The paper proposes a method for selectively querying language‑model advice in reinforcement learning by predicting the value of potential responses and only querying when the expected benefit outweighs the cost. It introduces a certified, response‑contingent metareasoning framework that guarantees near‑optimal advice usage under certain assumptions, and demonstrates that a calibrated controller with Qwen2.5 advisors can improve task performance while drastically reducing the number of advice calls on the BabyAI benchmark.
arXiv:2606. 29280v1 Announce Type: cross Abstract: We identify intervention bias as a previously unquantified failure mode of zero-shot large-language-model (LLM) educational advisory agents: without task-specific training, they recommend action when a hindsight-optimal oracle policy mandates inaction.
arXiv:2609.38142v1 Announce Type: new Abstract: A small trainable advisor can steer a frozen language-model executor using natural-language advice. In addition to learning from task rewards, the advi...
arXiv:2607. 13389v1 Announce Type: new Abstract: Reinforcement Learning (RL) post-training is increasingly used to adapt foundation models for reasoning, planning, and feedback-driven robot-learning pipelines, but constrained post-training resources are often summarized by a single total FLOP budget.
arXiv:2607. 26253v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) is bottlenecked by rollout generation, yet many sampled prompts produce saturated groups (all responses correct or all incorrect) whose zero reward variance yields no policy-gradient signal.
arXiv:2607.28077v2 Announce Type: replace Abstract: Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models, but prompt groups with identica...
Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models, but prompt groups with identical rollout rewards consume generation budget without effective learning signals. Pre-rollout prompt selection can reduce this waste by screening prompts before rollout generation.
BCPPO is a new variant of Proximal Policy Optimization that uses Bachelier-inspired cost‑prediction networks to generate a smooth penalty based on disagreement among critics. The method keeps temporal‑difference learning unchanged, applies a saturation‑aware controller to manage cost penalties, and deploys only the policy network. Across extensive experiments, BCPPO outperforms comparators in achieving higher mean returns while maintaining lower or comparable CVaR in all tested tasks.
The paper introduces Circuit Reasoning Score (CRS), a data‑selection signal for reinforcement learning with verifiable rewards that uses attention‑head activity from a frozen base model to gauge reasoning engagement. CRS is computed in a single forward pass without reward labels or rollouts, and it shows that selecting problems with the lowest reasoning‑circuit engagement can outperform random selection on several medium‑difficulty benchmarks. However, the benefit depends on domain, model scale, and reward conditions, indicating that data selection in this setting is regime‑dependent rather than a fixed ranking of problem quality.
MInTRL (Minimal Intervention Reinforcement Learning) expands exploration in on-policy reinforcement learning by inserting sparse, local corrections into rollouts via a judge-intervention policy. These interventions replace erroneous suffixes and immediately return control to the main policy, allowing the agent to explore beyond its natural trajectory while maintaining on-policy data. The method uses a sequence-level advantage-regression objective, avoiding importance sampling, and demonstrates superior performance on math and code benchmarks compared to standard on-policy and off-policy baselines.
arXiv:2608. 00220v1 Announce Type: new Abstract: We show that on-policy reinforcement learning with verifiable rewards (RLVR) can improve the current objective while making successful behaviors for later objectives too rare to sample and reinforce.
Online Self-Weighted Fine‑Tuning (OSW‑FT) augments standard supervised fine‑tuning by adding online, trajectory‑level weighting: for each query the model estimates its current success rate from a small number of inference‑only rollouts and rescales the SFT loss accordingly. The method keeps the optimization direction anchored to the expert trajectory while adapting the update magnitude online, and it is shown to be unbiased for any finite rollout count with a convergence analysis. Across Qwen3 models from 0.6B to 4B, OSW‑FT consistently outperforms plain SFT on challenging benchmarks such as AIME, achieving a favorable compute‑performance trade‑off with only two online rollouts.
arXiv:2607. 27610v1 Announce Type: new Abstract: Reinforcement learning (RL) finetuning significantly enhances the reasoning capabilities of large language models (LLMs), yet its effectiveness critically depends on selecting prompts of appropriate difficulty for the current policy.