<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Paper Radar · demo</title><link>https://github.com/Eliot5566/JEV-Paper-Radar</link><description>Papers selected by Paper Radar</description><lastBuildDate>Fri, 13 Nov 2026 16:26:24 +0000</lastBuildDate><item><title>★ A Survey of Uncertainty Estimation in Large Language Models</title><link>https://example.org/demo/2609.99005</link><guid isPermaLink="false">arxiv:2609.99005</guid><pubDate>Sat, 14 Nov 2026 00:00:00 +0000</pubDate><category>cs.CL</category><author>Grace Mensah, Daniel Kowalski</author><description>&lt;p&gt;&lt;b&gt;95%&lt;/b&gt; · LLM calibration 95%&lt;/p&gt;&lt;p&gt;We review methods for uncertainty estimation in large language models, covering verbalized confidence, sampling consistency, and probe-based approaches. We organize the literature by the kind of prediction being scored and discuss open problems in calibration for long-form generation.&lt;/p&gt;</description></item><item><title>★ Rerank Less, Retrieve Better: Query-Adaptive Depth for Retrieval-Augmented Generation</title><link>https://example.org/demo/2609.99003</link><guid isPermaLink="false">arxiv:2609.99003</guid><pubDate>Sat, 14 Nov 2026 00:00:00 +0000</pubDate><category>cs.IR</category><category>cs.CL</category><author>Samir Haddad, Elena Petrova</author><description>&lt;p&gt;&lt;b&gt;95%&lt;/b&gt; · RAG &amp; retrieval 95%&lt;/p&gt;&lt;p&gt;Retrieval-augmented generation pipelines typically rerank a fixed number of passages. We show that retrieval quality for question answering depends strongly on query difficulty, and train a lightweight controller that chooses retrieval depth per query. The controller cuts reranker calls by 58% while improving answer accuracy on four question answering benchmarks.&lt;/p&gt;</description></item><item><title>★ TraceGrade: Evaluating LLM Agents on Long-Horizon Tool-Use Tasks with Step-Level Verification</title><link>https://example.org/demo/2609.99001</link><guid isPermaLink="false">arxiv:2609.99001</guid><pubDate>Sat, 14 Nov 2026 00:00:00 +0000</pubDate><category>cs.CL</category><category>cs.AI</category><author>Mira Okafor, Tomás Lindqvist, Hae-won Seo, Priya Raman</author><description>&lt;p&gt;&lt;b&gt;95%&lt;/b&gt; · Agent evaluation 95%&lt;/p&gt;&lt;p&gt;Benchmarks for LLM agents usually score only the final answer of a multi-step episode. We introduce TraceGrade, a benchmark of 1,240 multi-step tool-use tasks across web, shell, and spreadsheet environments in which every intermediate action is verified against an executable checker. Evaluating 14 LLM agents, we find that final-answer accuracy overstates step-level correctness by up to 31 points. Code and tasks are publicly released.&lt;/p&gt;</description></item><item><title>★ Calibrated Abstention for Language Model Question Answering via Conformal Uncertainty</title><link>https://example.org/demo/2609.99002</link><guid isPermaLink="false">arxiv:2609.99002</guid><pubDate>Sat, 14 Nov 2026 00:00:00 +0000</pubDate><category>cs.CL</category><category>cs.LG</category><author>Jonas Brandt, Aiko Tanaka, Lucía Ferrer</author><description>&lt;p&gt;&lt;b&gt;95%&lt;/b&gt; · LLM calibration 95%, RAG &amp; retrieval 35%&lt;/p&gt;&lt;p&gt;Language model predictions are often overconfident, which makes abstention unreliable. We propose a conformal procedure that turns token-level signals into calibrated uncertainty estimation for open-domain question answering. Across five datasets, our method reduces expected calibration error by 43% and lets the language model abstain with a guaranteed error rate.&lt;/p&gt;</description></item><item><title>★ Position: Evaluating Autonomous Agents Needs Cost-Aware Metrics</title><link>https://example.org/demo/2609.99014</link><guid isPermaLink="false">arxiv:2609.99014</guid><pubDate>Sat, 14 Nov 2026 00:00:00 +0000</pubDate><category>cs.AI</category><author>Michael Brooks, Anika Sharma</author><description>&lt;p&gt;&lt;b&gt;94%&lt;/b&gt; · Agent evaluation 94%&lt;/p&gt;&lt;p&gt;We argue that evaluating LLM agents by success rate alone hides the cost of multi-step tool-use. We propose reporting success per dollar and per minute and show how rankings on popular agent benchmarks change under cost-aware metrics.&lt;/p&gt;</description></item><item><title>★ Do LLM Agents Know When They Are Wrong? Self-Evaluation of Tool-Use Trajectories</title><link>https://example.org/demo/2609.99009</link><guid isPermaLink="false">arxiv:2609.99009</guid><pubDate>Sat, 14 Nov 2026 00:00:00 +0000</pubDate><category>cs.AI</category><category>cs.CL</category><author>Laura Chen, Pedro Alves, Yuki Mori</author><description>&lt;p&gt;&lt;b&gt;94%&lt;/b&gt; · Agent evaluation 94%, LLM calibration 68%&lt;/p&gt;&lt;p&gt;We study whether LLM agents can evaluate their own multi-step tool-use trajectories. Agents&amp;#x27; self-reported confidence is poorly calibrated: across three agent benchmarks, trajectories rated 90% likely to succeed succeed only 61% of the time. A simple uncertainty estimation probe on hidden states recovers much of the gap.&lt;/p&gt;</description></item><item><title>★ SciRetrieve: A Benchmark of Retrieval Quality for Scientific Question Answering</title><link>https://example.org/demo/2609.99013</link><guid isPermaLink="false">arxiv:2609.99013</guid><pubDate>Sat, 14 Nov 2026 00:00:00 +0000</pubDate><category>cs.IR</category><category>cs.CL</category><author>Oluwaseun Adeyemi, Clara Dubois</author><description>&lt;p&gt;&lt;b&gt;93%&lt;/b&gt; · RAG &amp; retrieval 93%&lt;/p&gt;&lt;p&gt;We release SciRetrieve, a benchmark of 9,800 expert-written scientific questions with passage-level relevance labels, to measure retrieval quality for retrieval-augmented question answering. Dense retrievers and rerankers still miss a supporting passage for 27% of questions.&lt;/p&gt;</description></item><item><title>★ 4-Bit KV Cache Compression for Long-Context LLM Inference</title><link>https://example.org/demo/2609.99010</link><guid isPermaLink="false">arxiv:2609.99010</guid><pubDate>Sat, 14 Nov 2026 00:00:00 +0000</pubDate><category>cs.LG</category><category>cs.CL</category><author>Arjun Mehta, Sofia Rossi</author><description>&lt;p&gt;&lt;b&gt;89%&lt;/b&gt; · Efficient inference 89%&lt;/p&gt;&lt;p&gt;Long-context LLM inference is dominated by KV cache memory. We introduce a 4-bit quantization scheme for the KV cache with per-channel outlier handling that keeps perplexity within 0.1 of full precision and enables 2.6x cheaper inference at 128k context.&lt;/p&gt;</description></item><item><title>★ SpecTree: Draft Trees for Faster Speculative Decoding in Large Language Models</title><link>https://example.org/demo/2609.99004</link><guid isPermaLink="false">arxiv:2609.99004</guid><pubDate>Sat, 14 Nov 2026 00:00:00 +0000</pubDate><category>cs.CL</category><category>cs.LG</category><author>Wei Zhang, Olivia Grant, Rafael Moreno, Chen Li</author><description>&lt;p&gt;&lt;b&gt;89%&lt;/b&gt; · Efficient inference 89%, LLM calibration 35%&lt;/p&gt;&lt;p&gt;Speculative decoding accelerates LLM inference by verifying draft tokens in parallel. We propose SpecTree, which drafts a tree of continuations and prunes it with a learned acceptance model. SpecTree delivers 3.1x faster decoding than autoregressive sampling with identical outputs, and composes with quantization for cheaper inference. Code is available.&lt;/p&gt;</description></item><item><title>Faster Decoding for Speech Language Models via Speculative Unit Prediction</title><link>https://example.org/demo/2609.99044</link><guid isPermaLink="false">arxiv:2609.99044</guid><pubDate>Sat, 14 Nov 2026 00:00:00 +0000</pubDate><category>eess.AS</category><category>cs.CL</category><author>Hyun-woo Park, Isabel Novak</author><description>&lt;p&gt;&lt;b&gt;68%&lt;/b&gt; · LLM calibration 68%, Efficient inference 48%, RAG &amp; retrieval 35%&lt;/p&gt;&lt;p&gt;Speech language models generate discrete units slowly. We adapt speculative decoding to unit sequences and obtain 2.4x faster inference with no loss in speech quality.&lt;/p&gt;</description></item></channel></rss>
