SynAct samples repeated rollouts of a task, groups actions taken from the same environment state, and computes the mutual information between the action at a decision point and the terminal outcome class.
Expert feedback for LLM agents should target decisions, not whole tasks
SynAct ranks decision points by how strongly actions explain outcomes, then distills corrections into gated rules.
Big Tech
Ning Liu · Ian Li · Kristina Klinkner
Amazon
Research Digest··3 min read
The authors present SynAct, an active learning framework that allocates expert feedback to individual decision points rather than whole trajectories.
Why this paper
From Amazon
In one line
SynAct allocates expert feedback to decisions whose actions best predict terminal outcomes, improving held-out accuracy by 11.7 points on average.
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
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§