Extrapolative Weight Averaging Reveals Correctness-Efficiency Frontiers in Code RL
This paper investigates whether extrapolative weight averaging of RL-trained checkpoints can extend Pareto frontiers between competing objectives (correctness vs. computational efficiency) without additional training. Starting from a shared initialization, the authors train checkpoints under nested unit-test coverage regimes for competitive programming tasks, revealing a correctness-efficiency frontier where higher-coverage rewards reduce optimization failures but increase correctness failures. Extrapolation beyond trained endpoints produces complementary policies that, when ensembled, improve pass@250 on LCB/hard by 3.3% over the best single checkpoint at matched sample budget. Results hold across 7B and 32B model scales and three inference settings: pure reasoning, tool use, and agentic coding.
Related guides (4)
Agent and Tool EcosystemTopic guide
Agent and Tool Ecosystem: How the Infrastructure Layer Around LLMs Is Consolidating
Related events (8)
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.
RELEX: Extrapolating LLM RLVR Training via Rank-1 Parameter Trajectories
This paper demonstrates that RLVR weight update trajectories are extremely low-rank and near-linearly predictable, with a rank-1 approximation capturing most downstream performance gains. The authors propose RELEX, a compute-efficient method that observes a short training window, estimates the rank-1 subspace, and extrapolates future checkpoints via linear regression—requiring no additional training. Evaluated on Qwen2.5-Math-1.5B, Qwen3-4B-Base, and Qwen3-8B-Base, RELEX matches or exceeds full RLVR performance using as few as 15% of training steps, and can extrapolate up to 10–20× beyond the observed prefix. The authors attribute the method's effectiveness to a denoising effect from rank-1 projection that discards stochastic optimization noise.
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.
POW3R: Policy-Aware Rubric Rewards for More Efficient RLVR Training
This paper identifies a failure mode in rubric-based reinforcement learning with verifiable rewards (RLVR): static aggregation of criterion weights conflates human-assigned importance with current optimization utility, causing many criteria to be either already saturated or unreachable. The authors introduce POW3R, a framework that dynamically reweights criterion-level rewards during training using rollout-level contrast to emphasize criteria that currently differentiate policy outputs. Across three base policies and two datasets (multimodal and text-only), POW3R wins 24 of 30 comparisons on rubric reward and strict completion metrics, and reaches equivalent performance in 2.5–4× fewer training steps than vanilla GRPO with rubric rewards.
Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning
This paper introduces Equilibrium Reasoners (EqR), a framework that formalizes test-time compute scaling through learned task-conditioned attractors in latent space, where stable fixed points correspond to valid solutions. EqR scales along two axes—depth (more iterations) and breadth (aggregating stochastic trajectories)—without requiring external verifiers or task-specific priors. On Sudoku-Extreme, unrolling up to 40,000 equivalent layers boosts accuracy from 2.6% (feedforward baseline) to over 99%. The work provides a mechanistic lens for understanding why iterative latent models generalize beyond memorized patterns.
AXPO: Agent Explorative Policy Optimization Addresses Thinking-Acting Gap in Multimodal Agentic Reasoning
This paper identifies a structural asymmetry in agentic reasoning called the 'Thinking-Acting Gap,' where tool use is attempted in only ~30% of rollouts under standard RL training (GRPO), and all-wrong tool-using subgroups suppress learning signals. The authors propose AXPO (Agent eXplorative Policy Optimization), which fixes the thinking prefix and resamples tool calls for all-wrong subgroups, combined with uncertainty-based prefix selection. Evaluated across nine multimodal benchmarks on Qwen3-VL-Thinking at multiple scales, SFT+AXPO outperforms SFT+GRPO by +1.8pp on both Pass@1 and Pass@4 at 8B, with the 8B model surpassing the 32B baseline on Pass@4 using 4× fewer parameters.
RePro: Retrospective Progress-Aware Self-Refinement for LLM Agent Training
Researchers introduce RePro (Retrospective Progress-Aware Training), a framework addressing the gap between step-wise RL optimization and metacognitive task-progress awareness in LLM agents. The approach uses a forward-then-reflect rollout paradigm where agents execute actions online and then retrospectively assess step-wise progress given the completed trajectory and known outcome. Evaluated on WebShop, ALFWorld, and Sokoban, RePro achieves up to 12% absolute success rate gains over baseline Qwen-family models without requiring continuous external supervision.
Vector Policy Optimization: Training for Diversity Improves Test-Time Search
Vector Policy Optimization (VPO) is a new RL post-training algorithm for LLMs that replaces the scalar reward paradigm with vector-valued rewards, explicitly training models to produce diverse solution sets that specialize across different reward trade-offs. VPO is designed as a near-drop-in replacement for the GRPO advantage estimator and targets inference-scaling search procedures like AlphaEvolve. Across four tasks, VPO matches or outperforms scalar RL baselines on pass@k and best@k metrics, with advantages growing as search budget increases, and unlocks evolutionary search problems that GRPO-trained models cannot solve. The paper argues that diversity-optimized post-training may need to become the default as inference-time search becomes standard.