Research map / Large language models / Reasoning and chain of thought
Correcting in Hindsight: Editing Past Key-Value States for Robust LLM Reasoning
Mengfei Zhang, Yu Mi, Leijing Zhou
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 HEdit, a lightweight paradigm that edits past key-value states during generation so LLMs can correct early reasoning errors.
Problem
Autoregressive LLMs often fail on complex reasoning because early-stage errors cannot be corrected in later steps, due to the irreversibility of the Transformer's cached states.
Approach
Equips models with a hindsight-like capability for dynamic error correction by editing previously cached key-value states during generation, rather than only appending new tokens.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add HEdit KV-cache editing | Qwen3-0.6B · Math average | accuracy (average): 17.5 -> 28.2 | Improved |
| add HEdit KV-cache editing | Qwen3-8B · Math average | accuracy (average): 66.5 -> 72.8 | Improved |
| add HEdit KV-cache editing | DeepSeek-R1-Distill-Qwen-14B · Math average | accuracy (average): 70.1 -> 72.3 | Improved |
| use HEdit instead of naive backtracking | Qwen3-8B · Math average | accuracy (average): 67.7 -> 72.8 | Improved |
| add naive backtracking without editing KV cache | base models · Math average | accuracy (average): +0.6% to +1.9% | No clear effect |
| add entropy-and-similarity trigger detection (vs random trigger) (ablation) | HEdit · Math average | accuracy (average): 22.2 -> 28.2 | Improved |
| add trigger detection (vs random trigger) (ablation) | HEdit on Qwen3-8B · Math average | accuracy (average): 66.8 -> 72.8 | Improved |
| add anchor localization (vs random anchor) (ablation) | HEdit on Qwen3-0.6B · Math average | accuracy (average): 18.2 -> 28.2 | Improved |
| add anchor localization (vs random anchor) (ablation) | HEdit on Qwen3-8B · Math average | accuracy (average): 62.1 -> 72.8 | Improved |
| add trained MLP editor (vs random perturbation) (ablation) | HEdit on Qwen3-0.6B · Math average | accuracy (average): 12.9 -> 28.2 | Improved |
| add trained MLP editor (vs random perturbation) (ablation) | HEdit on Qwen3-8B · Math average | accuracy (average): 51.9 -> 72.8 | Improved |
| add trained MLP editor (vs random perturbation) (ablation) | HEdit on Llama3.1-8B-Instruct · Math average | accuracy (average): 9.7 -> 21.3 | Improved |
Closest papers by meaning
- GLoRe: When, Where, and How to Improve LLM Reasoning via Global and Local Refinements (ICML 2024 · similarity 0.89)
- A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning (NeurIPS 2025 · similarity 0.89)
- Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression (AAAI 2026 · similarity 0.89)
- Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs (NeurIPS 2025 · similarity 0.88)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)
- Reforming the Mechanism: Editing Reasoning Patterns in LLMs with Circuit Reshaping (ICLR 2026 · similarity 0.88)
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.
