lab NVIDIA Research · 4h ago · research

Mid-Harness improves terminal agent action reliability via test-time compute scaling

from English

Researchers introduce Mid-Harness, a method that samples and verifies candidate actions at the model-harness boundary to improve execution reliability in terminal agents. This approach allocates test-time compute by verifying multiple options before forwarding one for execution, keeping the generator and harness unchanged.

On TerminalBench-Lite, using a GPT-5.6 Sol verifier with 8 sampled actions raises Pass@1 from 50.00% to 68.03% compared to the base agent.
When TMAX-9B serves as the verifier, pairwise verification yields the best results among evaluated mechanisms.
Distilling responses from a stronger verifier into TMAX-9B further improves Pass@1 without changing the action generator.
Combining action and trajectory scaling with TMAX-9B achieves higher success at lower estimated token cost than generating more trajectories alone.

These findings identify action scaling as a promising target for test-time compute scaling in terminal agents, improving performance across additional models, benchmarks, and harnesses.

Importance 1/3 New harness with differentiators Trust 3/3 NVIDIA Research AI agents Research paper

Benchmarks

Benchmark	Model	Score
Terminal-Bench	GPT-5.6 Sol	68.03%
Terminal-Bench	TMAX-9B	50%

Related articles

arxiv arXiv cs.AI · 2d ago · 9 views

Agentic meta-reasoning scales inference through explicit control

The authors introduce agentic meta-reasoning, an inference-time harness that structures execution choices for long-horizon tasks. A controller manages the process by consolidating progress and dispatching work based on a compact run account rather than full history.

arxiv arXiv cs.LG · 7d ago · 7 views

GRASP framework improves strategic planning with multi-stage agentic AI

Researchers introduce GRASP, a strategy-aware, multi-stage planning framework designed to generate high-quality natural language executable plans for complex tasks by decoupling the pipeline into specialized modules. The system pre-compiles global macro-guidelines via GenPlan, explores localized strategies through RevPlan, and evaluates trajectories using a multi-criteria discriminator called VerPlan.

arxiv arXiv cs.AI · 7d ago · 7 views

GRASP introduces strategy-aware multi-stage planning to improve LLM reliability

Researchers introduce GRASP, a strategy-aware, multi-stage planning framework designed to generate high-quality natural language executable plans for complex tasks where standard Large Language Models typically degrade in reliability. The system decouples the planning pipeline into specialized modules: GenPlan for global macro-guidelines, RevPlan for exploring localized strategies, and VerPlan for independent trajectory evaluation.

lab Anthropic News · 8d ago · 41 views

Anthropic's Claude discovers novel ART enzyme system in bacteriophages

Anthropic has announced a new life sciences research group dedicated to using Claude to identify uncharacterized protein families and test hypotheses through laboratory experiments. In early results, the AI autonomously discovered a novel enzyme system called array-associated reverse transcriptases (ART) found in bacteriophages.

arxiv arXiv cs.AI · 9d ago · 26 views

AIDE^2 enables recursive self-improvement of AI research agents

Researchers present AIDE^2, a system that implements a loop of recursive self-improvement for an AI research agent by optimizing its own code. The system proposes changes to its internal logic, benchmarks modified versions on AI R&D tasks, and retains only those improvements that perform best on hidden evaluations.