acceptodds
Under review as a conference paper at ICLR 2027

Compact0: Learning Generalizable Compaction Intelligence

Abstract

Long-horizon LLM agents require context management that preserves task-critical information as interaction histories grow. Existing compaction strategies face two generalization challenges: (i) harness generalization, requiring a unified, sufficiently expressive management policy across heterogeneous harnesses; and (ii) domain generalization, requiring effective context management across diverse task types. We introduce Compact0, a dedicated compaction intelligence model for efficient context management across various agent harnesses and task domains. Technically, Compact0 uses a Context Abstraction Layer (CAL) that unifies native context stores through a shared management protocol, enabling harness generalization. For domain generalization, Compact0 learns agentic context organization through CAL to retain, compress, and recover evidence and execution state via supervised fine-tuning followed by reinforcement learning guided by downstream task rewards and compaction quality rubrics. Experiments spanning NL2Repo coding, deep research, streaming QA, and Claw-style benchmarks demonstrate generalization to unseen harnesses and domains, with performance and token-efficiency gains over default harness mechanisms and specialized context-management baselines. These results support compaction intelligence as a transferable and important capability for long-horizon agents.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agent, Agent Harness

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_S7bAOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

this paper trains a (small language) model for context compression in long horizon llm agents. to achieve this it introduces a context compaction layer (CAL) that first transforms each harness's history into a list of turns. then the model would label each turn as one of the following: visible, compressed, merged, hidden.

this model (used for turn labeling) is trained with SFT on some trajectories generated by a larger model (deepseek v4 pro) and then with RL. in the RL stage, a frozen executor would continue the task after context compaction. evaluations on coding, deep research, QA, and claw like personal ai benchmarks + two recent harnesses (opencode and pi agent) show good numbers over selected baselines while using fewer tokens.

Critical strengths

+ timely topic in the domain of agents: context compression is an important problem for ai agents, especially when horizon is long. many recent papers on this topic.

+ reasonable design. SFT+RL is a very legit process for the kind of turn labeling model authors want.

+ very detailed discussion of the training process (of the turn classifier model). this is strongly encouraged since the insights could help the community even if people might not follow the same training recipe or base language model or dataset.

+ reported results are decent across multiple chosen benchmarks (different applications as well) and language models.

Critical weaknesses

- choice of baselines: could you justify why you pick the current baselines? the practice in this research direction is to compare to published papers or well-known papers with high citations; most naturally, even if you build you own baselines, you can include a baseline where a strong llm just summarizes or drops things; and as i know there are a lot more, either trained compressors or test-time approaches; just to name a few, you don't have to present results for these, ACON (trained), LLM-Lingua and its follow-ups (both), etc.; is your resum baseline trained or not?; the current baselines are pretty simple or relatively less known ones (CoMem is the only one that gets formally published, at icml 2026). without legit baselines i am not sure how i could evaluate this work in the context of the general research direction.

- it is unclear to me if the transfer gains (to unseen harnesses and domains) are strong enough; the train -> transfer evidence is weak in current evaluation. for example, held-out harness is only tested on repo generation (table 1).

Questions for authors

See "weaknesses". I am always ready to be convinced.

Reviewer_NrwGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes Compact0, a dedicated context compaction model trained through supervised fine-tuning and reinforcement learning. It operates through a Context Abstraction Layer (CAL), which enables a shared compaction policy to be applied across different agent harnesses and task domains. Rather than requiring gold-standard compaction targets, the training pipeline uses compaction demonstrations generated by a stronger model for SFT, followed by RL using downstream task performance after compaction and rubric-based rewards.

Critical strengths
  • The idea of a shared, specialized context management model is interesting, and the compaction tool set is understandable. Exploring reinforcement learning to train such a model is a valuable contribution.

  • The experiments show strong results across diverse tasks and harnesses. I particularly appreciate the training-stage ablation in Figure 2. Compact0 outperforms competing compression methods in several long-horizon agent settings while substantially reducing token usage.

Critical weaknesses
  • The efficiency benefits need a more complete assessment of deployment resource requirements. In the Qwen executor settings, the compactor and agent share the Qwen3.6-35B-A3B backbone, although they use different trained weights. Given the compactor's additional KV-cache requirements, it remains unclear how much resource efficiency the method offers when both models must be hosted. Although Table 8 includes compactor token usage and model-call time for a subset of settings, GPU memory requirements are not established by these measurements. Reporting peak input context lengths for both models across tasks and executor settings would help characterize the KV-cache component of peak memory demand. It would also be useful to evaluate whether a smaller compactor could provide greater efficiency gains.

  • The compression interface itself lacks an ablation. While the training-stage ablation is informative, the necessity of the interface choices in Section 3.2 remains unverified. In particular, does the compactor need all four presentation states: visible, compressed, merged, and hidden?

Questions for authors
  • Could the authors provide more comprehensive efficiency metrics beyond the aggregate executor token counts in Table 2? In particular, please report input and output tokens separately, along with peak input context lengths, for both the executor and compactor across the evaluated settings.
Reviewer_VhHGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Compact0, a dedicated context-compaction model for long-horizon LLM agents. It proposes a Context Abstraction Layer (CAL) that provides a unified interface for managing context across different agent harnesses. The model is trained using supervised fine-tuning followed by reinforcement learning with downstream task rewards and temporal quality rubrics. Experiments across coding, deep research, streaming QA, and Claw-style benchmarks show improvements in task performance and token efficiency, including generalization to unseen harnesses and domains.

Critical strengths

1: Context management is an important challenge for long-horizon agents. The unified interface across heterogeneous harnesses is practically valuable, particularly because it does not require modifying the execution models.

2: The paper evaluates multiple execution models, harnesses, and task domains, including unseen configurations. The reported improvements are encouraging, and the training-stage ablations provide some evidence for the effectiveness of outcome-based RL and temporal rubric rewards.

Critical weaknesses

The central idea of training a dedicated context-management model to preserve task-relevant information is closely related to existing work Llike ACON (ICML 2026) and CoMem (ICML 2026). Learning from downstream execution feedback and using SFT followed by RL with outcome or rubric rewards are also established approaches. CAL is a useful engineering abstraction, but its inspect/update/finish operations largely formalize familiar memory-management mechanisms. The paper does not sufficiently establish which component represents a fundamentally new methodological contribution beyond combining existing ideas.

Questions for authors

1: What is the essential methodological difference between Compact0 and prior learned context-management approaches? 2: Could the authors provide a controlled comparison using identical compaction triggers, token budgets, and management-call limits? In particular, how much improvement comes from CAL's richer operations versus the trained model and its RL objective? 3: Can the authors report total token usage, including compactor overhead, across the main benchmarks rather than only the subset in Table 8? This would help establish whether the claimed efficiency advantages hold consistently

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.