Research map / Large language models / Reasoning and chain of thought

STRIDE: Learnable Stepwise Language Feedback for LLM Reasoning

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: RL with verifiable rewards for reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes STRIDE, which co-trains a generator and a generative verifier with outcome-only rewards and uses the verifier's stepwise language critiques to redirect failed reasoning trajectories at intermediate steps.

Problem

Scalar process rewards from PRMs create an information bottleneck and step-level annotations are costly, while language-critique methods rely on frozen or external critics and lack learnable, sustained policy improvement for LLM reasoning.

Approach

Interleaves three phases (generator GRPO, verifier GRPO with outcome reward, and guided trajectory redirection from the first point of failure using verifier language critiques), with advantages computed only from outcome correctness of redirected samples, so language feedback replaces scalar step rewards.

Key results

• STRIDE-Qwen-7B reaches 84.6 on MATH-500 and 51.1 average on math benchmarks versus 37.0 for GRPO and 46.1 for TANGO on the Qwen2.5 series (from the table).
• On the Llama-3.1-8B series, STRIDE averages 36.0 on math benchmarks versus 32.3 for GRPO and 33.5 for TANGO.
• Compute-matched GRPO variants (about 40 GPU-hours) reach only 74.4 and 75.2 on MATH-500, versus 84.6 for STRIDE.

Surprising findings

• Even the weakest frozen verifier (0 steps) variant slightly outperforms vanilla GRPO (75.2 vs 74.6 on MATH-500).
• Verifier is initialized from the base model without SFT, yet co-training bootstraps useful verification.

Limitations

• Added training complexity: two models and an interleaved three-phase schedule, roughly 25% more GPU hours than GRPO.
• Verifier step-level localization is imperfect and not explicitly regularized.
• Evaluated only on verifiable tasks; extension to open-ended tasks would need LLM-as-judge rewards with added variance.

Experimental setup

Datasets: Eurus-2-SFT-Data Eurus-2-RL-Data MATH-500 AIME 2024 AIME 2025 AMC 2023 OlympiadBench BoardgameQA CRUXEval StrategyQA TableBench

Benchmarks: MATH-500 AIME 2024 AIME 2025 AMC 2023 OlympiadBench BoardgameQA CRUXEval StrategyQA TableBench

Metrics: Pass@1 Correction Success Rate verifier F1 error localization accuracy critique helpfulness

Baselines: GRPO TANGO Llama-3.1-8B-Instruct OpenMath2-Llama3.1-8B Qwen2.5-7B-Instruct Qwen2.5-Math-7B-Instruct rStar-Math-7B Eurus-2-7B-PRIME

Builds on: GRPO TANGO PRIME Eurus-2

Closest papers by meaning

• Advancing LLM Reasoning with Natural Language and Numerical Feedback (ICML 2026 · similarity 0.92)
• Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.91)
• $\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving (ICLR 2026 · similarity 0.91)
• Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · similarity 0.90)
• Reinforce LLM Reasoning through Multi-Agent Reflection (ICML 2025 · similarity 0.90)
• Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (ICML 2026 · similarity 0.90)