arxiv arXiv cs.CL · just now · research

Coroutine-bridge harness wins CAR-bench Track 2 with 60.0% Pass^3

from English

A new coroutine-bridge harness for tool-using agents won Track 2 of the CAR-bench evaluation, achieving a 60.0% Pass@3 score on the official hidden test set. The system decouples model invocation from tool round-trips by having the model emit Python programs that block and resume across evaluator exchanges.

The agent uses a median of two model calls against seven agent turns per task, resolving tasks in a median of 1.8 seconds of model latency on Cerebras gpt-oss-120b.
It achieved 4.5 times the organizer baseline at the lowest estimated cost and fastest median task latency (3.14 s) among entries scoring above that baseline.
The unchanged harness reproduced an identical 60.0% Pass@3 on GPT-5.5 in the Open track, matching frontier-model agents.
A single static prompt served 78% of input tokens from cache, significantly reducing nominal input compute.

The approach enforces compliance at zero reasoning cost by encoding deterministic policies as logic in the tool layer rather than prompt rules.

Importance 2/3 New harness with differentiators arXiv cs.CL AI agents Inference efficiency

Related articles

arxiv arXiv cs.AI · 2d ago · 17 views

CliffCompaction reduces coding agent costs by 50% while maintaining performance

Researchers introduce CliffCompaction, an autocompaction technique designed to reduce costs for long-horizon coding agents by up to 50% under a bounded context. The method maintains or improves performance on Terminal-Bench and achieves state-of-the-art results on KernelBench by keeping compacted information faithful through truncation rather than rephrasing.

arxiv arXiv cs.LG · 2d ago · 15 views

CliffCompaction reduces long-horizon coding agent costs by up to 50% while maintaining performance

Researchers introduce CliffCompaction, an autocompaction technique designed for agents handling millions of tokens, which cuts costs by up to 50% under bounded context. The method maintains or improves performance on Terminal-Bench and achieves state-of-the-art results on KernelBench by keeping information faithful through truncation rather than rephrasing.

lab OpenAI News · 2d ago · 19 views

Parallel halves research time and cost using GPT-6 Astra

Parallel has integrated OpenAI's GPT-6 Astra into its AI agent infrastructure, achieving a 50% reduction in both research time and code costs compared to prior models. The company reports that the new model allows agents to complete complex knowledge work significantly faster while maintaining high-quality output.

arxiv arXiv cs.AI · 3d ago · 13 views

AgentRouter reduces agentic workflow costs by 72% via step-level model routing

Researchers propose AgentRouter, a lightweight classifier that optimizes multi-step agentic workflows by routing individual trajectory steps to appropriate model tiers rather than using frontier models for every task. The system addresses the inefficiency of existing solutions that ignore varying subtask complexity within a single agent session.

arxiv arXiv cs.CL · 3d ago · 13 views

AgentRouter reduces agentic workflow costs by 72% via step-level model routing

Researchers propose AgentRouter, a lightweight classifier that optimizes multi-step agentic workflows by routing individual trajectory steps to appropriate model tiers rather than using a single frontier model for all tasks. The system addresses the inefficiency of enterprise systems that waste 60-80% of their inference budget on subtasks that smaller models can handle equally well.