Andrey Goncharov, Daniil Vyazhev, Simon Epanov et al. · cs.LG
Compression 97%
Abstract
Distillation of reasoning chains from Large Language Models (LLMs) into their smaller siblings is a standard training approach. Frequently, researchers do not have full access to a teacher model and its internal representations, which limits the resulting performance. We propose a recursive offline distillation method combining two components: entropy-based sampling for selective distillation and answer mixing. Specifically, for three open models (3-4B), at each epoch, we sample reasoning chains weighted by entropy difference with an auxiliary medium model (70-72B) and, next, supplement thinking traces with single-token ground truth answers. Our approach consistently outperforms naive distillation with comparable data budgets ($57.15 \pm 1.01$ vs 51.61 on average) and offers a much steeper learning curve. Additionally, we present a comparative analysis of methods to generate and normalize distillation traces.
Sunghyun Wee, Geunjae Choi, Hyeonjin Kim et al. · cs.LG, cs.AI
Compression 97%Inference speed 94%
Abstract
Large Language Models (LLMs) demand efficient inference, where the multiply-accumulate (MAC) operations in linear projections dominate compute and energy cost. Post-training quantization (PTQ) reduces this cost by mapping weights and activations to low-bit integers. Power-of-Two (PoT) weight quantization promises efficient multiplier-free LLM inference via shift-add logic, but suffers significant accuracy degradation at low bitwidths due to its coarse base-2 quantization grid. We introduce Fractional PoT (FPoT), a base-$\sqrt{2}$ (half-octave) fractional PoT grid that achieves practical 4-bit weight quantization while preserving a fully integer shift-add datapath. The method combines (i) the base-$\sqrt{2}$ weight grid with uniform integer activations, (ii) a calibration-time grid-alignment refinement absorbed into the static per-channel scale, and (iii) a lightweight 1-adder $\sqrt{2}$ approximation whose error is absorbed during weight calibration via approximate-grid quantization, yielding a processing element (PE) with significantly fewer gates than a 4-bit integer multiplier. On Llama-3-8B with 4-bit weight and activation quantization, FPoT improves zero-shot commonsense reasoning performance by 8.72% relative to the PoT baseline, while maintaining performance close to the uniform integer baseline, with only a negligible 0.4%p degradation. These results hold across various configurations spanning seven models and three bit-widths. Furthermore, FPoT can be integrated with state-of-the-art LLM PTQ methods, demonstrating framework-agnostic applicability.
Bingxi Hou, Guochao Jiang, Guofeng Quan et al. · cs.LG, cs.CL
Compression 97%
Abstract
Policy-shift distillation reuses the difference between a teacher's log probabilities before and after reinforcement learning (RL) as supervision on student-generated trajectories. However, this endpoint policy shift cannot reveal whether local action preferences evolved consistently during teacher training. The same endpoint shift can result from early and late shifts that reinforce or partially cancel each other. We investigate whether agreement between early and late teacher policy shifts can guide selective transfer. We introduce Waypoint-guided On-Policy Distillation (Waypoint-OPD), which uses an intermediate teacher checkpoint from the same RL run to compare early and late shifts in relative action preferences at student-visited prefixes. We gate the endpoint shift according to the directional agreement between the early and late shifts, suppressing it when they point in opposing directions. Compared with endpoint-only transfer, our method requires only one additional frozen-teacher scoring pass. Using RL-trained Qwen3-4B teachers, we evaluate our method with Qwen3-8B, Qwen3-14B, and Qwen3-30B-A3B as student models in mathematical and code reasoning. In all six domain--student settings, Waypoint-OPD achieves the highest domain-average score among the compared distillation methods. Ablations with the Qwen3-8B student show that removing or shuffling gates reduces mathematical reasoning performance. These results support the use of directional agreement between early and late teacher policy shifts to weight the endpoint shift during student training.
Zhengxin Zhang, Maosen Tang, Ning Wang et al. · cs.LG, cs.AI
Compression 96%Inference speed 58%
Abstract
On-policy distillation (OPD) effectively transfers capabilities from a teacher to a student through feedback on student-generated rollouts. However, generating complete student rollouts and computing teacher feedback for every token make OPD computationally expensive. We introduce Stochastic Truncated On-Policy Distillation (STOPD), which accelerates OPD by randomly truncating student rollouts and correcting the resulting sampling bias with blockwise inverse-probability weighting (IPW). At each training step, we sample a stopping position for each rollout before generation and let the student generate until that position or natural termination. We obtain teacher feedback on every generated token and use blockwise IPW to compensate for the lower probability of retaining later tokens while controlling the weights. We prove that this weighting preserves Vanilla OPD's gradient sum in expectation and derive one-step update-error bounds for SGD, Adam, and Muon. We evaluate STOPD across multiple teacher-student model pairs, task domains, and optimizers. STOPD reduces generated tokens by 25-48% while maintaining accuracy comparable to Vanilla OPD, achieving 1.3-2.2x training speedups.
Vlad Alexandru S\u{a}nduleac, Shahariya Rasheed Rayamarakkarveettil Puthanpurayil, Noman Khan et al. · cs.LG
Compression 96%
Abstract
Grid foundation models (GridFMs) are graph transformers pretrained on a masked power-flow objective, and evaluate power flow orders of magnitude faster than an iterative solver. Their size, however, targets data-centre inference, whereas low-inertia operation increasingly asks for inference at the edge, next to the measurement, at batch size one and within a single synchrophasor (PMU) reporting frame. We distil a 20.1M-parameter, 12-layer GridFM teacher into a 0.62M-parameter, 6-layer student, a $32.5\times$ parameter reduction at half the message-passing depth, using an objective that combines the masked power-flow loss with output, physics-residual, latent-state and attention transfer. The physics-residual term, which matches the teacher's power-residual profile, has not previously been applied to typed power-grid graphs. We evaluate on a renewable variant of the IEEE 118-bus benchmark grid in which 26 of the conventional (synchronous) generators are replaced by inverter-based renewables with battery flexibility and standardised reactive-power limits (IEEE 1547), giving $97\,529$ physically valid (AC-OPF-feasible) operating scenarios, and we ablate all eight combinations of the transfer terms, reporting accuracy against ground truth and fidelity to the teacher separately. Distillation lowers student error on all four predicted quantities, by up to $48\%$ on active-power dispatch, and every distilled student is more accurate than its teacher on reactive-power dispatch. The gain is largely carried by the output matching term. The physics-residual term, in the layer-aggregated form used here, is consistently harmful, which we attribute to the spatial averaging it applies to a bus-level residual field. The resulting student is $32.5\times$ smaller in parameter count and has half the message-passing depth, making it a candidate for subsequent edge-device benchmarking.
On-policy distillation consolidates specialist models into one student at the price of rollouts, teacher passes and gradient steps for every teacher. We show that the consolidation does not need training: with forward passes only, the capabilities of separately tuned specialists can be aggregated into one model that is, on average, more accurate than either specialist alone. When specialists share a base model, distillation has an exact layer-wise solution: each linear layer of the fused model must reproduce every specialist's output on the inputs that specialist receives while reading its own answers, a least-squares problem whose solution needs only the second moments of those inputs. One forward pass over a few hundred rollouts provides them. We call this training-free on-policy distillation (\method). What the rollouts contain matters more than the solver: answers truncated in repetition loops are few among a mathematics specialist's rollouts but carry a large share of their tokens, and removing them lifts the fused model on AIME25 from task arithmetic's level to the specialist's at matched precision. On Olmo~3 7B, a few GPU-hours of forward computation per specialist yield a generalist that keeps most of each specialist's gain on each of seven mathematics and code benchmarks.
Shiyu Ji, Peng Ling, Chen Lin et al. · cs.LG, stat.ML
Compression 94%Inference speed 87%
Abstract
Multi-head latent attention (MLA) compresses the content of a key-value cache into a shared latent per token, making low-bit storage attractive for long-context inference. At two bits, errors in this representation perturb both attention matching and value readout, with effects that depend on error direction and reconstruction scale. We introduce RoLaQ (Rotation-calibrated Latent Quantization), an offline method for joint rotation and range calibration. RoLaQ refines a shared rotation within subspaces defined by the frozen key and value projections, and jointly learns quantization ranges whose thresholds and reconstruction gains are coupled to control the shared signal's response. Prediction matching and attention-output supervision guide calibration with model weights frozen. Across three MLA models and six benchmarks, RoLaQ achieves the highest average accuracy among the compared 2-bit methods both without and with token protection. On DeepSeek-V2-Lite, an SGLang implementation provides 3.96 times the BF16 KV capacity under the same memory budget and 2.51 times its decode throughput with 256K input tokens.
Residual context is key to unlocking the speed-quality Pareto frontier of diffusion language models (DLMs) by carrying forward predictive distributions across denoising steps. Yet turning this seemingly simple mechanism into an effective learning paradigm is far from simple: there is no clear recipe for how residual context should be learned. We introduce ORL, a framework for on-policy residual learning in DLMs. Algorithmically, we identify that residual learning is inherently on-policy. ORL learns from model-induced residual trajectories on the fly and introduces an entropy-contraction objective that sharpens predictions as the model revisits its own predictions. System-wise, exact residual supervision requires capturing, retaining, and reprocessing many trajectory-dependent states. ORL co-designs the algorithm with asynchronous residual capture and transport, specialized block-diffusion attention, and packed compute-balanced training, making it practical at scale. On a 4B DLM converted from Qwen3-4B, ORL improves tokens per forward (TPF) by up to 6.2x and, with our dlmrt engine, end-to-end throughput by approximately 6.9x at comparable or better quality. In single-request serving on one H100, this mode achieves a stronger speed-quality Pareto frontier than autoregressive models with EAGLE3 and DFlash speculative decoding under both strict and relaxed verification, running 1.74x faster than DFlash at matched accuracy. ORL also transfers to existing DLMs: on LLaDA-2.1-mini, 36 H100-hours and 4.2M self-generated tokens, about 10x less data than prior LLaDA recipes, yield up to 1.5x higher TPF and up to 3x higher throughput at comparable quality, and on DiffusionGemma, ORL raises TPF by 1.1-1.2x. With dlmrt and NVFP4 quantization, ORL models decode a single request at up to 5,046 tokens/s on one B200, 7,708 on eight B200s, and 2,788 on a consumer RTX 5090, with comparable quality.
We expand the exponential kernel of attention into a truncated Taylor series of $n$ terms and factorize each term into a tensor-power inner product, which collapses the causal history into exactly $n$ recurrent states, each updated by one rank-one outer product per token and independent of context length $N$. The recursion is lossless with respect to the truncated kernel ($2.2 \times 10^{-16}$ in float64), and the states are additive across chunks ($5.7 \times 10^{-14}$), so prefill admits a parallel scan. Measured exponents are $t \propto N^{2.28}$ for softmax against $N^{0.91}$ for states, with a crossover at $N \approx 728$ and $20.8\times$ at $N = 8192$; per-token decoding latency is constant at $7.3\,\mu$s against $N^{1.04}$ for softmax, giving $76.0\times$ at $N = 65536$. Accuracy improves exponentially in $n$ while cost grows combinatorially, since the $m$-th state is $\binom{d+m-1}{m}$ wide---a bound confirmed by SVD---so the optimum is second order, $n = 3$, at $6.5 \times 10^{-4}$ error and $9{,}945$ elements per head. Query and key normalization is a prerequisite: without it the argument exceeds the convergence radius and adding terms stops helping.
Reusing the attention KV cache across requests is a standard way to cut prefill cost in long-context LLM serving, and production systems increasingly back the per-engine cache with a remote tier shared by every engine. We study such a tier in a disaggregated deployment of GLM-5.3, an FP8 mixture-of-experts model with multi-head latent attention (MLA), on 8xB300 nodes running prefill context parallelism (PCP8). Enabling the remote tier (LMCache with a Valkey store) left median time-to-first-token unchanged but raised the 16k-token p99 from 3.96 s to 17.3 s. Replaying a production request stream on an isolated engine, we find that a remote hit took 0.88 s against 0.31 s to recompute the prompt, and that synchronous loading turned a hit-heavy stream into an unstable queue: p99 83.32 s with the GPUs 7% busy. A two-term cost model explains both: a hit pays off only if the tier delivers more than B* = sR (2.53 GB/s for this model and GPU), and a fixed per-hit engine cost (T0 ~ 0.33 s) means that hits on prompts shorter than about 17k tokens cannot beat recompute at any bandwidth. Guided by the model, we (i) shard the MLA KV store across PCP ranks, with a proof that the existing minimum-over-ranks lookup returns the longest cached prefix and a load protocol that can never desynchronize collectives; (ii) partition the remote tier by rank, one server instance per rank with no coordination, which raises GET bandwidth 6.3x to 10.76 GB/s; (iii) add asynchronous loading to the sharded store; and (iv) fix three client bugs that caused unbounded stalls. On the hit-heavy stream, with 98.5% of prompt tokens served from the remote tier, p99 falls from 83.32 s to 0.67 s. The sharded store raises prefill throughput by 14%, and by 34% together with an upstream kernel fix. All changes are submitted upstream to LMCache.
Aref Jafari, Vansh Bhatia, Matt Qin et al. · cs.LG
Inference speed 90%
Abstract
Speculative decoding pairs a cheap drafter with parallel verification against the target. At ultra-long context it stops paying, because a drafter that attends over the full prefix inherits the target's context-dependent cost, so the speedup it was meant to provide first erodes and then reverses. The standard response is to build a long-context drafter, by training or fine-tuning one, or by using a multi-token-prediction head shipped inside the target checkpoint. We argue this is unnecessary, because the drafters in use are target-conditioned, reading hidden states the target computed over the entire prefix, so distant context already reaches them without their own attention, and verification corrects what they miss. We therefore freeze released short-context drafters and run them far past their training length under two inference-time mechanisms, one bounding the attention cache they see, which fixes their cost, and one re-basing their rotary position indices into the trained range. No weight is touched and the only tunable saturates early, so one vLLM-side configuration spans 4k to two million tokens, costing at most 0.16x of speedup at short length. We prove when a common index shift leaves the drafter's computation exactly unchanged, and decompose acceptance into two measurable gaps with sharp bounds. Whatever speedup a frozen drafter delivers at short context, it retains out to two million tokens, across three target-drafter pairs and five corpora. The claim concerns the flatness rather than the multiple, and removing either mechanism eliminates it. The gain is conditional on the speculative cost per emitted token staying below the matched target-only cost.
Oguzhan Buyuksolak, Osman Erman Okman, Mehmet Gorkem Ulkar · cs.LG
Compression 90%
Abstract
Fixed-function CNN accelerators make edge inference fast and energy-efficient, but they execute only networks that obey a precise numerical contract. On the MAX7800x family, this contract requires power-of-two, per-tensor, symmetric scales, 8-bit activations and biases, integer-only arithmetic, and one statically configured shift per layer. We present ShiftQAT, the quantization-aware training (QAT) flow we developed for this family, which produces networks satisfying the contract by construction. It is the default training path of ai8x-training, the family's open-source training toolchain. ShiftQAT resolves residual, concatenation, and weight-sharing dependencies so that activation scales propagate statically and each layer's requantization reduces to one offline-configured shift. It also trains weights and biases on a joint power-of-two grid that fits the 8-bit bias budget. Among the compared methods, only ShiftQAT satisfies all of the contract's scaling constraints, including signed 8-bit activation storage. Even so, on ImageNet it matches the W8A8 accuracy drop of less constrained QAT methods and stays within 2.9 percentage points of FP32 at W4A8. On MAX78002 silicon, ShiftQAT-trained MobileNetV2 and ResNet-18 classify ImageNet images in 14.6 and 8.4 ms at 5.1 and 2.8 mJ per inference.