ICA: Information-Aware Credit Assignment for Visually Grounded Long-Horizon Information-Seeking Agents
Cong Pang, Xuyu Feng, Yujie Yi, Zixuan Chen, Jiawei Hong, Tiankuo Yao, Nang Yuan, Jiapeng Luo, Lewei Lu, Xin Lou · 2026
Read ICA: Information-Aware Credit Assignment for Visually Grounded Long-Horizon Information-Seeking Agents by Cong Pang, Xuyu Feng, Yujie Yi, Zixuan Chen, Jiawei Hong, Tiankuo Yao, Nang Yuan, Jiapeng Luo, Lewei Lu, Xin Lou (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-2602-10863.txt.
Description
Cong Pang and ten co-authors improve credit assignment for web information-seeking agents in a 2026 paper. Long-horizon reward-based learning struggles because terminal rewards say whether the final answer is right, not which acquired information enabled it. Text-derived webpage views worsen the problem, as parsing and summarization produce unstable content models across paths.
Their evidence-centric system represents acquired information as identifiable units comparable across paths. Fetched webpages become rendered snapshots, keeping layout and mixed content as stable views. On those units, information-aware credit assignment propagates rewards post hoc, estimating turn-level utility from rollout success rates. The method tells agents not just that they succeeded, but what they found that mattered.
Catalog record
| CREATOR | Cong Pang, Xuyu Feng, Yujie Yi, Zixuan Chen, Jiawei Hong, Tiankuo Yao, Nang Yuan, Jiapeng Luo, Lewei Lu, Xin Lou |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-10863