AIAny
Icon for item

FastContext: Training Efficient Repository Explorer for Coding Agents

Moves repository search into a dedicated exploration subagent that issues parallel read-only READ/GLOB/GREP calls and returns compact file:line citations. Trained (4B–30B) with SFT+RL, it reduces main-agent token use up to ~60% and raises end-to-end success by up to ~5.5%.

Introduction

Oops! Something went wrong

[next-mdx-remote-client] error compiling MDX: Expected a closing tag for `<final_answer>` (5:299-5:313) before the end of `paragraph` 3 | ## Key Findings 4 | - Separation pays: moving exploratory reads outside the solver reduces the main agent's token consumption by up to ~60% while improving end-to-end resolution rates by up to ~5.5% in reported benchmarks. This means the solver keeps a cleaner context window and spends more tokens on planning/editing rather than file hunting. > 5 | - Lightweight, trainable explorers: the paper trains 4B–30B exploration models (bootstrapped with supervised trajectories and refined with task-grounded RL) to optimize for broad first-turn coverage, multi-turn evidence gathering, and precise citation generation. Trained explorers produce compact <final_answer> blocks of file paths and line ranges instead of pasting large code snippets. | ^ 6 | - Practical tool set and workflow: FastContext exposes only read-only tools (READ, GLOB, GREP) and supports parallel tool calls per exploration turn. The intended contract is simple: the main coding agent invokes FastContext for context; FastContext returns focused citations; the main agent edits/tests/answers using that evidence. 7 | More information: https://mdxjs.com/docs/troubleshooting-mdx

Information

  • Websitearxiv.org
  • OrganizationsMicrosoft
  • AuthorsShaoqiu Zhang, Maoquan Wang, Yuling Shi, Yuhang Wang, Xiaodong Gu, Yongqiang Yao, Rao Fu, Shengyu Fu
  • Published date2026/06/12

Categories

More Items

Synthesizes a deployment-specific safety harness (natural-language policy plus executable logic) for a frozen LLM in a target domain, using model behavior, domain specs, and adversarial review to optimize safety–utility trade-offs and transfer across agent benchmarks.

Defines the Discovery Certification Protocol (DCP) to audit claims of discovery by AI research agents, converting claims into executable recovery and feedback tests. Specifies multi-gate certification, Core control requirements, and a deterministic offline verifier; validated in two controlled audits.

Evaluates whether AI agents can autonomously use Sparse Autoencoders (SAEs) to discover interpretable features and causally steer model behavior. Introduces a 20-task benchmark that has agents design contrastive probes to search a 131K+ feature dictionary in Gemma-2-9B-IT and evaluates activation rank, selectivity, and causal steering against expert references.