ACC: Compiling Agent Trajectories for Long-Context Training

2026-05-02 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Agent Context Compilation (ACC) is a novel method that addresses the costly long-context training data problem for large language models (LLMs) used in agents. ACC converts multi-turn agent trajectories from search, software engineering (SWE), and SQL agents into long-context Question-Answering (QA) pairs. This process compiles the original question with all tool responses and environment observations into a single context, enabling direct supervision of long-context reasoning without additional human annotation. Training Qwen3-30B-A3B with ACC significantly improved performance on long-range dependency modeling benchmarks, achieving 68.3 on MRCR (+18.1) and 77.5 on GraphWalks (+7.6). These results are comparable to the much larger Qwen3-235B-A22B model, while preserving general capabilities on GPQA, MMLU-Pro, AIME, and IFEval. The method leverages 10,802 trajectories with context lengths up to 128K tokens.

Key takeaway

For Machine Learning Engineers developing agent-powered LLMs, ACC offers a scalable and effective approach to enhance long-context reasoning without expensive manual annotation. You should consider integrating ACC into your supervised fine-tuning pipelines, especially when working with multi-turn agent trajectories from search, software engineering, or database querying. This method can significantly boost performance on complex long-range dependency tasks like MRCR and GraphWalks, potentially matching larger models with fewer parameters, while maintaining general capabilities.

Key insights

Agent Context Compilation (ACC) transforms multi-turn agent trajectories into long-context QA pairs, directly supervising long-range reasoning.

Principles

• Masking tool responses creates a supervision blind spot.
• Explicitly compiling scattered evidence improves reasoning.
• Diverse trajectory types offer complementary coverage.

Method

ACC compiles an agent trajectory τ=(q,(r₁,a₁,o₁),…,(rₖ₁,aₖ₁,oₖ₁),(rₖ,y)) into a QA pair (x,y) where x=(q,o₁,…,oₖ₁), training the model to answer y directly from x.

In practice

• Apply ACC to search, SWE, or SQL agent trajectories.
• Combine ACC with existing long-context extension methods.
• Include distractors in compiled contexts for localization.

Topics

• Agent Context Compilation
• Long-Context LLMs
• Supervised Fine-Tuning
• Agent Trajectories
• MRCR Benchmark
• GraphWalks Benchmark
• Qwen3-30B-A3B

Code references

• gkamradt/LLMTest_NeedleInAHaystack

Best for: NLP Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Latent Context Compilation: Distilling Long Context into Compact Portable Memory — Latent Context Compilation distills long LLM contexts into portable buffer tokens using a disposable LoRA and self-aligned optimization.
• How AI Agents Navigate Massive Skill Graphs — Skill-DAG enables LLMs to dynamically interpret and evolve a typed skill graph for scalable, context-aware skill selection, moving beyond semantic similarity.
• WRIT: Write-Read Intensive Trajectory Synthesis for Multi-Turn User-Facing Agents — WRIT synthesizes agent training trajectories balancing sequential execution and evidence-gathering complexity for robust decision-making.
• ACCORD: Action-Conditioned Contextual Grounding for Language Agents — ACCORD improves LLM agent performance by actively grounding underspecified instructions with environmental and trajectory context.
• Connect the Dots: Training LLMs for Long-Lifecycle Agents with Cross-Domain Generalization Via Reinforcement Learning — LLMs can achieve continuous self-improvement and cross-domain generalization by learning to self-update context via end-to-end RL.
• LongAct: Harnessing Intrinsic Activation Patterns for Long-Context Reinforcement Learning — High-magnitude activations in query/key vectors are critical for long-context LLM reasoning and RL optimization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.