Context Pruning for Coding Agents via Multi-Rubric Latent Reasoning
cs.AI, cs.CL
Submitted: 2026-05-14
Updated: 2026-09-26
Comments: COLM 2026 Workshop on Efficient Reasoning
License: http://creativecommons.org/licenses/by/4.0/
The gist: LLM-powered coding agents spend the majority of their token budget reading repository files, yet much of the retrieved code is irrelevant to the task at hand.
Terminology
Abstract
LLM-powered coding agents spend the majority of their token budget reading repository files, yet much of the retrieved code is irrelevant to the task at hand. Existing learned pruners compress this context with a single-objective sequence labeler, collapsing all facets of code relevance into one score and one transition matrix. We show that this formulation creates a modeling bottleneck: a single CRF transition prior must serve heterogeneous retention patterns, including contiguous semantic spans and sparse structural support lines. We propose LaMR (Latent Multi-Rubric), a structured pruning framework that decomposes code relevance into two interpretable quality dimensions, semantic evidence and dependency support, each modeled by a dedicated CRF with dimension-specific transition dynamics. A mixture-of-experts gating network dynamically weights the per-rubric emissions conditioned on the query, and a final CRF layer on the fused emissions produces the aggregate keep-or-prune decision. To supervise each dimension without additional annotation cost, we derive multi-rubric labels from the existing training corpus via AST-based program analysis, simultaneously denoising the teacher's binary labels. By effectively filtering distracting noise, LaMR frequently matches or even outperforms unpruned full-context baselines. Experiments on four benchmarks (SWE-Bench Verified, SWE-QA, LCC, LongCodeQA) show that LaMR wins 12 of 16 head-to-head multi-turn comparisons. It saves up to 31% more tokens on multi-turn agent tasks and improves Exact Match by up to +3.5 on single-turn tasks, while performance is frequently enhanced by denoising the context, and any remaining drops are marginal.
Sources
- Lost in the Middle: How Language Models Use Long Contexts
- Retrieval-Augmented Generation for Large Language Models: A Survey
- LongCodeZip: Compress Long Context for Code Language Models
- UniXcoder: Unified Cross-Modal Pre-training for Code Representation
- Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models
- SWE-QA: Can Language Models Answer Repository-level Code Questions?
- LongCodeBench: Evaluating Coding LLMs at 1M Context Windows
- RepoBench: Benchmarking Repository-Level Code Auto-Completion Systems
- Auto-Rubric: Learning From Implicit Weights to Explicit Rubrics for Reward Modeling
- Scaling LLM Multi-turn RL with End-to-end Summarization-based Context Management
- Scaling Long-Horizon LLM Agent via Context-Folding
- The Complexity Trap: Simple Observation Masking Is as Efficient as LLM Summarization for Agent Context Management
- ACON: Optimizing Context Compression for Long-horizon LLM Agents
- AgentFold: Long-Horizon Web Agents with Proactive Context Management
- COMPASS: Enhancing Agent Long-Horizon Reasoning with Evolving Context
Related papers
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Agents' Last Exam
- The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
- MindHelper: Closed-Loop Embodied Mental-State Reasoning for Precision Intervention
- Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems