{"paper":{"title":"AIS: Adaptive Importance Sampling for Quantized RL","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Adaptive Importance Sampling corrects non-stationary bias from low-precision rollouts while keeping their speed gains in LLM RL.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"stat.ML","authors_text":"Jiajun Zhou, Lingchao Zheng, Ngai Wong, Wei Shao, Yuwei Fan","submitted_at":"2026-05-13T03:36:57Z","abstract_excerpt":"Reinforcement learning (RL) for large language models (LLMs) is dominated by the cost of rollout generation, which has motivated the use of low-precision rollouts (e.g., FP8) paired with a BF16 trainer to improve throughput and reduce memory pressure. This introduces a rollout-training mismatch that biases the policy gradient and can cause training to collapse outright on reasoning benchmarks. We show that the mismatch is non-stationary and acts as a double-edged sword: early in training it provides a stochastic exploration bonus, exposing the gradient to trajectories the trainer would otherwi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"AIS matches the BF16 baseline on most tasks while retaining the 1.5 to 2.76x rollout speedup of FP8 by combining weight reliability, divergence severity, and variance amplification into a per-batch mixing coefficient that interpolates between uncorrected and importance-weighted gradients.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The three real-time diagnostics can be combined into a single mixing coefficient that reliably preserves early-training exploration benefits while suppressing later destabilizing bias across different models, tasks, and training stages without introducing new instabilities.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"AIS adaptively corrects non-stationary policy gradient bias in quantized LLM RL, matching BF16 performance while retaining 1.5-2.76x FP8 rollout speedup.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Adaptive Importance Sampling corrects non-stationary bias from low-precision rollouts while keeping their speed gains in LLM RL.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2fd68b1d443977d73aaceb481c406a1a25ca12a27a183992f92b4c727281f2b5"},"source":{"id":"2605.13907","kind":"arxiv","version":1},"verdict":{"id":"df2c41d6-3fc1-4b2d-a308-ba890af62447","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T03:09:29.288537Z","strongest_claim":"AIS matches the BF16 baseline on most tasks while retaining the 1.5 to 2.76x rollout speedup of FP8 by combining weight reliability, divergence severity, and variance amplification into a per-batch mixing coefficient that interpolates between uncorrected and importance-weighted gradients.","one_line_summary":"AIS adaptively corrects non-stationary policy gradient bias in quantized LLM RL, matching BF16 performance while retaining 1.5-2.76x FP8 rollout speedup.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The three real-time diagnostics can be combined into a single mixing coefficient that reliably preserves early-training exploration benefits while suppressing later destabilizing bias across different models, tasks, and training stages without introducing new instabilities.","pith_extraction_headline":"Adaptive Importance Sampling corrects non-stationary bias from low-precision rollouts while keeping their speed gains in LLM RL."},"references":{"count":33,"sample":[{"doi":"","year":null,"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","ref_index":1,"cited_arxiv_id":"2110.14168","is_internal_anchor":true},{"doi":"","year":null,"title":"DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models","work_id":"07c85cc5-4086-4abc-823b-6d0f4ff784d0","ref_index":2,"cited_arxiv_id":"2512.02556","is_internal_anchor":true},{"doi":"","year":null,"title":"AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning","work_id":"60017c83-cbbd-4807-b8e4-f8149a6aeaf0","ref_index":3,"cited_arxiv_id":"2505.24298","is_internal_anchor":true},{"doi":"","year":null,"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","ref_index":4,"cited_arxiv_id":"2501.12948","is_internal_anchor":true},{"doi":"","year":null,"title":"OlympiadBench: A Challenging Benchmark for Promoting AGI with Olympiad-Level Bilingual Multimodal Scientific Problems","work_id":"19abed3b-0ff6-409b-aded-a50205319aa3","ref_index":5,"cited_arxiv_id":"2402.14008","is_internal_anchor":true}],"resolved_work":33,"snapshot_sha256":"1097be1a26b192962051ee3eb5e922b9174846614d9d282c3c1ddc711a0ca721","internal_anchors":18},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}