The authors define “direction drift” as a response that completes the apparent task but recommends a course of action inconsistent with the user’s current state.
Wrappers keep LLM coaching aligned with live user state
State-dependent rule layers improved grounding across multi-match game-coaching sessions while reducing first-token latency relative to a tool-using agent baseline.
Chinese Tech
Qi Liu · Xiaoyang Yuan · Yubin Ruan · Zhuomeng Zhang · Wenjin Wang · Di Wu · +5 more
Tencent
Research Digest··2 min read
Thread:Agent Harness Optimization
Liu et al.
Why this paper
From Tencent · Part of Agent Harness Optimization, now 61 papers
In one line
State-Grounded Conditioning uses three wrappers to boost turn-level grounded accuracy from 61.1% to 96.7%.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§