Research map / Large language models / Reasoning and chain of thought
Patching Gaps In LLM Reasoning With Interventional Training
Matthew Y. R. Yang, Hao Bai, Ian Wu, Gene Yang, Amrith Setlur, Aviral Kumar
Read the paperCheck your idea against 137,217 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 interventional training to address RL stalling on hard problems where LLM policies never sample correct rollouts, by patching reasoning gaps (based only on the abstract).
Problem
RL training of LLMs stalls on hard problems because the policy rarely generates non-zero-reward rollouts, so it receives no learning signal; the abstract notes hard problems often remain unsolved due to repeated generation of incorrect reasoning steps.
Approach
Uses interventions on the reasoning process during training to inject correct steps and patch gaps in LLM reasoning so that hard problems yield non-zero-reward rollouts (details not stated in the abstract).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| condition rollouts on a self-proposed intervention instead of the original erroneous step (ablation) | Qwen3-4B-Instruct-2507 · Omni-MATH hard subset (334 problems) | accuracy: 0.0713% -> 1.56%; coverage 29/334 -> 80/334 | Improved |
| clone the prefix before the intervention during SFT (ablation) | InT SFT · 235 DeepScaleR problems | accuracy: 2.87% -> 7.71%; coverage 162/235 -> 202/235 | Improved |
| also clone the correct suffix in SFT (ablation) | InT SFT · 235 DeepScaleR problems | accuracy: 7.71% -> 2.31%; coverage 202/235 -> 111/235 | Hurt |
| filter interventions to those yielding a correct rollout in 32 samples (ablation) | InT SFT · 235 DeepScaleR problems | accuracy: 5.06% -> 7.71%; coverage 196/235 -> 202/235 | Improved |
| give the reference solution when proposing interventions (ablation) | intervention generation · Omni-MATH hard subset (334 problems) | accuracy: 0.171% -> 1.56% | Improved |
| use larger Qwen3-30B-A3B-Instruct to propose interventions (ablation) | intervention generation · Omni-MATH hard subset (334 problems) | accuracy: 1.56% -> 2.87%; coverage 80 -> 101 | Improved |
| use reasoning-tuned Qwen3-4B instead of instruct model to propose interventions (ablation) | intervention generation · Omni-MATH hard subset (334 problems) | coverage: coverage 80/334 -> 73/334; accuracy 1.56% -> 2.65% | Mixed |
| combine hints with interventions (ablation) | intervention rollouts · hint-guided pass@16=0 subset | coverage: 18/176 -> 25/176 | Improved |
| add InT SFT before RL | RL on Qwen3-4B-Instruct-2507 · IMO-AnswerBench | accuracy: 23.46 -> 25.62 | Improved |
| SFT on reference solutions before RL | RL on Qwen3-4B-Instruct-2507 · average of four benchmarks | average score: 21.17 -> 20.76 | Hurt |
| SFT on self-reflection rewrites before RL instead of InT | RL on Qwen3-4B-Instruct-2507 · AMO-Bench and IMO-AnswerBench | pass@8 / accuracy: AMO-Bench 36.72 vs 36.16; IMO-AnswerBench 15.53 vs 25.62 | Mixed |
| InT + RL versus standard RL on average score | Qwen3-4B-Instruct-2507 · average of four benchmarks | average score: 28.26 -> 33.72 | Improved |
Surprising findings
- Many hard problems remain unsolved because the policy repeatedly generates incorrect reasoning, per the abstract.
Closest papers by meaning
- $\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving (ICLR 2026 · similarity 0.92)
- General-Reasoner: Advancing LLM Reasoning Across All Domains (NeurIPS 2025 · similarity 0.91)
- Advancing LLM Reasoning with Natural Language and Numerical Feedback (ICML 2026 · similarity 0.91)
- Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective (NeurIPS 2025 · similarity 0.91)
- The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning (NeurIPS 2025 · similarity 0.91)
- Curriculum Reinforcement Learning from Easy to Hard Tasks Improves LLM Reasoning (ICLR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
