When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression
Ranking
Overall
82
Content
95
Popularity
N/A
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - ICLR is a training-free, online method for compressing long-horizon agent context by selectively removing prior reasoning while retaining actions, tool calls, and observations. It reduces token usage while modestly improving task reward, suggesting reasoning can be forgotten once relevant state is reliably externalized.
- ICLR ranks reasoning blocks for removal using frozen proxy entropy.
- On 260 WorkBuddyBench tasks, average reward increased from 0.699 to 0.718.
- Input, output, and cache-read tokens fell by 25.5%, 14.4%, and 33.3%, respectively.
- Ablations show “trajectory amplification”: deleting local reasoning can nonlinearly affect later actions and total computation.
Sources (1)
When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression
Public signals
N/A
TL;DR - ICLR is a training-free, online method for compressing long-horizon agent context by selectively removing prior reasoning while retaining actions, tool calls, and observations. It reduces token usage while modestly improving task reward, suggesting reasoning can be forgotten once relevant state is reliably externalized.
- ICLR ranks reasoning blocks for removal using frozen proxy entropy.
- On 260 WorkBuddyBench tasks, average reward increased from 0.699 to 0.718.
- Input, output, and cache-read tokens fell by 25.5%, 14.4%, and 33.3%, respectively.
- Ablations show “trajectory amplification”: deleting local reasoning can nonlinearly affect later actions and total computation.