Research map / Large language models / Reasoning and chain of thought

Log-Augmented Generation: Scaling Test-Time Reasoning with Reusable Computation

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 Log-Augmented Generation (LAG), which stores KV caches of a selected subset of tokens (last response) from past agent logs, encoding the full reasoning context, and reuses them via positional re-encoding to augment test-time generation.

Problem

LLM agents fail to retain and reuse reasoning from previous tasks; textual logs are noisy and long, and distilled reflections lose reasoning context.

Approach

Store only the KV values of the last model response's tokens while encoding the full reasoning trace, then retrieve and reattach these KV caches (with RoPE re-positioning) to augment generation, improving accuracy and efficiency rather than just speed.

Key results

• On Musique (unseen), LAG-KV reaches 32.2 EM / 45.0 F1 with 2.68 iterations vs. 27.0 EM / 37.3 F1 and 3.90 iterations for standard agentic.
• On GPQA (unseen), LAG-KV achieves 30.4 EM vs. 18.5 for standard agentic.
• Storing only last-action KV reduces average KV storage by 6.30x (to 6.35GB) while still outperforming the no-log agentic baseline on most datasets.

Surprising findings

• Storing full textual reasoning traces (LAG-text-all) often underperforms the no-log agentic baseline on Musique and GPQA, suggesting noise in full traces hurts.
• Adding more content for encoding does not increase KV storage size, since size depends only on stored tokens.

Limitations

• Only applicable to open-weight models where KV states can be accessed and modified.
• Benefit depends on similarity between new tasks and previously logged tasks.
• Static offline log store; no online updating evaluated.

Experimental setup

Datasets: Musique 2WikiMultiHop GPQA MMLU-Pro

Benchmarks: Musique 2WikiMultiHop GPQA MMLU-Pro

Metrics: exact match F1 number of reasoning iterations KV storage size

Baselines: Standard agentic (ReAct) Dynamic Cheatsheet Block Attention LAG-text-all LAG-text-last

Builds on: ReAct Dynamic Cheatsheet Block Attention

Closest papers by meaning

• ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.90)
• Stop When Further Reasoning Won’t Help: Attention-State Adaptive Generation in Reasoning Models (ICML 2026 · similarity 0.89)
• Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (NeurIPS 2025 · similarity 0.89)
• LoongRL: Reinforcement Learning for Advanced Reasoning over Long Contexts (ICLR 2026 · similarity 0.88)
• ArborKV: Structure-Aware KV Cache Management for Scaling Tree-based LLM Reasoning (ICML 2026 · similarity 0.88)
• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.88)