🚨 New paper: "Self-Compacting Language Model Agents"
LM agents build up long traces of reasoning and tool calls. As the trace grows, old mistakes and stale info stick around and anchor everything that follows. We ask: can the model itself decide when to clean up?
Users praised the self-compacting language model agents paper because automating context compaction by default improves agent performance based on their experience with manual compaction.
No Digg Deeper questions have been answered for this story yet.
Results across 3 benchmarks in search agents:
+5 to 9 pts across BrowseComp, BrowseComp Plus and DeepSearch QA 30-70% lower per-question cost.
One interesting finding: we're often compacting too late, not too early. Once a search agent finishes a subtask, it's safe to compact right away. SelfCompact's summaries often fire well before the standard 30%-of-context threshold. The fixed schedule retains stale tokens long past when the model has moved on.
We propose SelfCompact, which turns context compaction from a reactive trigger (e.g., firing only when the window is nearly full) into a proactive tool call 🚀
Check out Tianjian's summary for more ⬇️
🚨 New paper: "Self-Compacting Language Model Agents"
LM agents build up long traces of reasoning and tool calls. As the trace grows, old mistakes and stale info stick around and anchor everything that follows. We ask: can the model itself decide when to clean up?
Huge thanks to my collaborators: @jackjingyuzhang, @williamjurayj, @xidulu, @chuanyang_jin and my wonderful advisors: @MFarajtabar, Eric Nalisnick, and @DanielKhashabi.
More details in our paper: https://arxiv.org/pdf/2606.23525
Current systems compress the context on a fixed schedule (every N tokens). Problem: a fixed timer doesn't know what the model is doing. It might fire mid-derivation and wipe facts the model just verified.
We propose SelfCompact: give the model a summarization tool + a lightweight rubric that says when it's safe to fire (sub-task resolved, trajectory converging) and when to hold off (mid-derivation, stuck in a loop). No fine-tuning, no external verifier. Same model does everything.
Amazing work led by @tli104 on letting LLM agents decide for themselves when to compact their own context.
The key idea is pairing a compaction tool with a lightweight rubric that says when to trigger and when to wait, no fine-tuning needed. It's surprising that models still can't tell when their own context is rotting, and that a lightweight rubric closes the gap.
🚨 New paper: "Self-Compacting Language Model Agents"
LM agents build up long traces of reasoning and tool calls. As the trace grows, old mistakes and stale info stick around and anchor everything that follows. We ask: can the model itself decide when to clean up?
@tli104 Nice work! Seems sensible to have this by default. I find agents perform much better in my use when I do manual compaction at good moments.
Another thing I do quite often is rewind + summarize from here. Would be interesting to also allow the LLM to compact the past N turns.
@tli104 The "compacting too late" finding is the spicy one — fixed schedules optimize for memory pressure, not subtask closure. Tying it to tool-graph state rather than token count seems like the implication.
