recursive-evidence-replay-reasoning

name	recursive-evidence-replay-reasoning
description	ReContext framework — recursive evidence replay for long-context reasoning, mapping LLM attention to associative memory mechanisms from neuroscience

Recursive Evidence Replay for Reasoning (ReContext)

Description

ReContext (Recursive Evidence Replay as LLM Harness for Long-Context Reasoning) is a training-free inference method that improves long-context reasoning by constructing a query-conditioned evidence pool from model-internal relevance signals, then replaying it before final generation. Its theoretical framework maps directly to neuroscience concepts of associative memory: context as memory store, question as retrieval cue, attention as cue-trace association, and replay as trace reactivation.

Activation Keywords

• recursive evidence replay
• long-context reasoning
• associative memory reasoning
• evidence replay LLM
• context retrieval reasoning
• ReContext framework
• trace reactivation reasoning
• cue-trace association
• 递归证据回放
• 长上下文推理
• 联想记忆推理
• memory store reasoning

Core Concepts

Associative Memory Mapping

The ReContext framework establishes a precise correspondence between LLM mechanisms and biological memory:

LLM Component	Neuroscience Analog	Function
Context window	Memory store	Repository of stored information traces
Query/question	Retrieval cue	Signal that activates relevant memory traces
Attention weights	Cue-trace association	Strength of connection between cue and stored memory
Evidence replay	Trace reactivation	Re-presentation of activated memories to strengthen retrieval

Key Mechanisms

1. Evidence Pool Construction: Uses model-internal relevance signals (attention patterns, hidden state similarity) to identify and extract the most relevant evidence from long context
2. Recursive Selection: Iteratively refines evidence selection by re-running the relevance computation on the selected subset
3. Trace Reactivation: Replays selected evidence before final generation, preserving full original context while emphasizing key information

Usage Patterns

Pattern 1: Long-Context Reasoning Enhancement

When an LLM needs to reason over documents >100K tokens:

1. Extract relevance signals from the model's attention patterns
2. Build a query-conditioned evidence pool from top-k relevant segments
3. Recursively refine by re-computing relevance on the evidence pool
4. Concatenate: [original context] + [refined evidence pool] + [generation prompt]
5. Generate final answer with enhanced evidence utilization

Pattern 2: Evidence-Based QA

For questions requiring specific evidence from long documents:

1. Identify query keywords and semantic similarity to context segments
2. Score each segment by attention-based relevance to query
3. Select top segments forming the evidence pool
4. Replay evidence before generation to prime the model

Pattern 3: Neuroscience-Inspired Memory Systems

When building AI memory architectures:

1. Model the context window as a Hopfield-like associative memory
2. Use attention as the retrieval mechanism (cue-trace association)
3. Implement replay cycles to strengthen weak memory traces
4. Apply recursive selection to handle interference between competing traces

Instructions for Agents

Step 1: Evidence Identification

• Parse the input context into meaningful segments (paragraphs, sections, or logical units)
• For each segment, compute relevance to the query using:
  • Attention weight aggregation from transformer layers
  • Cosine similarity between query embedding and segment embeddings
  • Cross-attention scores if available

Step 2: Evidence Pool Construction

• Rank segments by relevance score
• Select top-k segments (k typically 5-20 depending on context length)
• Ensure diversity: avoid selecting redundant segments that cover the same information

Step 3: Recursive Refinement

• Re-compute relevance scores using the evidence pool as the new context
• This allows the model to identify sub-evidence within already-selected evidence
• Repeat 1-2 times (diminishing returns beyond 2 iterations)

Step 4: Generation with Replayed Evidence

• Format: [Full original context] + \n\n**Key Evidence:**\n + [Refined evidence pool] + \n\n + [Original query]
• The full context preserves completeness; the evidence pool primes relevant information

Error Handling

Evidence Pool Too Small

• If recursive refinement reduces evidence pool below 2 segments, stop recursion
• Use the pool from the previous iteration instead

Irrelevant Evidence Selected

• Apply a relevance threshold (e.g., top 5% of attention weights)
• Use diversity constraints to prevent concentration on a single topic

Context Window Overflow

• If evidence pool + original context exceeds token limit, prioritize evidence pool
• Truncate original context to fit while preserving evidence pool integrity

Resources

• Paper: "ReContext: Recursive Evidence Replay as LLM Harness for Long-Context Reasoning" (arXiv: 2607.02509)
• Code: https://github.com/Yanjun-Zhao/ReContext
• Related neuroscience: Associative memory theory, Hopfield networks, memory consolidation

Related Skills

• agent-memory-framework — memory architecture design
• brain-inspired-memory-agents — neuroscience-inspired memory systems
• context-selective-multimodal-memory — context-selective memory retrieval
• worldkv-world-memory — world model memory systems
• hippocampal-replay-credit-assignment — hippocampal replay mechanisms