ByteDance Paper: Keeping Raw Attempt History Outperforms Summaries for LLM Reasoning
rohanpaul_ai · x · 2026-08-31
A new ByteDance paper explores test-time computation optimization, introducing the Chain-of-Experience (CoE) method. Contrary to traditional approaches that summarize past attempts into concise memories, CoE advocates keeping the full history of earlier attempts and feedback in context, asking the model to try again based on this raw data.
Across 6 benchmarks covering math, coding, and knowledge, the CoE method with self-feedback achieved an average score of 71.0%, compared to 66.8% for iterative solving without feedback. This suggests that retaining the 'messy' history of attempts can be more effective than compressing it into 'neat' summaries for specific tasks.
More from Research
- DreamX-Creator: Native 2K Audio-Video Generation via 7B Model — GD-ML · 2026-09-01
- Matrix-Game 3.5: Enhancing Real-Time Streaming Interactive World Models with Patch Memory — Runjia Qian · 2026-09-01
- ByteDance Releases Lucida for Composable Real-to-Sim Scene Modeling — ByteDance-Seed · 2026-09-01
- Qwen Team Analyzes Qwen3.8-Next Architecture Design — Qwen · 2026-09-01
- Normalized LoRA Stabilizes Training Without Extra Cost — Jiale Kang · 2026-09-01
- Google Launches PaperBanana-Interact for Scientific Diagram Refinement — google · 2026-09-01