Paper diagnoses RL collapse in multi-step tool-use training and proposes supervisory signal fixes
A new arXiv preprint identifies a failure mode in reinforcement learning for LLM tool use: catastrophic collapse caused by probability spikes in control tokens that disrupt structured execution while leaving underlying tool-use capability intact. The authors systematically evaluate supervisory signals—including off-policy supervision, hint-based guidance, and erroneous example supervision—under synchronous and interleaved training schemes. Interleaving SFT with RL improves stability but degrades performance under out-of-distribution format and content evaluation. Code is released as Tool-RL-Box.
Related guides (2)
Related events (8)
Two is better than one: A Collapse-free Multi-Reward RLIF Training Framework
This paper proposes a multi-reward reinforcement learning from internal feedback (RLIF) framework that decomposes training signals into an answer-level reward via cluster voting and a completion-level reward via token-wise self-certainty. To address reward hacking and entropy collapse common in single-reward RLIF, the authors introduce GDPO-based normalization and KL-Cov regularization targeting low-entropy token distributions. Evaluated on mathematical reasoning and code-generation benchmarks, the method achieves stability and performance approaching supervised RLVR methods without requiring external ground-truth supervision. The work advances scalable unsupervised RL training for LLM reasoning.
Peak-Then-Collapse: RLVR Tool-Use Failures on Knowledge-Graph APIs
This paper investigates RLVR-based tool-use training (GRPO on Qwen2.5-7B-Instruct) on a minimal knowledge-graph API (Freebase over Complex WebQuestions) and documents a 'peak-then-collapse' pattern where tool-grounded answer rates rise then fall to zero within 50 steps, replicated across four seeds and seven reward designs. The authors identify a key structural difference between knowledge-graph APIs and other tool types (Python, web search, JSON): sparse, non-natural-language feedback signals (e.g., empty brackets '[]') prevent the model from recovering via pretraining-familiar error signals. A direct oracle ablation shows relation selection is not the bottleneck—95.4% of errors are retrieval-composition failures—and self-distillation reaches 40% EM at 7B, with capacity scaling to 14B yielding only marginal gains, suggesting an interface-bound ceiling.
ExpRL: RL-based mid-training using human QA data as reward scaffolds for LLM reasoning
ExpRL proposes an automated approach to LLM mid-training that replaces manually curated reasoning traces with large corpora of human-written QA data used as reward scaffolds rather than imitation targets. Reference solutions are hidden from the policy and used only to construct problem-specific grading rubrics, enabling dense process-level rewards that reinforce partial progress and intermediate reasoning steps. On challenging math reasoning benchmarks, ExpRL outperforms SFT, sparse-reward GRPO, and self-distillation as an RL initialization strategy, with additional mixed-domain experiments suggesting broader applicability.
STARE: Token-level advantage reweighting to prevent entropy collapse in GRPO-style RL training
Researchers introduce STARE, a method addressing policy entropy collapse in GRPO-style reinforcement learning from verifiable rewards (RLVR) for LLM post-training. Through first-order gradient analysis, they identify a token-level credit assignment mismatch and propose selectively reweighting advantages for entropy-critical tokens using batch-internal surprisal quantiles plus a closed-loop entropy gate. Evaluated across 1.5B–32B models on short/long chain-of-thought and multi-turn tool use tasks, STARE outperforms DAPO and other baselines by 4–8% on AIME24/25 while sustaining stable training over thousands of steps.
Mechanism-driven internal monitors detect LLM training instability thousands of steps before loss divergence
A new arXiv preprint proposes mechanism-driven monitoring signals derived from the functional roles of critical modules (low-precision flash attention, MoE routers) to detect training instability before it manifests in loss or gradient norms. The authors derive monitors such as spectral entropy of a QK bilinear decomposition and MoE router indicators, showing via fault-injection experiments that these signals trigger thousands of steps ahead of loss divergence. The work targets a high-cost failure mode in frontier LLM training where instability can persist undetected for thousands of steps on expensive accelerator fleets.
ACPO: Adaptive Clip Policy Optimization improves RLVR training for LLM reasoning
A new arXiv preprint provides theoretical analysis of Reinforcement Learning from Verifiable Rewards (RLVR) updates, identifying off-policy degree and gradient expectation as key factors governing update dynamics. The authors show that differences in gradient steps per rollout substantially affect importance sampling ratio distributions and which tokens dominate updates. Based on this analysis, they propose Adaptive Clip Policy Optimization (ACPO), which adjusts clipping boundaries per token group by empirical variance of importance sampling ratios, outperforming DAPO and CISPO baselines on 3B and 7B models across math, tabular QA, and logic benchmarks.
SafeCtrl-RL: Inference-Time Adaptive Behaviour Control for LLMs via RL-Driven Prompt Optimisation
SafeCtrl-RL is a framework for controlling LLM safety at inference time without retraining or modifying model parameters. It formulates dialogue generation as a sequential decision process where an RL agent dynamically selects prompt adjustment strategies based on contextual feedback, iteratively suppressing unsafe outputs. The authors frame this as 'inference-time behavioural unlearning' and report improvements in safety and response quality across multiple LLMs and unsafe dialogue scenarios, outperforming existing prompt-based optimisation baselines.
Signal Collapse and Reward Hacking in Checker-Guided RAG for Biomedical QA
This paper investigates why NLI-based claim checkers used as process rewards in RL-trained medical RAG agents succeed or fail during training. The authors find that a checker's output distribution during training—not its held-out accuracy—determines whether it provides useful gradient signal, with LLM log-probability scoring causing near-total signal collapse (97%+ neutral labels) while a calibrated MedNLI classifier avoids this. A key finding is that stronger checkers can trigger reward hacking cascades (ultra-short answers, search avoidance, language collapse), while moderate-signal local classifiers yield better final model quality (+12% BERTScore over zero-shot). The work frames these as boundary conditions for verifier-as-reward systems in RLVR pipelines.