We've trained a model to achieve a new state-of-the-art in mathematical problem solving by rewarding each correct step of reasoning (“process supervision”) instead of simply rewarding the correct final answer (“outcome supervision”). In addition to boosting performance relative to outcome supervision, process supervision also has an important alignment benefit: it directly trains the model to produce a chain-of-thought that is endorsed by humans.
arXiv:2605. 02395v2 Announce Type: replace Abstract: Process reward models (PRMs) rely on high-quality process supervision data, yet existing construction methods often provide limited control over error location, error type, and trajectory consistency.
The paper introduces Reasoning State Propagation (RSP), a method that models each reasoning prefix with a binary validity state and learns transitions between successive states. RSP predicts break and repair probabilities to connect intermediate reasoning states to the final outcome, enabling outcome supervision to guide learning of earlier steps. Experiments on reasoning search, response selection, and reinforcement learning show RSP consistently outperforms existing Process Reward Models, achieving notable gains over Qwen2.5-Math-PRM.
arXiv:2609.36641v1 Announce Type: cross Abstract: Process reward models (PRMs) have become a key component for LLMs, as their step-level feedback supports both post-training and test-time reasoning....
The paper proposes a self‑improvement loop for reasoning models that leverages hindsight: when a problem is too hard, a supplied solution can help the model learn useful solution ideas. The method trains a single model to (1) predict solution ideas from problems alone, (2) reverse‑engineer ideas from problems and known solutions, and (3) solve problems using provided ideas. The loop alternates between reverse‑engineering ideas from problems with supplied solutions and using these ideas as additional supervision for joint training of all three capabilities, with a formal specification and an instantiation for interactive theorem proving in Lean.
arXiv:2603.17815v2 Announce Type: replace Abstract: Understanding and evaluating multi-step reasoning in LLMs at the level of individual steps remains a key challenge. Process reward models (PRMs) pr...
arXiv:2604.02341v2 Announce Type: replace-cross Abstract: Mathematical reasoning in large language models has improved substantially with reinforcement learning using verifiable rewards, where final...
SPEAR (Symbolic Process Evaluation and Alignment Reward) is a training‑free, plug‑and‑play reward method for on‑policy distillation in reinforcement learning. It converts natural‑language reasoning traces into domain‑adaptive symbolic milestones and uses the longest common subsequence to align student exploration with teacher milestones, producing a dense, order‑aware reward that enforces logical consistency without an external neural verifier. Experiments on math, science, and commonsense reasoning tasks show that SPEAR effectively bridges the reasoning gap between student and teacher models through sequence‑level distillation with efficient dense process rewards.
arXiv:2605. 03862v4 Announce Type: replace Abstract: Reinforcement learning with verifiable rewards has become a common way to improve explicit reasoning in large language models, but final-answer correctness alone does not reveal whether the reasoning trace is faithful, reliable, or useful to the model that consumes it.
arXiv:2509.13334v2 Announce Type: replace Abstract: Chain-of-thought (CoT) reasoning improves LLM accuracy on complex tasks, but CoTs are often unfaithful; that is, the model does not rely on the sta...
arXiv:2606. 31800v1 Announce Type: new Abstract: Despite recent progress, the reasoning capabilities of large multimodal language models (MLLMs) remain fundamentally constrained by static supervision, where fixed prompts, rules, or reward models provide non-adaptive guidance throughout training.
The paper introduces Prototype-Mediated Process Supervision (PMPS), a method that uses learnable reasoning prototypes to provide structural supervision for latent chain-of-thought embeddings, addressing representation collapse and uneven information distribution. PMPS aligns latent and explicit CoT embeddings in a shared prototype space via many-to-many soft assignment and employs a Progressive Sequential Alignment module to guide training from positional priors to adaptive matching. Experiments show PMPS reduces output token length to under 50% of explicit CoT on GSM8K-Aug and improves accuracy by 2.08% over SIM-CoT, even surpassing CoT-SFT on GPT-2 and achieving the highest accuracy among latent reasoning methods on larger models and harder tasks.
arXiv:2606. 13125v1 Announce Type: cross Abstract: Reinforcement learning has rapidly emerged as a key component in the training of reasoning and coding models, yet it remains poorly understood from a mechanistic perspective.