Same Signal, Opposite Meaning: Direction-Informed Adaptive Learning for LLM Agents

Ziming Li, Jiatan Huang, Xiaoguang Guo, Guilin Wang, Chuxu Zhang {ziming.li,jiatan.huang,xiaoguang.guo,chuxu.zhang}@[email protected] · 2026

Read Same Signal, Opposite Meaning: Direction-Informed Adaptive Learning for LLM Agents by Ziming Li, Jiatan Huang, Xiaoguang Guo, Guilin Wang, Chuxu Zhang {ziming.li,jiatan.huang,xiaoguang.guo,chuxu.zhang}@[email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-06908.txt.

Description

Ziming Li and four co-authors show that extra test-time computation helps or harms language model agents depending on direction, not just amount. Adaptive methods gate extra rollouts by confidence or uncertainty, assuming the signal's meaning is fixed. The authors demonstrate the same signal predicts benefit in one environment and harm in another, even with the task unchanged. Uncertainty can mark states where comparison helps, or states where intervention cannot.

Their DIAL method learns the utility direction of state features separately for each environment and model backbone, trained from counterfactual exploration that does not presuppose the signal's meaning. Across six environments and three backbones it achieves a stronger success-versus-cost trade-off than fixed-direction gates.

Catalog record

CREATORZiming Li, Jiatan Huang, Xiaoguang Guo, Guilin Wang, Chuxu Zhang {ziming.li,jiatan.huang,xiaoguang.guo,chuxu.zhang}@[email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE106 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.06908

Browse the full library · About the Tower of Babel