Quantization, distillation, pruning and serving work aimed at the same accuracy for less memory, latency and money.
arXiv:2608.30439v1 Announce Type: cross Abstract: Inference with transformer-based large language models (LLMs) is often limited by the memory-bound KV cache and quadratic attention cost. State-space...
arXiv:2608.29018v1 Announce Type: cross Abstract: We determine the sharp restricted isometry threshold for recovery at global minima of the rank-restricted matrix LASSO. For target rank $r_{\star}$,...
arXiv:2605.28042v2 Announce Type: replace-cross Abstract: Modern large language models (LLMs) achieve state-of-the-art machine translation performance, but they do so as broad generalists largely tra...
arXiv:2605.06387v4 Announce Type: replace-cross Abstract: On-policy distillation (OPD) trains a student on its own trajectories with token-level teacher feedback and often outperforms off-policy dist...
WiSP (Working‑Set Paging) is a routing‑aware expert pager that allows Mixture‑of‑Experts models to run on GPUs that cannot hold the entire expert pool by paging experts in and out of VRAM while preserving byte‑identical outputs. On a 24 GiB RTX 3090, WiSP doubles decode throughput compared to static offload when the model does not fit, and its companion policy MV‑WSA allocates VRAM between resident experts and KV cache based on marginal latency benefit, reducing end‑to‑end time by up to 1.19× without altering model outputs.
arXiv:2608.29088v1 Announce Type: new Abstract: Multimodal question answering remains sensitive to noisy, incomplete, and weakly grounded evidence. Long unstructured contexts can introduce redundancy...
arXiv:2608.29099v1 Announce Type: new Abstract: Temperature scaling is a core component of knowledge distillation, yet its role and effect are still not fully understood. Transformed Teacher Matching...
The paper investigates on‑policy distillation (OPD), showing that teacher supervision during OPD contains significant noise that grows with teacher size, yet the student policy remains largely unaffected by this noise. It finds that OPD’s gains stem mainly from suppressing low‑log‑probability tokens, a process that can be replicated without a teacher. Building on this insight, the authors propose On‑Policy Self‑Adaptation (OPSA), a supervision‑free method that uses entropy‑adaptive negative advantages to improve performance on several benchmarks, outperforming both the base model and OPD.
arXiv:2608.28725v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly used as graders, verifiers, and process auditors, but most mathematical evaluations still emphasize final...
The paper introduces TASPO, a method that transforms privileged information (PI) into outcome‑grounded action credit for language‑model agents. TASPO constructs decision‑applicable PI from verified successful experience, aggregates PI‑induced likelihood shifts at the executable‑action level, and converts relative action support into positive, bounded, mean‑preserving weights on the original trajectory advantage. Experiments on three agentic benchmarks show TASPO improves over GRPO by 10.6% and generalizes better to unseen tasks, while reducing supervision mismatch and stabilizing policy optimization.
arXiv:2601.10015v2 Announce Type: replace Abstract: Personalized Federated Learning (PFL) excels at tailoring client-specific models, which is particularly critical for decentralized and heterogeneou...
arXiv:2602.08136v2 Announce Type: replace-cross Abstract: Vision-Language Models (VLMs) are now a core part of modern AI. Recent work proposed several visual jailbreak attacks using single/ holistic...
arXiv:2604.13287v2 Announce Type: replace Abstract: Weight pruning is a common technique for compressing large neural networks. We focus on the challenging post-training one-shot setting, where a pre...
The paper introduces CHAP, a personalized generative retrieval framework that aligns query semantics with item representations through a hierarchical semantic alignment module and models user behavior using both discrete Semantic IDs and continuous representations. It also proposes a Residual Cascading Generation mechanism to reduce inference latency by limiting the Transformer decoder to a single pass. Experiments on multiple datasets and online A/B tests show that CHAP outperforms existing methods, demonstrating its practical value.
The paper introduces a framework that combines world models, which generate concrete visual rollouts of possible futures, with multimodal large language models (MLLMs) that perform abstract reasoning. It proposes a controlled concrete reasoning approach and a new training method called Privileged‑Future On‑Policy Self‑Distillation (PF‑OPSD), which uses ground‑truth future videos as privileged teacher context during training while the student model never sees true futures at test time. Experiments on two human‑verified benchmarks, VRQABench and OpenWorldQA, show that PF‑OPSD improves performance by about 10–11% over baselines and enhances robustness to noisy or conflicting rollouts.
arXiv:2510.22876v4 Announce Type: replace-cross Abstract: Inference optimizations are routinely evaluated by throughput alone, without verifying output correctness. We conduct a forensic analysis of...
arXiv:2608.31102v1 Announce Type: cross Abstract: Industrial post-training is a brownfield regime. Teams inherit a deployed checkpoint and must land targeted improvements under fixed compute and mixt...
The paper introduces a GPU‑optimized retrieval framework for LinkedIn’s semantic search, partitioning embeddings into eight category‑supervised segments and applying a min/median aggregation rule aligned with the existing relevance policy. A lightweight Stage‑1 scorer generates high‑recall candidates, while a two‑stage GPU architecture—FP8 coarse ranking followed by FP16 re‑ranking—boosts throughput and recall, achieving 99.6‑99.8% of full‑FP16 recall at over 500 QPS per shard. In A/B testing, the system raises exploratory‑query Precision@10 from 63.7% to 79.0% and navigational Precision@1 from 65.5% to 74.7%, with human evaluation confirming the improvement.
The paper presents a target‑centric survey of Quantization‑Aware Training (QAT), a technique that simulates quantization during model training to produce low‑bit models with accuracy comparable to full‑precision ones. It systematically reviews existing QAT methods using a target‑centric taxonomy, highlighting differences in error characteristics, numerical formats, and strategy transferability across targets. The survey also summarizes QAT evaluation paradigms, discusses optimization and deployment challenges, and outlines potential future research directions.
LiteSearch‑VL demonstrates that distilling released agent trajectories into small vision‑language backbones can transfer the agent’s behavioral contract, enabling a 2B model to produce usable answers in 28.4% of cases on multimodal benchmarks. The approach uses parameter‑efficient LoRA adapters and synthetic step‑level preferences derived from GPT‑5 hard negatives to refine tool use and query quality. While synthetic preference learning and tool distillation provide incremental improvements, the main bottleneck identified is answer verification rather than search depth.