Saved in:
Bibliographic Details
Main Authors: Wang, Jingjing, Chen, Xiwen, Zhu, Wenhui, Li, Huayu, He, Zhengxiao, Cai, Feiyang, Carreon-Rascon, Ana S., Dong, Xuanzhao, Luo, Feng
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.15315
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910222595915776
author Wang, Jingjing
Chen, Xiwen
Zhu, Wenhui
Li, Huayu
He, Zhengxiao
Cai, Feiyang
Carreon-Rascon, Ana S.
Dong, Xuanzhao
Luo, Feng
author_facet Wang, Jingjing
Chen, Xiwen
Zhu, Wenhui
Li, Huayu
He, Zhengxiao
Cai, Feiyang
Carreon-Rascon, Ana S.
Dong, Xuanzhao
Luo, Feng
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2605_15315
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Context Pruning for Coding Agents via Multi-Rubric Latent Reasoning
Wang, Jingjing
Chen, Xiwen
Zhu, Wenhui
Li, Huayu
He, Zhengxiao
Cai, Feiyang
Carreon-Rascon, Ana S.
Dong, Xuanzhao
Luo, Feng
Artificial Intelligence
Computation and Language
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.
title Context Pruning for Coding Agents via Multi-Rubric Latent Reasoning
topic Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2605.15315