agent drives ten stages and an oracle scores each stage from platform-recorded facts. The central silent failure is the run that trains but does not learn (TBDL): loss falls, every signal stays green, and the delivered model is no better than the base. An operator-run acceptance gate catches every such run before payment, and a detector calibrated on known-corrupted runs flags severe corruption mid-run. We ran four frontier agents (Claude Opus 5, GPT-5.6-luna, Gemini 3.7 Flash, DeepSeek V4-Pro) end to end on metered L40S, A100, and H200 GPUs across 8B to 70B open bases, certifying every scenario before scoring. We also ran a human FDE arm under the same oracle and compare every agent against it.

Real-time voice assistants must reason over evolving requests, execute actions, and follow conversational rules. Qwen-Audio-3.1-Realtime brings these requirements together through Think, Act, and Speak and Coordinate. Think combines Core-Cocktail supervised fine-tuning with Multimodality and Multi-Teacher On-Policy Distillation (M$^{2}$-OPD) to transfer language capabilities and develop native audio skills. Act uses self-evolving executable environments and multi-granularity rollouts for Group Relative Policy Optimization (GRPO), teaching the model to use tools, interpret feedback, and complete tasks. Speak and Coordinate aligns how, when, and whether the assistant speaks or acts. We evaluate audio reasoning, multilingual understanding, tool use, conversational behavior, full-duplex interaction, and safety. Compared with Qwen-Audio-3.0-Realtime, 3.1 raises overall task success from 78.4% to 82.0% on our half-duplex speech-to-text adaptation of $\tau$-Voice. On speech-to-speech Full-Duplex-Bench v1.5, the response rate to background speech falls from 73.0% to 13.0%. We also present a separate Voice Harness prototype, using Qwen-Audio-3.0-Realtime as its foreground, that extends spoken interaction to persistent tasks through foreground--background coordination and memory.

Retrieval over a personal memory store is retrospective: it surfaces what resembles the query, and it is blind to what the user has committed to do. We describe a prospective term for memory retrieval that costs no inference at query time. Commitments are held in an explicit ledger as dated or trigger-conditioned entries; memory items linked to a firing entry receive a salience boost, blended multiplicatively into embedding-based retrieval so that relevance remains sovereign. On a synthetic prospective-memory task set modeled on TriggerBench's published structure (48 blind-authored dialogues, 175 tasks), the term raised recall@5 on the hard stratum from 0.000 to 0.955 at the default blend weight and to 1.000 under a floor variant, with zero false boosts across 53 resolved-commitment tasks. Blind authorship also produced a scope finding: only 17-29% of naturally phrased commitment-trigger pairs defeat embedding similarity, so the term matters on a real minority of cases and must do no harm on the rest, which it does not. We position precomputed commitment linkage as the always-on floor of a layered design whose expansion layer is query-time prospection. Results are preliminary: the evaluation set is author-constructed, and evaluation on TriggerBench proper is committed follow-up work once its data is released.

LLM agents often use generative models for bounded decisions, raising the question of when these decisions can be handled more efficiently without reducing task success. We study REFLEX, an agent architecture that uses Jev as a fast, typed decision layer and calls a strong LLM when confidence is low, or generation is required. On a frozen 100-task benchmark, REFLEX achieves 95% success with 72.7% fewer strong-model calls than a strong-only agent, with reductions persisting across three fallback families. Controlled interventions show that reliability depends on action-set size and near-valid alternatives near authorization boundaries. External BFCL and $\tau$-style evaluations reveal limited advantages over a cheap generative cascade when ordinary routing is already highly accurate. These findings identify when selective control with Jev can reduce computation and where its benefits are limited.

Peer review plays a central role in scholarly publishing, yet verifying whether reviewer claims are supported by manuscript evidence remains a largely manual and time-consuming process. We present Peerify, a pipeline for manuscript-grounded verification of peer-review claims. Given a manuscript and a review comment, the Peerify pipeline decomposes reviews into atomic claims, retrieves relevant manuscript evidence, and determines whether each claim is supported by the paper. To support the development and evaluation of the pipeline, we construct a benchmark of 800 claims derived from authentic peer-review interactions collected from NeurIPS 2024 and ICLR 2024, including a 300-claim hand-labeled subset used to audit the automated supervision. We evaluate state-of-the-art language models and retrieval strategies within the Peerify pipeline, together with entailment baselines. Our results demonstrate the importance of retrieval-centered verification and claim decomposition, while highlighting the challenges posed by ambiguous and interpretive reviewer claims. Automated labels agree with human consensus on 90.3% of audited claims ($\kappa = 0.87$), while off-the-shelf entailment models stay below 0.24 macro-F1.

Simulation-based tutor selection can be unstable when predictively adequate learner models imply different policy rankings. RankCert certifies one of eight equal-budget tutoring policies only when model-averaged utility, probability-best, posterior regret, cross-domain rank, family coverage, and leave-one-domain-out and leave-one-visible-family-out averages support the same candidate; otherwise it abstains. We evaluated RankCert in 1,280 frozen held-out settings spanning five rotating held-out oracle families, 64 scenarios per family, and four cohort sizes. Calibration used a licensed, de-identified EdNet-KT1 derivative with 5,000 learners and 590,056 retained responses; all five family representatives passed the frozen adequacy gate. Minimum-domain mean pairwise top-1 agreement was 0.272917 (95% CI [0.253646, 0.293229]), showing substantial structural disagreement. Cohort-noise variance decreased from n = 30 to n = 300, while the structural family share remained nonzero. RankCert reduced total held-out decision loss relative to full-coverage point selection by 0.006605 normalized-outcome units (95% CI [0.004859, 0.008407]). At comparable coverage, however, it did not reduce selective risk relative to a confidence-gated point certificate (difference -0.000213; 95% CI [-0.003238, 0.002384]; Holm p = 0.929654). Certification occurred in 3.75% of settings and only in stable scenarios; RankCert abstained in every ambiguous, misspecified, and structural-conflict setting. "Safe" denotes only benchmark-scoped decision certification under the declared utility and uncertainty set; no human-learning, causal, deployment-effectiveness, or general-safety claim is made.

Climate campaigns are often evaluated through attention and mobilization, but less is known about the well-being language that accompanies them. Whether campaign periods alter positive affect and hope, and whether happiness aligns with action language, remains unresolved. We analysed 364,118 public Twitter/X posts from Earth Day, Earth Hour, Global Climate Action Day and World Environment Day in 19 occurrence-years, using 30-day pre-event, event and post-event windows. A versioned weighted lexical model estimated happiness, future-oriented hope, collective capability, distress and action language. Event-period happiness prevalence was 9.02 percentage points higher than the pre-event baseline , whereas paired occurrence contrasts showed a 10.75-point decline in action language, indicating a happiness--action divergence. The happiness estimate remained positive across composition and text-deduplication checks, but was less precise under a 19-cluster wild bootstrap. Happier source posts had lower odds of an observed matched retweet cascade.

Analyzing neuroimaging data requires specialized coding and statistical expertise, which limits accessibility for researchers without computational backgrounds. We present the AI Neuroscientist, a language agent for interactive data exploration. The system integrates a large language model (LLM) with a neuroimaging toolset to perform quality control, modeling, and visualization. This allows researchers to query data quality and specify analysis parameters directly in natural language, providing a transparent and interactive alternative to conventional scripted pipelines for small-scale data exploration. We demonstrate these capabilities using functional near-infrared spectroscopy (fNIRS) data, and evaluate the agent on a custom fNIRS benchmarking suite against general-purpose LLM agents with code sandboxes. Future extensions will generalize the architecture to additional modalities, including functional magnetic resonance imaging (fMRI) data, and expand the benchmarking suite to additional fNIRS tasks.

Water treatment research is expanding rapidly, but much of the knowledge acquired from this research remains scattered across unstructured literature. The field still lacks a dedicated language model that can efficiently capture water treatment-specific domain semantics for large-scale literature mining. Here, we address this by developing WaterBERT, a domain-adapted encoder model designed for semantic representation and structured information extraction from water treatment texts. WaterBERT was developed by continual pretraining on a large-scale water treatment corpus comprising about 2.97 billion tokens. Three fine-tuned models based on WaterBERT were systematically evaluated on downstream tasks, achieving the best overall performance among general-purpose and domain-specific BERT models, with F1 scores of 90.12% for multiclass treatment process classification, 79.50% for named entity recognition, and 74.04% for relation extraction. Beyond these benchmark tasks, we further demonstrated WaterBERT's advantages for large-scale literature processing. Applied to 5,144 Environmental Science & Technology articles, WaterBERT-BERTopic identified coherent, diverse, and domain-specific research topics without predefined categories. Building on WaterBERT, we processed 693,211 abstracts at substantially lower cost than commercial LLMs while retaining competitive extraction performance to construct a structured water treatment knowledge graph. The knowledge graph was then integrated with lexical and dense retrieval to develop a Water Knowledge-Enhanced Retrieval System (WaterKERS), which achieved a relevance score of 77.7, substantially outperforming text-based retrieval baselines (54.7-64.5). Through WaterBERT, this study provides a compact and scalable semantic foundation for large-scale information processing and evidence mapping in water treatment research.

We introduce SWE-Serve, a benchmark for evaluating agents on production inference engineering tasks. Implementing an inference feature can require coordinating multiple changes across the serving stack, including model support, runtime execution, and public APIs. Existing benchmarks provide limited coverage of production inference engineering: repository-level software engineering benchmarks do not target inference, while general terminal-agent benchmarks include only a few inference tasks. Dedicated inference benchmarks, meanwhile, focus primarily on isolated kernel generation or performance optimization rather than repository-scale production feature implementation. SWE-Serve provides 53 repository-grounded tasks derived from recent production changes to SGLang, spanning six inference engineering families. Each task executes on either CPU or a single GPU (H100) and is evaluated with hidden functional and regression tests, including, where applicable, end-to-end (E2E) serving tests and calibrated performance gates. Executable no-op and oracle controls, adversarial verifier review, and closed-book execution support task validity and evaluation integrity. Across 11 models and 31 model-effort configurations, the best-performing configuration achieves 75% mean pass@1. SWE-Serve exposes a substantial gap between completing tasks locally and achieving production correctness. On 19 tasks with end-to-end coverage, model-serving E2E tests reject roughly one-third of patches that pass every other test (45.9% under the verifier versus 69.4% with E2E tests excluded from scoring), with pass rate increasing for each model's best-performing configuration. By making the production correctness gap directly measurable, SWE-Serve enables the field to track whether future agents move beyond completing tasks locally to achieving production correctness.

Long-context LLMs focus on retrieving distant evidence from extensive context, yet existing work has largely focused on overcoming distance alone. In this work, we identify the Proximity Trap, insufficient attention to distant evidence often arises less from distance itself than from cumulative competition with abundant, task-irrelevant proximal background. To address the Proximity Trap, we introduce LYRA (Long-context heavY-tailed Relevance Alignment), a t-distributed directional matching mechanism that reshapes the context retrieval distribution, directing more attention mass toward task-relevant evidence, while preserving the relative positional information encoded. Extensive experiments on LongBench-v2, RULER, and LongBench demonstrate consistent improvements across context lengths and task categories. We further introduce ProxBench, a multi-level fine-grained benchmark for evaluating distant evidence utilization under increasing proximal background interference. Project page: https://xiaoyuyoung.github.io/LYRA/

In this report we present results of the ICDAR2026 Competition on Multimodal Reasoning over Documents in Multiple Domains. This competition aimed to advance research in document understanding through the task of Visual Question Answering (VQA). Building upon previous DocVQA benchmarks, this competition introduces challenging reasoning questions over a diverse collection of documents spanning eight domains, including business reports, scientific papers, slides, posters, maps, comics, infographics, and engineering drawings. The competition concluded with 20 valid submissions from 8 teams spanning zero-shot VLMs, OCR and parser-augmented pipelines, agentic retrieval systems, multi-agent ensembles, and fine-tuned multimodal models. The results show that the strongest systems move beyond single-pass prompting and instead rely on structured evidence extraction, retrieval, verification, and orchestration across multiple components.

Scientific investigations into microbial natural products (NPs) present significant challenges for novices, largely due to the complexity of microbial systems, biochemical diversity, technical skill requirements, and the demands of bioinformatics and data analysis processes. To address these issues, we introduce ChatT2, a large language model (LLM)-based agent that is specifically tailored to the unique characteristics of bacterial type II polyketides. These polyketides form a structurally distinct and therapeutically important NP family. ChatT2 was developed within an autonomous multiagent framework composed of a mentor, an executor, and an evaluator, each with defined responsibilities. The mentor acts as an intermediary between ChatT2 and the user, utilizing chain-of-thought prompting to refine the intent of the user. Under the guidance of the mentor, the executor synthesizes multimodal information via retrieval-augmented generation techniques and seamlessly integrates bioinformatics and cheminformatics tools. The evaluator ultimately assesses the output of the executor to ensure the richness and accuracy of the retrieved information. Our research highlights how ChatT2, designed with this multiagent framework, addresses the challenges faced by general LLMs in terms of understanding limited, specialized corpora and complex biological information and provides both experts and novices with a valuable tool for exploring various NPs of interest. The ChatT2 webserver can be accessed at https://chatt2.site/#/chat.

Learned-memory methods store information in an explicit table and consume it through a separate reader, allowing addressing, storage, and reading to be modified independently. We study whether useful memory can also be generated rather than only retrieved. MemoryAthena uses three pathways: direct Engram retrieval (E), generation from retrieved Engram cues (GE), and generation from causal backbone states without consulting the memory table (GH). Generated memory is conditionally useful: it can complement E in one context but interfere with it in another. MemoryAthena therefore treats E as an anchor and learns when a generated representation should intervene. With the backbone, memory, generators, and readers frozen, a lightweight causal routing head is trained from counterfactual future-token likelihood advantages of GE and GH relative to E. At inference time, an admitted candidate modifies the E residual through bounded interpolation, while rejection recovers the direct pathway exactly. On question answering, MemoryAthena raises the five-task average from 37.65 to 39.28 over the direct pathway of the same checkpoint, while the six-task general-NLP average increases from 76.73 to 79.13. The complete memory-side system contains approximately 201M parameters, excluding the frozen backbone. Further analyses show complementary strengths among E, GE, and GH across tasks and inputs. These results support generated memory as a selective correction to direct retrieval and highlight routing when, which, and how strongly to intervene as the central challenge.

In this paper, we show that LLM and human coding are observationally equivalent in terms of annotation quality: recent LLMs agree with expert coders at rates comparable to those observed among experts themselves. We demonstrate this through replications of text-classification tasks from 14 peer-reviewed political science studies, in which ten LLMs, three human experts, and 165 crowdsourced workers independently classify the same texts using identical codebooks. We find that this equivalence is driven by ambiguity in the texts and coding rules. When LLMs disagree with experts, experts are also more likely to disagree with one another, and clarifying coding rules reduces disagreement among both experts and sufficiently capable LLMs. Thus, there is little empirical basis for preferring human coding on the basis of annotation quality alone, while LLMs offer substantial advantages in speed and cost. We therefore argue that the central challenge of text annotation is no longer choosing between human and machine coders, but developing coding rules that minimize ambiguity and accounting for the ambiguity that remains. To this end, we propose using disagreement across LLMs to identify difficult cases and refine codebooks, and we develop ambiguity-aware bounds for downstream inference when a unique annotation cannot be defined for every text.

Autonomous agents that automatically build artificial intelligence (AI) models could broaden access to AI across science and engineering. A popular line of such agents frames model building as a code search problem and solves it by tree search, in which each node is a candidate program and the tree grows by generating a child program from a parent, and these agents now approach the capability of experienced AI engineers on realistic benchmarks. However, these agents have three weaknesses in efficiency that have not been fully addressed. First, only a small number of candidates can be executed within a realistic budget, so search rules that rank nodes by executed rewards, such as Monte Carlo-style tree search, rely on few and noisy scores and select the next node to explore less effectively. Second, no resource-aware strategy is used to schedule training jobs, which can lower hardware utilization and training efficiency. Third, every agent call is served by a single powerful model, which inflates inference cost. Here we introduce AIBuildAI-2.5, an agentic system that carries out the tree search with LLM agents and addresses each of the three issues. AIBuildAI-2.5 proposes a novel LLM-guided tree search, in which a judge scores each candidate on its expected improvement, grounding, and feasibility, and a selector ranks the pool of candidates from these scores and the state of the search. In addition, AIBuildAI-2.5 comprises a scheduler that launches training jobs with the current hardware resource status taken into account and a router that assigns lower-cost LLMs to less demanding tasks while reserving the most capable LLM for the most challenging sub-tasks in the AI model building workflow. AIBuildAI-2.5 ranks first on MLE-Bench with a medal rate of 73.3%, and outperforms a strong baseline on six autonomous AI research tasks from AIRS-Bench.

Structured-data agents in enterprise settings must reason over complex data environments whose relevant evidence is distributed across schemas, relationships, policies, and recurring business roles. Modern agentic systems often address this burden through reusable markdown-style memory or skill files that preserve previously discovered information for later queries, reducing the need to rediscover the same structure repeatedly. This is useful, but it obscures a natural division of labor: agents are well suited to semantic reasoning, while learned systems are well suited to predicting and organizing recurring structure. We introduce latent equivalence learning to bridge this gap. The framework separates persistent task-relevant identities from their dataset-relative realizations. In our realization, supporting and opposing evidence shape support-realized Gaussian prototypes that learn how those identities are expressed in a particular data environment, while soft-membership profiles retain distinctions lost under a hard assignment. A separate learned query-prototype system represents recurring evidential requirements and maps them through a learned compatibility function into the same persistent identity structure. This identity-factorized, query-conditioned routing materializes the relevant dataset-specific evidence for downstream reasoning, allowing the agent to operate over an already organized evidential state rather than reconstructing cross-schema structure at every query. On the Data Agent Benchmark, spanning 54 queries across 12 heterogeneous datasets, our full implementation achieves 94.67% dataset-macro stratified Pass@1 over five complete trials and 258/270 successful raw query attempts, compared with 55.51% for the benchmark's Claude Opus 4.6 reference agent, ranking first among 40 leaderboard entries at submission.

Dialogue failures in language models are usually framed as memory failures: context too long, summaries lossy, a constraint forgotten. We argue this misses a deeper problem: in many conversations the model does not forget, it commits too early. An ambiguous early turn collapses into a single hidden interpretation, and later clarification is filtered through that commitment. We call this early posterior collapse: unresolved user intent collapsing into a committed task state before ambiguity is resolved. We study it with controlled dialogue tasks in writing, planning, and coding using Gemini-2.5-Pro and Gemini-2.5-Flash. Across thousands of trials, the same information in different orders yields different outcomes, even when the final dialogue contains equivalent task-relevant information. This order effect suggests later clarification is treated as extra context rather than a corrective signal: it refines a stale task state without invalidating it. Coding tasks are especially vulnerable, suggesting early assumptions get embedded in structured artifacts such as interfaces and control flow. Standard prompting and memory strategies do not reliably help: summaries can collapse ambiguity, and chain-of-thought can reduce explicit wrong commitment in reasoning traces without improving final task success. These findings motivate uncertainty-preserving state management. If assistants cannot let go of early interpretations, robustness cannot rely on post hoc correction alone; it must keep ambiguous early turns from hardening into one task state. Assistants should hold tentative hypotheses while ambiguity remains, ask before executing when high-impact ambiguity persists, and rebuild from a revised state when later evidence invalidates an earlier reading. Rather than one prompting fix, we aim to redirect research for interactive LLMs from retaining more context toward preserving uncertainty.

Marketers now deploy generative AI agents as synthetic consumers to pretest visual assets such as logos, packaging, and advertising at a fraction of human-panel cost. However, this procedure assumes that a model seeing a visual cue can also perceive its consumer meaning, which is largely untested. We stress-test the assumption using six canonical visual marketing experiments, varying the two levers managers control: model generation (GPT-4o-mini vs. GPT-5.4-mini) and input format (plain text vs. JSON). Every resulting configuration passed the manipulation checks; however, none of the configurations reproduced more than two of the six human effects, and the remainder were nonsignificant. The one exception was a significant reversal of the human pattern. Providing conceptual or empirical evidence through in-context learning steers average responses toward the human effect. Yet steering has a limit: even when it succeeds, a configuration reproduces less than half of the natural spread of human responses and so understates consumer heterogeneity. We integrate these results into an AI governance protocol (Calibrate, Intervene, Deploy) that delineates when synthetic consumers can responsibly screen creatives and when human panels remain necessary.

Post-training quantization and pruning rely on a small calibration corpus. Whether specialized domains such as finance require domain-matched calibration data remains unsettled. We argue that the answer depends on the task-level damage caused by compression rather than on domain mismatch. If compression preserves the target capability, changing the calibration corpus has little effect. If compression causes large losses, task-formatted calibration can recover part of the loss. We test this hypothesis across two model families, six compression configurations, three token-matched calibration corpora, and ten financial classification and numerical question-answering tasks. The results support this hypothesis. Quantization largely preserves task performance, and calibration choice has little effect in this case. Pruning reduces numerical QA accuracy by over 40 points. In these damaged settings, another generic corpus does not help, while FinMix, a mixture of financial task examples, recovers a large part of the loss. The link between damage and recovery holds across model families and scales. These findings support a practical rule. Measure task-specific compression damage first, and construct specialized calibration data only when the damage is large.

Retrosynthetic planning advances through decisions that reshape the remaining chemical problem: a locally plausible disconnection can leave precursors whose chemoselectivity constraints complicate the rest of the route. Existing planners often channel model proposals through search or template procedures, leaving open whether a general-purpose large language model (LLM) can itself sustain and revise route strategy. We developed Rachel, a stateful environment that executes and checks LLM-directed chemistry but prescribes neither a search policy nor a stopping rule. Without supplied reference routes or route-level solutions, GPT-5.5 achieved strict closure for 111 of 120 PaRoutes120 targets and 24 of 25 targets in the separate RF25 difficult-target cohort. RF25 was drawn largely from studies published after GPT-5.5's reported knowledge cutoff. Closure required complete routes and independent source resolution of every terminal precursor after planning. On a shared PaRoutes subset, forward-model support exceeded that of most comparator methods, and Rachel received the highest mean overall route score from both method-blinded LLM evaluators. Recorded trajectories showed continued model-proposed chemistry, with revised strategies carried into subsequent steps. Replacing LLM route decisions with fixed policies reduced strict closure to 6-15/120 despite continued local chemical execution; restricting planning support also reduced closure in RF25. Within Rachel, a general-purpose LLM coordinated successive chemical choices and revised its strategy as earlier decisions reshaped the remaining problems.

Deploying Large Language Models for runtime operational triage incurs prohibitive latency (>100-500 ms), high VRAM requirements (>4-8 GB), and excessive energy dissipation. Extending Mandelbrot Fractal Neural Synthesis (Dagli et al., 2026), this paper presents the Universal Fractal Natural Language Decision Map, realized via the werr machine-native edge reflex runtime and the production answerr platform (https://answerr.me). Operating entirely without stored weight tensors (0 Bytes VRAM), the engine synthesizes deterministic decisions---noul (Boolean), choice (categorical), and score (ordinal)---by dynamically modulating 24-byte coordinate seeds along the chaotic boundary of the Mandelbrot set and evaluating 4-quadrant escape dynamics. Drawing inspiration from biological System-One reflex arcs, the engine introduces: (i) an Auto-Seed Router with domain projector Phi_D yielding a +28.8% accuracy gain over linear baselines; (ii) an Information-Theoretic Acoustic Damping Filter grounded in token entropy and phonetic spectral density that insulates against prompt injections (0.0% empirical bypass; 95% Wilson CI: [0.0%, 30.8%]) while pruning escape iterations by 45.8% (accelerating throughput 2.5x to 3.31 ms latency); and (iii) an Organic Dynamic Calibration framework using O(1) Exponential Moving Average (EMA, alpha=0.03) and quadrant phase rotation to eliminate positional bias. Benchmarked on bare-metal infrastructure (api.answerr.me:4431) across 1,150+ verified decisions (3,200+ questions) and ranked World #1 on the independent JevBench suite (81.65%), the framework achieves 92.6% macro-accuracy (95% CI: [90.8%, 94.1%]) with 7.08 ms median CPU latency. We provide an OpenAI-compatible API (/v1/chat/completions) and demonstrate feasibility on microcontrollers and 32-byte EVM smart contracts.

Soccer tactics are interactive: an attacking action changes the opponent's defensive problem, and the observed response depends on the multi-agent match state. We introduce SambaGraph, an action--reaction spatio-temporal graph dataset and benchmark for soccer tactical response modeling. From tracking and event data for all 64 matches of the 2022 FIFA World Cup, we curate 4,070 action-centered episodes represented as temporally aligned 23-node player--ball graph sequences with attack/defense views, response labels, and 26,270 split-safe attack--defense pairs. We study three questions: whether observed responses can be classified from graph episodes, whether successful defenses can be retrieved for a query attack, and whether graph-derived summaries support grounded LLM reasoning. A compact signature MLP obtains $0.796\pm0.007$ macro-F1 for response classification, while a fused graph--signature dual encoder reaches $0.471\pm0.029$ Hit@5 and $0.655\pm0.051$ Hit@10 for full-bank defensive retrieval. Hard negatives maximize pair discrimination but not retrieval quality. Local LLMs underperform supervised encoders for direct classification and do not improve over a strong original order in eight-candidate reranking, but they provide grounded tactical rationales. These results position SambaGraph as a reproducible benchmark for graph-based soccer strategy-response research. Code and dataset are available at: https://github.com/areyesan/SambaGraph.

Recent code generation research has moved from isolated function completion toward repository-level generation in existing codebases. To implement a target function correctly, an LLM must identify reusable repository dependencies such as existing functions, APIs, and cross-file definitions. Existing retrieval methods provide such context through code similarity search, persistent whole-repository graphs, or LLM-driven graph exploration, but often incur high graph construction, reasoning, and token costs. Feature-oriented methods offer a natural view of software functionality, yet they mainly support requirement decomposition, planning, or feature editing rather than code dependency retrieval. This paper presents \textbf{FeatLens}, a feature-guided dynamic code graph construction and retrieval approach for repository-level code generation. FeatLens builds a feature index that links natural-language feature descriptions to function-level code entities. Given a generation task, it dynamically constructs a task-specific seed graph from the feature index and applies semantic-structural graph reasoning with personalized PageRank to select a compact reasoning graph. This design replaces persistent whole-repository graph maintenance and LLM exploration with deterministic and lightweight dependency retrieval. Experiments on DevEval and EvoCodeBench show that FeatLens achieves the best DR@15 among sparse, dense, and graph-based baselines (0.501 and 0.460). On DevEval generation, it obtains the highest DIR@1, reaching 52.91\% with DeepSeek-V3.2 and 53.58\% with GPT-5-mini, while maintaining competitive Pass@1 and producing shorter code. Compared with the strongest graph-based baseline, FeatLens reduces graph nodes by 61.0\%, edges by 86.2\%, and total token overhead by 45.9\%, with no LLM tokens used during retrieval.

A multi-agent system can reduce latency on complex tasks by executing work concurrently. Several pioneering harness frameworks support multi-agent systems. However, the scalability of current multi-agent harnesses is often constrained by a central orchestrator's capacity to allocate tasks and coordinate workers. To address this limitation, we introduce Agensh, a scalable self-organized multi-agent harness without a central orchestrator: concurrent workers execute a multi-agent cooperation loop, continuously gathering context, claiming and self-assigning sub-tasks, taking action and sharing findings, verifying results, and merging progress in an asynchronous manner. The loop is supported by the agentic organization infrastructure comprising three components: a shared workspace holds proposed, ongoing, and completed work; a message interface lets workers communicate; and shared context retains reusable findings and work intentions. To test the scalability of Agensh, we evaluate it on the five hardest ProgramBench tasks with GPT-5.6-sol (high). Scaling from 1 to 128 agents raises the mean final test-pass rate from 19.31% to 28.78%, an approximately 49% relative improvement. Larger organizations reach comparable test-pass rates earlier. On pandoc, scaling from 1 to 1,024 agents raises the final test-pass rate from 33.89% to 55.06%. Worker trajectories further show that different forms of self-organized cooperation gradually emerges and standardizes as the organization grows. These results reveal the number of agents as a new scaling dimension for multi-agent organizations to expand the frontier of general intelligence, offering a practical solution for complex tasks under hard latency constraints or time budgets.

QMSum provides no scorer, making query-focused meeting summarization results difficult to compare. We rescore or generate 15 systems under one implementation. Through a common inference port, a released 406M Fusion-in-Decoder specialist loses 6.30 ROUGE-1 when moved from capped long input to 2,000-word retrieved spans. Fine-tuning it on this span regime recovers the loss. On test it scores 36.33 ROUGE-1 versus 35.41 for our 1.2B system; the meeting-cluster 95% interval for the difference is [-0.27, +2.22], so QMSum does not statistically separate them. The smaller system uses about one-third as many total parameters and less than half the peak inference memory. Within the fixed 1.2B base, span-regime fine-tuning adds 5.29 [+4.02, +6.56], while replacing the first 4,500 transcript words with 2,000 retrieved words adds 1.55 on test and 0.29 on validation. Separately, under one concise prompt and reference-overlap scorer, a released 406M specialist exceeds five proprietary hosted models by at least 6.2 ROUGE-1, but output length and absent human or factuality evaluation limit this ordering. Conclusions are limited to QMSum and automatic metrics.

Vision-Language Models (VLMs) offer promising capabilities for automated sports coaching but face a fundamental limitation: they implicitly compare against professional standards, making their feedback impractical for developing players. We present CricRAG, a retrieval-augmented framework that aligns VLMs with skill-appropriate benchmarks for personalized cricket coaching. Our key insight is that by retrieving similar-but-better techniques as reference points, we can guide VLMs to provide developmentally appropriate feedback that mirrors human coaching practices. We contribute: (1) a labelled dataset of 288 cricket technique videos spanning multiple skill levels, (2) an efficient motion retrieval pipeline using contrastive learning that achieves 78% top-3 retrieval accuracy, (3) a frame sampling technique that reduces inference costs, and (4) a retrieval-augmented approach that significantly improves feedback alignment with coaching principles, achieving up to 94% agreement with professional assessments compared to 67% without retrieval context.

Runtime traces can appear transparent, but a closed-loop policy determines which states are visited and which failures become visible. We study a simulated hotel-pricing agent mapping time, inventory, and market state to discrete price actions under varying demand regimes. A fault may leave no aggregate trace when the policy rarely visits affected cells. We treat entry into aggregate-only fault interpretation as a diagnosability decision preceding scoring or localization. A reference-map gate requires repeated clean-policy support; a matched runtime gate then requires joint support in clean and current streams. Signal analysis occurs only after both pass. We calibrate false admission on a disjoint clean stream at the physical-component level and model detection by affected clean traffic rather than nominal cell coverage. In a frozen one-shot heldout, 55/72 (76.4%) regime-component units were reference-admitted, representing 20 physical components; 54/55 passed matched runtime admission, while the rejected unit abstained. Stable false admission was 0/20, with a one-sided exact 95% upper bound of 0.1391, meeting the frozen 0.20 criterion. Across 540 repeated unit-arm rows nested in those 20 clusters, affected clean traffic reduced negative log likelihood by 29.3% relative to cell coverage, a gain of 0.1264 nats per row (cluster-bootstrap 95% interval [0.0593, 0.1918]). Adding mask family and its interaction improved log loss by 0.0015 nats per row (one-sided upper bound 0.0066), below the frozen 0.01 practical-sufficiency margin. A development audit found that exact minimum hitting set and greedy selection chose identical supports in 12/12 scenarios because singleton evidence had resolved the conflicts. The result is a bounded rule for interpreting aggregate agent behavior: first establish exposure, then score change, and abstain when the trace cannot support the claim.

We introduce Qwen3.8-Omni-Flash, a natively multimodal agentic model for real-world multimodal productivity. Compared with previous omni models, which primarily emphasized perception and interaction, Qwen3.8-Omni-Flash substantially improves multimodal understanding and reasoning, as well as performance on long-horizon agentic tasks. These capabilities are supported by a native multimodal co-training strategy that preserves strong text-domain capabilities while facilitating the transfer of agentic capabilities from text to audio and video tasks. The model inherits the sparse mixture-of-experts (MoE) architecture of Qwen3.8-Next and extends the context window to one million tokens, supporting long-context multimodal reasoning and long-horizon planning. These advances enable integration into production workflows as a primary agent or a specialized sub-agent, supporting video editing, long-form audio and video translation, music-conditioned music video or movie generation, and video-based note or omni-skill creation. To address the lack of native audio and video support in existing agent harnesses, we release Qwen-MM-Plugins, a lightweight open-source plugin framework for multimodal productivity. We further frame real-time multimodal interaction as a system-level challenge requiring orchestration of context and memory management, tool use, and sub-agent delegation. Accordingly, we release Qwen-Live-Harness, an open-source framework for building responsive, real-time multimodal agents based on Qwen3.8-Omni-Flash. Extensive evaluations demonstrate that Qwen3.8-Omni-Flash achieves strong performance across multimodal understanding, reasoning, long-horizon agentic execution, and video productivity tasks. These results and the accompanying open-source tools support Qwen3.8-Omni-Flash as a practical foundation for deploying natively multimodal agents in research and production.

Large language models (LLMs) exhibit a parametric vulnerability to adversarial swarm consensus. To mitigate this sycophancy, we introduce Contrastive Epistemic Decoding (CED), a zero-shot inference intervention. Unlike standard Contrastive Decoding (CD) which relies on a weaker secondary model, CED utilizes a dual forward-pass on a single architecture to isolate conformity bias. By introducing a novel asymmetric, zero-bounded probability clamp and discrete top-k truncation mask, CED mathematically suppresses toxic consensus tokens without causing grammatical collapse. Evaluated across 7,200 paired trajectories on complex benchmarks (GAIA, SWE-bench, Multi-Challenge) using Gemma-2 (9B), Llama-3.1 (8B), and Mistral v0.3 (7B), CED successfully neutralizes architectural and positional biases. By reducing cognitive loafing by up to 33.00% absolute, CED drives significant performance gains, yielding up to a 30.75% accuracy recovery. Regaining sovereignty induces distinct architectural behaviors---passive task-focus in Gemma-2 and active refutation of the simulated swarm in Llama-3.1---showing CED decouples compliance from capability without fine-tuning.

Active sequential hypothesis testing studies how to identify an unknown hypothesis with a given set of sensing actions. We study this in the setting of identifying large language models (LLMs), \textit{i.e.}, if a user is conversing with an LLM drawn from a known set of models, how can they identify which one is in use? Here, the available sensing actions (evaluations) are themselves a design choice: an evaluator must first decide which environments and prompt families to construct, and only then decide how to use them sequentially. We formalize these two levels as an outer probe-design problem and an inner identification problem. Simply put, the outer stage selects a set of probes to be sent to the entire set of models, creating a kind of fingerprint dataset. This is followed by the inner stage, which sequentially sends a budget-minimizing set of those probes to identify the model in use. For the outer problem, we show that selecting which evaluations to construct at minimum cost, so that every pair of candidates is distinguished, is exactly a weighted set cover problem. Since the response distributions of the candidate models are not known exactly but only through calibration samples, we give a one-shot procedure that estimates the cover instance from these samples. For the inner problem, we bound the number of evaluations needed to identify the unknown model in terms of how well the available evaluations distinguish each pair of candidates.

Multimodal large language models (MLLMs) often struggle with fine-grained visual perception when processing complete images, as critical evidence may only appear in local regions. On-policy self-distillation (OPD) enables transferring privileged visual knowledge from informative views to full-image policies, but querying the teacher for every rollout introduces substantial supervision costs. In this work, we propose BAS-OPD, a budget-aware selective OPD framework that allocates teacher supervision under limited query budgets. Instead of querying all rollouts, BAS-OPD selects informative samples while maintaining full-batch student generation. We explore random, uncertainty-based, and learned utility-based selection strategies, where the learned selector estimates query value from detached rollout statistics and online utility signals derived from student--teacher agreement and teacher confidence without additional student forward passes. BAS-OPD only changes training-time supervision allocation and preserves single-pass full-image inference. Experiments on fine-grained multimodal perception benchmarks demonstrate that BAS-OPD achieves strong performance while substantially reducing teacher supervision costs, highlighting the effectiveness of selective OPD under constrained budgets.

As automated scientific discovery advances, Large Language Models (LLMs) can now generate research ideas at an unprecedented scale, shifting the bottleneck from idea generation to idea evaluation. Existing evaluators mainly rely on parametric LLM knowledge or unstructured retrieval, producing judgments that lack the experience-grounded reasoning used by human instructors. To address this, we propose DeepInstructor, an agentic framework that formulates idea evaluation as reasoning over structured scholarly experience. DeepInstructor constructs an Experience Graph from 58,607 peer reviews and employs a ReAct-based agent to retrieve dimension-specific evidence for traceable evaluation. We further introduce DeepInstruct, a dataset with controlled pairwise comparisons across novelty, significance, and feasibility. Experiments show that DeepInstructor substantially outperforms existing baselines, improving Hit@1 and Hit@2 alignment with human judgments by 24.4% and 29.7%, respectively. Our findings suggest that scientific idea evaluation can be grounded in explicit reasoning over structured scholarly experience

Context compression is widely proposed as a way to cut the token bill of LLM coding agents, and public benchmarks report that aggressive compression preserves task-solving quality. These two facts do not imply the third one commonly assumed: that compressing file reads saves money in a real multi-turn agent. We instrument a production compression gateway (Paritok) between coding agents (Claude Code, Codex) and frontier LLMs (Claude Sonnet, GPT-5), and decompose the token bill of real sessions into three independent levers: tool-schema filtering, content compression of file reads and tool output, and history summarization. Measured in isolation under controlled A/B runs, the three save at fundamentally different rates. Tool-schema filtering removes a fixed block every turn, roughly 21K-57K tokens on a typical turn; it is linear in the turn count N and the only unambiguously and reproducibly positive lever. Content compression saves only about 2% of the cache-priced prefix per turn, but compressed reads accumulate in history and are re-sent on every later turn, so its cumulative saving grows quadratically, about 3350*N^2 tokens (measured), overtaking the fixed tool-filter saving within roughly 6 turns until the context window caps it. A non-destructive gateway lets the agent pull original bytes back on demand; each recall re-sends exactly the one segment just compressed away, so its cost is fixed and bounded rather than a multiplicative blowup, and heavy recall spends the accumulated saving back one segment at a time. Finally, a strong single-shot compression benchmark - 86.5% of SWE-bench quality retained at a 25.7% compression rate, achieved by the model this gateway deploys (Paritok-4B, reported separately) - is orthogonal to multi-turn agent cost and must not be cited as a cost-saving argument. We distill the results into an actionable recipe for where token-saving effort pays off.

A classical question in statistics is which observable quantities to condition on when drawing inferences about unobservable targets. For conformal prediction in nonparametric uncertainty quantification, standard marginal validity offers limited resolution at the prediction values on which decisions are based, and fully conditional guarantees with respect to the covariates are provably unattainable. We address this gap by introducing a prediction-based conditioning framework that we refer to as Prediction-Interval-Conditional Prediction Intervals (PICPIs). Formally, a PICPI is an interval $I$ satisfying a self-consistency condition: $$\mathbb{E} [Y \mid p(X) \in I] \in I,$$ for predictive model $p$, contextual covariate $X$, and outcome $Y$. Thus, an interval simultaneously defines a stratum of prediction values and certifies that the mean outcome in that stratum lies in the same interval. This self-consistency condition yields data-adaptive strata without altering the original prediction. Such intervals can be constructed using practical algorithms. Under regularity of the prediction distribution, the constructed intervals cover all but an arbitrarily small fraction of prediction values and have widths that decrease at rate $n^{-1/3}$, up to logarithmic factors and the prediction error. Moreover, identifying these locally calibrated intervals can, in turn, inform downstream decision-making. We derive inference procedures for PICPIs in probabilistic prediction and multi-class classification, accompanied by theoretical guarantees. Empirical results are provided that compare PICPIs with existing interval-based baselines.

Graphical User Interface (GUI) Agents autonomously interact with software to fulfill user requests, where GUI navigation stands out as the most critical and challenging capability. Mastering this capability demands a complex synergy of step-wise decision-making, state-action alignment, and long-horizon planning. While directly mixing these corresponding navigation tasks seems intuitive to simultaneously acquire these skills, such a direct combination is severely bottlenecked by inconsistent optimization objectives and profound data heterogeneity. To overcome these barriers, we propose the MaP (stands for ``\textbf{M}asked Tr\textbf{a}jectory \textbf{P}rediction''), a unified framework that seamlessly harmonizes divergent GUI navigation tasks. By modeling multi-turn GUI interactions as a trajectory and defining training objectives through component masking and prediction, MaP shifts the optimization from task-specific marginal distributions to a consistent objective. Furthermore, to handle the data heterogeneity across multiple navigation tasks, we design a role-aware adapter learning module that dynamically routes each token to a specialized representation space. Extensive experiments on five representative GUI navigation benchmarks demonstrate that MaP effectively mitigates gradient conflicts and significantly outperforms the direct mixture training, establishing a robust paradigm for multi-task GUI navigation.

Pathologists integrate morphology across magnifications and across the slides of a patient case, whereas pathology foundation models encode thousands of tiles from single slides and aggregate their features. Here we present WILSON, a vision--language foundation model that represents whole-slide images and multi-slide cases as single multi-magnification composite images, trained on approximately 189k slides from Mayo Clinic spanning 42 organs and 829 diagnostic entities using pathology reports as supervision. Without task-specific training, WILSON exceeded a dedicated case-level model on all internal cohorts (macro-F1 0.52 versus 0.38) and matched slide-level models up to 9.4 times larger at 272- to 2,155-fold lower compute. End-to-end fine-tuning on 508 triple-negative breast cancer cases improved histologic subtyping and stromal tumor-infiltrating lymphocyte grading by 0.16 and 0.11 macro-F1. WILSON retrieved matching diagnostic text at 75.6% recall@1 (PRISM, 58.1%) and generated captions closer to report-derived references than PRISM and PRISM2 on the internal cohort and on most external comparisons. Composite images thus offer a compact, clinically aligned computational unit for pathology.

Large language model agents are increasingly capable of conducting research autonomously, producing research documents alongside the code and experiments that ostensibly support them. Yet whether the reported findings are consistently supported by corresponding implementations and execution evidence remains largely unexplored: existing review practices primarily assess textual quality and cannot reliably identify inconsistencies such as hard-coded metrics, unimplemented methods, or unsupported experimental results. We present ReAgent, an automated auditing framework for assessing the consistency between agent-generated research documents and their associated repositories. ReAgent constructs structured representations of scientific claims from research documents and uses them to guide repository analysis and evidence collection. Static auditing examines whether claimed methodologies, implementations, and experimental configurations are consistently reflected in the repository, while dynamic auditing executes relevant experiments and collects execution evidence to assess empirical findings. By combining static analysis with dynamic evidence, ReAgent identifies inconsistencies that may remain hidden under either perspective alone, such as experiments that reproduce reported numbers while deviating from the claimed methodology. The collected evidence and audit decisions are organized into a structured repository-level audit report, enabling transparent evidence traceability. We evaluate ReAgent on a manually curated benchmark of agent-generated research document--repository pairs and compare it against representative static and reproduction-based baselines. Experimental results demonstrate that ReAgent effectively identifies inconsistencies between reported research findings and their supporting repository evidence.

Legal argument mining supports passage classification, retrieval, and argument completion. This work introduces an expert-annotated dataset of 42 U.S. federal tax opinions on corporate reorganizations under I.R.C. {\S}368. To our knowledge, it is the first expert-annotated, tree-structured argument corpus for this domain. Explicit spans receive one of five functional labels: Rule, Analysis, Conclusion, Background Facts, and Procedural History. Rule, Analysis, and Conclusion spans can be linked into directed support trees, while Background Facts and Procedural History serve a contextual function. The corpus provides span-based, sentence-based, flat, and tree-structured representations. Agreement analysis shows that functional node labels are more reliable than directed support edges and implicit intermediate conclusions. Directed-path agreement is stronger than direct-edge agreement, which indicates that broad reachability is more stable than exact local decomposition. Classification experiments show that functional labels are learnable under case-disjoint evaluation. Retrieval experiments show that supervised fine-tuning improves within-case retrieval. However, cross-case generalization remains weak. The dataset supports legal passage classification and provides a conservative benchmark for structured argument mining in U.S. federal tax case law.

Uniform discrete flow permits repeated updates at every generation position. While continued revision supports correction of wrong tokens, it also exposes correct intermediate predictions to later errors. An experiment on Sudoku puzzles shows that 9.4% of generated cells are correct at an intermediate step but incorrect in the final output. We introduce generation order into uniform discrete flow through selective absorption, which fixes chosen predictions while preserving the uniform-flow velocity at active positions. To prevent absorbing incorrect predictions, we propose Low-Entropy Discrete Flow (LEDFlow), a training-free sampler that adaptively orders absorption by local entropy. By decomposing absorption error into joint dependence and conditional prediction terms, we show that selecting the lowest-entropy positions under a fixed absorption budget minimizes an upper bound on the conditional term. We further support the choice of local entropy by showing that the decision-error bound of global lookahead grows with the lookahead window under an imperfect denoiser. Across reasoning benchmarks, LEDFlow attains 0.845 Nikoli Sudoku solve accuracy, with the largest gains on strongly constrained tasks. On text-to-image generation it attains the best overall score, and on multimodal understanding it improves over the native sampler on all six benchmarks, at an inference cost comparable to standard flow sampling.

Causal tabular foundation models amortize effect estimation across synthetic mechanisms, but latent-effect supervision rewards posterior shrinkage instead of directly encoding the repeated-sample response needed in a fixed deployment population. We introduce fluctuation-supervised pretraining (FSP): each synthetic table is labeled by its average treatment effect plus its efficient influence-function fluctuation, while deployment remains a single frozen forward pass. Along the path $T_{\lambda,P}=\theta(P)+\lambda P_n\psi_P$, we prove an endpoint transition: every fixed $\lambda<1$ retains label ambiguity of order $(1-\lambda)^2/n$, whereas full fluctuation makes the Gaussian label observable and reduces optimal finite-stratum causal label-prediction risk to order $n^{-2}$. One finite-pretraining bound combines label, network, episode-sampling, and optimization errors; its resulting sampling defect controls fixed-mechanism bias, mean squared error, variance, Gaussian approximation, and, with variance-head accuracy, studentized coverage. Complementary lower bounds separate the local $n^{-1}$ ATE risk that deployment observations cannot erase from the $\log N/M$ excess risk of a generic finite-dictionary episode-learning problem. Experiments trace the learned sampling response. With a raw-row/column backbone, FSP reduces large-effect-shift RMSE by 69.8% relative to latent supervision and by 39.5% relative to a released CausalPFN checkpoint on matched tables. Continuous-covariate experiments, known-effect semisynthesis and two randomized-study evaluations separate sampling-law fidelity from point-risk shrinkage and expose weak-overlap errors in both learned heads.

Collaboration topology shapes both the performance and execution cost of LLM-based multi-agent systems. Because tasks differ in complexity and required capabilities, recent approaches generate task-specific collaboration graphs that specify agent participation and information flow. However, representative topology generators use either individual agents or predefined groups throughout an organization, overlooking differing collaboration needs across subtasks. Our key insight is to select granularity locally for each functional role, combining fine-grained control with reusable collaboration patterns within one organization. Learning such organizations requires exploring a combinatorial construction space with limited intermediate feedback from final-answer rewards. Therefore, we propose MAGIC, a dense-reward reinforcement learning framework for mixed-granularity graph generation. Specifically, MAGIC constructs a mixed-granularity agent graph by sequentially selecting a functional role, instantiating it as a single agent or reusable group, and connecting it to existing units. We directly optimize the construction policy using returns from trajectories sampled under the current policy and use potential-based reward shaping to provide intermediate feedback from probe-based utility and structural signals while preserving the cumulative task reward. MAGIC outperforms state-of-the-art baselines across eight benchmarks and demonstrates strong inference efficiency in our efficiency study.

For a long-horizon LLM agent, the memory question is not what was once recorded but what \emph{currently holds}. Most designs answer it only indirectly: every interaction is stored, and the present is reconstructed at query time by retrieving and reconciling records, so stale values re-enter and the same conflicts are re-litigated. Committing the current value at write time avoids this, but existing write-time (CRUD) memories overwrite, so a wrong update is unrecoverable and prior state is lost. We take the missing combination---\emph{commit on arrival while retaining what is displaced}---and formalize it as \textsc{Memory of Memory} (MoM): memory tracks not only content but the provenance, status, and history of its own entries. We instantiate MoM as \textsc{Provenant Memory} (P-Mem), a typed provenance graph whose \emph{active frontier} exposes one current value per resolved key while displaced values are retained as provenance; typed operations decide whether a new observation supports, supersedes, contests, rejects, revokes, or resolves an existing value. P-Mem's decisive gain is validity rather than accuracy: its turn-level read matches the strongest retrieval memory in accuracy at $\sim$4$\times$ fewer read tokens---a retrieval-granularity effect---while graph-guided turn pruning cuts the knowledge-update stale-answer rate (19.4\%$\rightarrow$10.9\%); on revision chains it stays at 100\% where query-time reading collapses to 25\%, and, because displaced values are retained rather than overwritten, it recovers committed errors a CRUD memory cannot (100\% vs.\ 0\%).

On repeated work, agents are inconsistent. We ran 42 tasks three times each and found that, depending on the model, 38% to 74% returned answers that did not agree. Consistency is what a buyer, an auditor, or a regulator requires, and agents do not have it. They are wasteful too: 95.3% to 97.2% of what an agent generates goes to re-deriving a plan the system already knows. We propose skill habit formation. An agent mines its own execution history for candidate skills, deterministic variants that compete against the incumbent rather than replacing it. A candidate declares the region of input space it claims, so the common case runs as a script and the rest falls through to reasoning. Four gates of ascending cost admit candidates; the central one tests a candidate's execution trace against a retained reference, within a tolerance measured from that reference's own run-to-run variability. On text-to-SQL, three of four reasoning arms reproduced their own output on 11 to 13 of 42 repeated questions and the fourth on 26 of 42, while a habit-formed variant reproduced on all 456 dispatches we repeated and was non-inferior to every arm it replaced (p<0.0001). It also used 14% to 56% fewer tokens, turning net positive after 7 to 53 reuses. We measured what this costs in accuracy. The guard admitted work it should have deferred on 2.6% of natural paraphrases and 26% of inputs near its boundary, and 11 of 13 such failures were invisible to the trace-conformance gate at any threshold. Deterministic errors repeat exactly: a bad habit is as reliable as a good one, and that is the price of the property that makes the system auditable. Separating routing from parameter extraction raised end-to-end accuracy from 0.888 to 0.952 at 43% of the cost.

Biomedical Entity Linking disambiguates mentions to entities in a knowledge base (KB), making it the cornerstone of information extraction pipelines. While embedding-based models are a popular approach for the task, they suffer from a key limitation. They compress mentions (and entities) into a single vector, forcing the model to average away crucial fine-grained differences. We present BELXTR, a novel embedding model based on the multi-vector (a.k.a. late interaction) architecture, which allows to leverage token-level matching information. BELXTR extends the original XTR model to biomedical entity linking by integrating an existing task-specific training objective and exploring active query expansion. Experiments across ten corpora and five KBs show that BELXTR improves upon current state-of-the-art in half of the corpora with an average improvement of 5pp recall@1. The largest gains are reported on the challenging cross-species gene disambiguation subtask, where BELXTR outperforms an LLM-powered retrieve-and-rerank pipeline and closely approaches a specialized rule-based system. Our results highlight multi-vector models as a practical alternative to hard-to-maintain rule-based systems or in scenarios where LLM-based reranking is too costly as in PubMed-scale mining. The code to reproduce our experiments can be found at: https://github.com/sg-wbi/belxtr.

"An agent that writes its conclusions into a store it later retrieves from closes a loop usually reported as one-way contamination. Taking the loop to the infinite-tenure limit against an append-only store gives a different picture: because writing never deletes, the reachable state space has a hard upper edge at (n-1)/n, so the outcome is a choice between two edges rather than a decay. At f_0 = 0.9 the interval between the two modes holds 3.6% of 220 runs where a uniform spread would put 20.6%, and is strictly empty on the first 15; the pooled mean describes 8.2% of the runs it summarises, the median 68.2%. Everything the model contributes is carried by one measured primitive with no fitted parameter, the copy function \gamma(\phi): on 36 Wikidata facts, sign(\hat{\gamma} - \gamma_{crit}), with \gamma_{crit} = 1/k at r = 0, w = 1, predicts the direction of drift on 353 of 360 real-fact runs (39 of 40 synthetic in the same batch). Scale does not rescue the store: pooled frontier capture is 0.850, with claude-sonnet-4.5 captured on 20 of 20 seeds against our registered prediction of <0.5. What the interval tests is distinguishability rather than count: on the real facts, multi-valued runs have 6.4x its occupancy of the rest. It survives at f_0 in {0.1, 0.3, 0.5}, capture peaks at f_0 = 0.5, and of four interventions with criteria frozen first, timing dominates fraction at matched budget while a consistency gate drives every model to 0.993. The resampling unit is the seed, at a design effect of 3.75 on a pooled level: under a 44-seed control the ordering supporting claim 4 collapses from Spearman +0.98 at three seeds to +0.31-0.80 at forty-four, while claim 2's ordering is exact there (+1.00, p = 0.017). All 87 graded rows are in Appendix W, 37 of them graded withdrawn, failed, self-correcting, undecidable or an acknowledged limit, against 50 that are not."

On-device assistants require function-calling models that map natural language to local system actions, but existing resources emphasize web APIs or narrow mobile-action catalogs. We extend FunctionGemma 270M-it to practical Android workflows by introducing MOBILEACTIONSEXTENDED, a synthetic, schema-validated dataset of ~9,500 conversations covering fifteen device-control categories, including messaging, phone calls, camera/screenshot, brightness control, device-status queries, flashlight control, and application management. We fine-tune the 270M model with TRL supervised fine-tuning under completion-only loss, producing an extended specialist and a combined model trained jointly with Google's MOBILEACTIONSGOOGLE. On MOBILEACTIONSEXTENDED, end-to-end accuracy improves from 29.3% for the base model and 17.2% for Google's Mobile-Actions variant to 76.5%. The combined model retains 76.5% on MOBILEACTIONSEXTENDED and reaches 82.3% on MOBILEACTIONSGOOGLE, down from the 90.3% of Google's Mobile-Actions specialist, representing an 8.0-percentage-point trade-off in return for doubling category coverage. We release the dataset, fine-tuned models, reproducible training/evaluation pipeline, and an Android demo, highlighting compact local function calling as a practical path towards low-latency and privacy-preserving mobile assistants.

Diffusion Large Language Models (dLLMs) have recently emerged as a promising alternative to conventional Auto-Regressive (AR) Large Language Models (LLMs). By leveraging bidirectional attention and parallel decoding, dLLMs enable more efficient generation. However, they require a carefully designed denoising scheduler at inference time (absent during training) whose choice significantly impacts generation quality. While confidence-based heuristic schedulers have shown strong empirical performance, they suffer from two critical failure modes: EOS Overflow and Proximal Bias. Through in-depth analysis of the Transformer's attention patterns, we reveal that these failures stem from certain positions assigning disproportionately high attention weights to invalid tokens (e.g., [MASK] and [EOS]), which produce misleading confidence signals. Building on this insight, empirical evidence shows that valid attention scores can provide complementary guidance to conventional confidence-based heuristics, yet no single metric consistently excels across all scenarios, implying that the optimal denoising trajectory is highly context-dependent. To address this problem, we propose a lightweight evolutionary heuristic scheduler optimized using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). Our scheduler dynamically integrates multiple heuristic features with a contextual mean-field embedding, while requiring only 393 trainable parameters. Evaluated on LLaDA and Dream across four reasoning and planning benchmarks, our method consistently outperforms strong baselines, including conventional heuristics, block auto-regressive methods, and recent State-Of-The-Art (SOTA) approaches. To the best of our knowledge, it represents the most parameter-efficient neural scheduler to date. Our code is available at https://github.com/RS2002/Evo-Denoise .

Long-horizon and multi-turn agents typically generate short actions and process long observations from tools and environments. This growing context demands efficient prefill, compact KV-cache storage, and accurate long-context retrieval. To meet these demands, we introduce HySparse2, a hybrid sparse attention architecture with two-level KV sharing. At the outer level, KV Bridging adopts a YOCO-style self-decoder and cross-decoder structure, but bridges only full-attention layers. The self-decoder uses hybrid sliding-window attention (SWA), while the cross-decoder uses hybrid sparse attention. The KV caches for full-attention layers in the cross-decoder are generated from the hidden states of full-attention layers in the self-decoder. At the inner level, HySparse2 retains HySparse's core KV Reuse design with two refinements. First, it replaces block-level sparsity with token-level sparsity for finer long-context retrieval. Second, it removes the separate SWA branch from sparse layers and instead forces a sliding window of recent tokens into the sparse selection. This two-level KV sharing allows all cross-decoder KV caches to be constructed from self-decoder hidden states. Prefill can therefore exit after the self-decoder, skipping all cross-decoder layers. On an 80B-A3B MoE model, HySparse2 outperforms HySparse and Hybrid SWA on long-context retrieval and multi-turn agentic tasks, while substantially reducing prefill computation and KV-cache storage.

We introduce Knowledge Pull Requests (KPRs), a framework for continual document authoring that makes each change interpretable. Documents require ongoing revision as new knowledge surfaces from other sources, languages, or times, but existing approaches either edit with no account of what knowledge changed or regenerate from scratch. A KPR integrates new knowledge into a document by extracting claims, filtering and routing them to sections, and flagging conflicts with existing content, producing a ChangeLog that separates what knowledge changes (claim proposal) from how the text changes (document diff). We evaluate KPRs on revising Wikipedia across languages and updating query-driven reports on RAGTIME. KPRs integrate more information and better preserve existing content than rewriting from sources or regenerating from scratch, while adding the most information per token generated. A KPR-revised article also grounds question answering better than a frontier model with search, which does not surface knowledge documented only in other languages.

When an agent handling a payment exception loses money, the agent-step attribution methods this paper compares against will name one of its actions. They will do so even when a settlement message was dropped and the agent never had a chance: they intervene on agent actions and do not expose infrastructure faults as intervenable variables, so every dollar they explain is charged to a decision. We take a benchmark whose fault process is explicit and replayable, decompose each episode's realised delivery schedule into named, individually repairable messages, and intervene on both the agent's choices and the infrastructure's. A telescoping identity splits any policy's loss exactly three ways: an infrastructure effect, a policy differential against the best implementable policy, and a reference-policy residual. Two of the three can be negative, so none is a share; Shapley then divides the first into signed allocations over individual messages. One result is structural and needs no corpus: an agent-only baseline identifies no infrastructure cause, because its model contains no variable that could name one. What 545 planted episodes across 3 policies measure is the size of that consequence. It misfiles 100% of infrastructure episodes and charges $114,383.40 to the agent. Repairing what it names recovers 0.0% of the available loss; repairing a minimal sufficient set recovers 100.0%. Scoring messages one at a time is not merely imprecise: 27.8% (95% CI: 23.3--32.3%) of episodes do not decompose additively. We evaluate deterministic programmatic policies rather than language-model agents, which is what makes replay exact and which limits external validity to stochastic agents. The prevalence figures are properties of this generator, not field rates.

Structured multi-agent workflows exchange intermediate messages whose content and form can reveal private state even when the final output is safe. We identify selection-channel leakage: after authorization fixes what may be released, a private-state-aware choice among semantically valid realizations creates an additional inference channel. We introduce the selection-invariant communication compiler(SICC), which constrains this post-authorization representation kernel rather than prescribing templates. Any deterministic or independently public-randomized generator satisfying the invariant is valid; requirement-indexed canonical forms are one auditable implementation. We prove a compositional communication-layer guarantee: authorization, public-only form generation, and a dependency-safe utility gate make the emitted transcript reveal no information beyond the complete authorized view. Private-state-aware selection remains vulnerable after surface-disjoint and length-matched controls. Across 132 AgentLeak communication replays and 100 executable LangGraph tasks, deterministic SICC retains complete protocol utility without a positive excess-gain signal; independent public randomization preserves the same result in AgentLeak and 480 controlled cases.

Children increasingly interact with AI chatbots, making trust calibration essential to AI literacy. Prior research has examined children's trust in AI mainly as users evaluating systems built by others, rather than as designers of their own chatbots. We developed a chatbot-building environment with adjustable trust-relevant traits (e.g., confidence, transparency, formality, assertiveness), rules, and persona. We conducted mixed-methods study with 115 learners (ages 8-18) who made 119 chatbots. We examined how children configured their chatbots, reasoned about trustworthiness, and how closely chatbot behavior aligned with their designs. Younger students (age 10-13) set significantly higher confidence than older students (age 14-18), and some deliberately built chatbots that gave wrong answers on purpose, yet still called them trustworthy, arguing that a chatbot does what it was built to do. Younger students equated trust with purpose-fulfillment, while older students linked it to transparent, calibrated design. Students also calibrated academic chatbots to be more transparent and formal than hobby chatbots. We identify seven design dimensions describing what children believe makes a chatbot trustworthy, and discuss implications for AI literacy tools.

Evaluation of large language models (LLMs) for safety, security, and privacy (SSP) relies heavily on static benchmarks, which suffer from score saturation, data contamination, and aggregation artifacts, and fail to capture sensitivity to linguistic variation. As a result, models that perform well on fixed test sets often fail under semantically equivalent rephrasings. We introduce SSP-Bench, a dynamic benchmarking framework that generates evaluation instances on demand while preserving domain consistency. The framework ensures label validity through externally grounded sources, enforces scope via service-specific validation, and calibrates difficulty using a multi-model steering panel. Benchmark construction is formulated as a multi-objective optimization problem over difficulty, separability, novelty, and diversity. Across 24 models and four SSP services, SSP-Bench reveals systematic failures of static evaluation, including near-zero correlation in safety rankings due to construct mixing, strong safety--over-refusal coupling, and hidden within-family regressions. These results show that static benchmarks can misrepresent model behavior, motivating dynamic, deployment-relevant evaluation.

Large language model (LLM) inference is constrained by the quadratic scaling of self-attention and the linear scaling of the KV cache, increasing latency, energy consumption, and GPU memory demand as context length scales. Existing soft-compression methods either lack query-guided memory selection at inference time, train without answer-targeted supervision, or couple compression tightly to a specific decoder architecture. We propose a Context-to-Answer-Aligned Memory Compression (CMC) framework, which compresses long input contexts into compact Context Memory Embeddings (CMEs) aligned to any frozen decoder's embedding space, reducing inference costs without modifying decoder weights. CMC introduces a two-tier KV cache that combines question-guided CME selection with a local context window, and trains the compressor with answer-targeted distillation from a frozen LLM. Experiments across nine encoder-decoder combinations and four QA benchmarks show that CMC consistently outperforms the baseline, achieving up to 7.3 EM and 4.0 F1 point gains on SQuAD, while reducing inference time and energy consumption by up to 20% and peak reserved GPU memory by up to 50% at 3,000 generation tokens. Ablation studies confirm that each architectural component and training objective contributes to the performance.

Text anomaly detection, the task of identifying text instances that deviate from normal language patterns, is crucial for language-driven applications. However, most existing methods can only perform document-level anomaly detection, making it hard to locate harmful phrases or support targeted prevention. Recently, there has been an emerging trend toward token-level text anomaly detection, which aims to address the above limitation by identifying anomalous words or fragments within a document. Nevertheless, one representative method mainly relies on representation-space distance measurement, neglecting the complementary roles of different anomaly cues in capturing diverse abnormal patterns. To bridge the gaps, we propose a Dual-evidence framework with adaptive Fusion and Aggregation (DiFA) for token-level anomaly detection. DiFA derives anomaly scores from form-structural and semantic views to capture visible structural abnormality and contextual inconsistency, respectively, thereby providing complementary evidence for identifying diverse anomalies. To combine these two scores with varying numerical scales, DiFA incorporates a calibration and fusion mechanism to adaptively balance the two views. Moreover, to obtain a discriminative document-level score, a multivariate aggregation method is designed to summarize token-level anomaly scores from multiple perspectives, preventing rare anomalous tokens from being diluted. Extensive experiments across various text anomaly detection benchmarks demonstrate that DiFA consistently achieves top performance while maintaining strong efficiency, robustness, and interpretability. The code and scripts are available at: https://github.com/qyy11-com/DiFA.

Entropy over chain-of-thought tokens decides which tokens receive the policy gradient, which get pruned, and whether a run has collapsed, yet each such statistic reads a next-token distribution mixing three choices: whether to emit connective scaffolding, which connective, and what the substantive continuation should be. Designating a scaffold vocabulary subset separates the three, exactly, for entropy, Kullback--Leibler divergence, and the first-order entropy velocity of a softmax policy. We prove the raw and content conventions disagree about which position is the larger fork on an explicit open region, and bound answer diversity by the content channel plus a leakage term a measured witness certifies. Across twenty-three configurations the scaffold side carries up to 41% of the raw high-entropy set; on a matched-tokenizer ladder, coupling changes only at the math-corpus step while the scaffold's entropy share keeps growing through distillation; a closed-form forecast from one channel correlation tracks selection retention over a 54-point range to five points, unfitted. On compression, the content convention beats raw surprisal in every cell; an answer-leakage audit then corrects our own headline control: re-fed chains earn a quarter to a half of their accuracy from restated answers, and once stripped, no token scorer beats a random contiguous block.