NVIDIA Mid-Harness improves terminal agents

// 1h agoRESEARCH PAPER

NVIDIA Mid-Harness improves terminal agents

NVIDIA researchers introduce Mid-Harness, which samples candidate terminal actions and verifies them before an unchanged harness executes one. On TerminalBench-Lite, GPT-5.6 Sol verification raises Pass@1 from 50% to 68.03% with eight candidates.

// ANALYSIS

The important idea is placing a judge before irreversible shell actions, making test-time compute more targeted than rerunning entire trajectories.

–Samples multiple actions without modifying the generator or harness
–Pairwise verification outperforms listwise and pointwise methods
–Distilled TMAX-9B reaches 57.14% Pass@1, narrowing but not eliminating the frontier-verifier gap
–Combines with trajectory scaling at lower estimated token cost
–Command semantics and execution feasibility remain major failure modes

// TAGS

mid-harnessagentcoding-agenttool-useevaluationresearch

DISCOVERED

1h ago

2026-10-04

PUBLISHED

2h ago

2026-10-04

RELEVANCE

9/ 10

AUTHOR

lunkertw

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

UPDATE39m ago

Synara Simplifies Worktree Setup to One Click

Synara plans to replace its worktree picker with a simple checkbox, making isolated Git environments easier to enable for parallel coding-agent tasks. Its local-first workspace combines provider sessions, worktrees, orchestration, diffs, and pull-request delivery.

TUTORIAL42m ago

Retro Rad Packs 71 AI Vintage Looks

Retro Rad is a Higgsfield image-to-image app that applies 71 preset styles, including CRT, VHS, glitch, collage, paint, and film-burn effects. Its preset-first workflow lets creators generate retro visuals quickly without constructing complex prompts.

RESEARCH1h ago

Finetuning with Sampling Makes SFT Rival RL

Harvard researchers introduce an MCMC-based projection-sampling method that reshapes expert off-policy traces toward a model’s own distribution before supervised fine-tuning. Across chemistry, math, and medical QA tasks, the approach reportedly improves generalization and reduces forgetting versus strong RL and self-distillation baselines.