HarnessTTT separates the system into a proposer and a frozen executor.
Learning the search harness improves frozen models’ scientific problem solving
HarnessTTT trains a proposer to construct task-specific prompts, tools and workflows, then transfers those learned search procedures across language-model executors.
Big Tech
Yingzi Ma · Yulong Cao · Nanxi Li · Fang Wu · Yizhe Zhang · James Zou · +2 more
University of Wisconsin-Madison · NVIDIA · Johns Hopkins University · Stanford University · Meta
Research Digest··3 min read
Ma et al.
Why this paper
From NVIDIA and 4 others
In one line
HarnessTTT learns to generate executable harnesses at test time, allowing a frozen small executor to outperform fine-tuned systems across scientific discovery tasks.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§