The authors measured output tokens per response across six models and found reasoning intensity declines as score gains diminish.
Alternating AI agents across sessions beats single agents on machine learning tasks
A workflow that caps each session at k experiments and resets context reaches state-of-the-art any-medal rates on MLE-bench.
Big Tech
Changye Li · Zijian Zhang · Muheng Li · Yiduo Pang · Tingjun Wu · Menghan Li · +9 more
Tsinghua University · NVIDIA · Zhejiang University · MIT
Research Digest··2 min read
The authors identify two failure modes in long-horizon machine learning agents: long-horizon deceleration, where reasoning intensity drops as a session stretches, and harness staleness, where custom tooling helps less as base models improve.
Why this paper
From NVIDIA and 3 others
In one line
Alternating two LLM agents with context resets and experiment caps outperforms single-agent workflows on long-horizon machine learning engineering.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓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.
§