| name | recontext-evidence-replay |
| description | Recursive Evidence Replay as LLM Harness for Long-Context Reasoning. Training-free inference method that uses model-internal relevance signals to construct query-conditioned evidence pool and replays before final generation, improving long-context reasoning without training, external memory, or context pruning. |
| trigger | recursive evidence replay, ReContext, long context reasoning, evidence retrieval, context utilization |
| category | ai_collection/collection/skills |
ReContext: Recursive Evidence Replay for Long-Context Reasoning
Overview
RECONTEXT (Recursive Evidence Replay as LLM Harness for Long-Context Reasoning) is a training-free inference method that separates evidence organization from answer generation to improve LLM long-context reasoning.
Core Mechanism
- Evidence Pool Construction: Use model-internal relevance signals to identify query-relevant passages from the full context
- Recursive Replay: Replay the evidence pool before final generation while preserving the original full context
- Associative Memory Theory:
- Context = memory store
- Question = retrieval cue
- Attention = cue-trace association
- Replay = trace reactivation
Implementation Steps
- Feed the full context + query to the LLM
- Extract relevance signals from model attention weights or internal states
- Construct a query-conditioned evidence pool from the most relevant passages
- Prepend this evidence pool to the original prompt and generate the final answer
- Optionally repeat the process (recursive) for multi-hop reasoning
Key Benefits
- Training-free: No model fine-tuning required
- No external memory: Works with existing model parameters
- No context pruning: Preserves full original context
- Consistent improvement: Works across Qwen3-4B, Qwen3-8B, and Llama3-8B
- 128K context: Tested on long-context datasets
Activation Keywords
recursive evidence replay, long context reasoning, evidence pool, context utilization, attention retrieval, associative memory
Source
arXiv: 2607.02509 (2026-07-02)