The authors developed Interaction Aware Compression for Long Horizon Reasoning (ICLR), a training-free online method that uses entropy estimates from a frozen proxy model to rank reasoning blocks for removal.
Agents Can Drop Old Reasoning After Externalizing Task State
A training-free compression method selectively removes reasoning history while retaining actions, tool calls, observations, and externally stored task state.
Top University
Mingxuan Wang · Fei Luo · Bo Wang · Guorun Yao · Yinglong Guo · Chao Ning · +3 more
TierFlow Team · Renmin University of China · Tsinghua University
Research Digest··2 min read
Wang and colleagues tested whether long-running language-model agents need to retain all of their earlier reasoning.
Why this paper
From Tsinghua University and 2 others · Part of Context Engineering for Agents, now 23 papers
In one line
Long-horizon agent reasoning can be compressed by ranking blocks with frozen proxy entropy, improving reward while reducing token cost.
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.
§