acceptodds
Under review as a conference paper at ICLR 2027

REDUCE: Risk-Aware Dual-Level Context Compression for Efficient Terminal Agents

Abstract

The performance of large language model (LLM) agents is increasingly constrained by the accumulation of verbose context. Existing context management methods typically treat rule-based observation filtering and model-based history compression as isolated stages, neglecting the cross-stage risk coupling: Aggressive filtering may remove evidence that downstream summarization cannot recover, whereas conservative filtering passes excessive history to the summarizer, increasing its cost and making it harder to preserve task-relevant details. To bridge this gap, we propose a training-free, self-evolving framework REDUCE, Risk-aware Dual-level Context Compression for Efficient Terminal Agents. It features a risk-aware controller that dynamically governs rule activation, history retention, and compression thresholds by synthesizing online signals, such as context token pressure, summary grounding quality, and rule complaint feedback. Furthermore, it maintains dual timescale persistent knowledge: a global observation rule pool accumulates reusable filtering behaviors, while a statistical memory derives protective priors from historical validation outcomes. Experiments across multiple terminal-agent benchmarks and backbone models show that REDUCE reduces end to end token consumption by 16.1%–63.0% while maintaining or improving task accuracy.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Context Compression,Terminal Agents,Large Language Model Agents, AI Agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_x4k2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

REDUCE is a training-free two-level context compression framework for terminal agents. It filters redundant terminal output via an updatable rule pool, then compresses history through fact-verified structured summaries. It designs a risk-aware controller to dynamically adjust filtering, retention windows, and compression thresholds based on token pressure, verification results, and rule feedback.

Critical strengths
  1. REDUCE combines observation filtering and history compression, dynamically adjusting compression intensity based on shared risk signals rather than fixed thresholds.
  2. Summary validation, deferred retries, and extractive fallback help prevent unreliable summaries from replacing important historical information, improving the robustness of context compression.
Critical weaknesses
  1. Limited comparison with related methods. REDUCE does not include a direct comparison with ACON, which makes it difficult to assess the benefits of the proposed risk-aware coordination.
  2. Insufficient ablation of key components. The paper does not isolate the contributions of grounded validation, deferred retries, extractive fallback, or adaptive control.
  3. Limited analysis of compression risks. The experiments mainly report task accuracy and token consumption. A case study of compression risks will be beneficial.
  4. Limited evaluation of efficiency and robustness. The paper reports substantial token savings but does not evaluate end-to-end latency or actual inference costs.
  5. Insufficient experiments. Only three models are evaluated on three datasets, and on CompileBench and DevEval only the results of MiniMax-M3 are reported, with no comparison provided.
  6. Format issues. Figure 3 is hard to read and understand.
Questions for authors
  1. Can the token savings transfer into lower end-to-end latency and inference costs? How much overhead is introduced by summarization, validation, and rule evolution?
  2. REDUCE has many heuristic designs, and the actual effectiveness of each design remains unclear. Moreover, the performance of these designs varies across benchmarks and models, further limiting the method's generalizability.
Reviewer_BkFeOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work proposes REDUCE, a risk-aware framework for dual-level context compression in terminal agents. It addresses the risk that overly aggressive observation filtering may remove information that subsequent history summarization cannot recover. REDUCE combines rule-based observation filtering with structured history compression through a threshold-based controller. The controller makes compression more aggressive when token pressure is high and quality risk is low, and prioritizes information preservation when risk increases. The experiments report modest improvements in task performance while reducing total token usage.

Critical strengths
  • The proposed method reduces token usage on TerminalBench 2.0 while modestly improving task performance.
Critical weaknesses
  • The method is difficult to understand and reproduce. The central motivation concerns the interaction between observation filtering and history compression in long-horizon agents. However, substantial portions of Section 3 describe heuristic designs for the individual compression stages, and their connection to the central coordination problem is insufficiently explained. Although Section 3.3 provides a quality-risk update rule, it does not specify the concrete severity scores, success/failure update functions and coefficients, or controller thresholds, nor how these settings were selected or tuned. The operational criteria for declaring an emergency are also insufficiently specified. Observation rules are adapted through LLM-generated revisions and heuristic feedback rather than explicit task-reward optimization or a defined search procedure, leaving the justification for this adaptation process unclear. The rationale for the six-section summary schema in Section 3.2 is also insufficiently established.
  • The experimental evidence does not sufficiently establish the reliability or source of the improvements. The reported TerminalBench 2.0 results are single-pass evaluations on 89 tasks, making the modest performance gains difficult to assess reliably. I recommend at least three independent attempts per task, reporting both pass@1 and pass@3. Although Table 3 includes component ablations, further controlled comparisons are needed to establish whether the gains over TACO arise from the observation-filtering rules in Section 3.1, the history-compression design in Section 3.2, or the risk-aware coordination in Section 3.3.
Questions for authors
  • How sensitive are the results to the observation-filtering rules and history-compression design in Sections 3.1 and 3.2? Among the proposed components, which contribute most to reducing token usage while preserving task performance?

  • What are the key methodological differences between REDUCE and TACO, and which differences account for the reported gains? Why are other compression methods discussed in the introduction, such as ACON, ReSum, Context Folding, and AgentFold, not included in the experimental comparisons?

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

This paper proposes REDUCE, a training-free context compression framework for terminal agents. It has three parts.

Observation-level rule pool. It starts from six hand-written seed rules. An LLM creates, revises and prunes rules across tasks, tracking lineage and a complaint-based reliability score. Progressive history compressor. It writes a six-section structured summary (TASK/STATE/DONE/ERRORS/FILES/NEXT). A grounding validator checks each candidate summary before it replaces the memory, with deferred retry and extractive fallback. Risk-aware controller. It uses token pressure, a quality-risk state and safety events to choose one of four history profiles and to turn observation filtering on or off.

Experiments use Terminus-2 on Terminal-Bench 2.0 with three backbones, plus CompileBench and DevEval with MiniMax-M3. The paper reports 16.1%–63.0% end-to-end token reduction and 2.2–3.4 point accuracy gains over the uncompressed baseline. A cross-task memory transfer experiment from Terminal-Bench 2.0 to DevEval is also reported.

Critical strengths

The problem is practical. Terminal outputs are verbose, and the motivating example in Figure 1 (six near-identical Valgrind reruns versus a two-line root cause buried in a patch dump) shows clearly that observation filtering and history compression interact. Treating the two levels as coupled is a reasonable view.

Validating a summary before it replaces the memory, and keeping raw history when validation fails, is a sensible safety design. The extractive fallback cannot invent facts. Appendix B gives a concrete trace of a rejected candidate and a recovered one, which helps readers understand the mechanism.

The token reductions on Terminal-Bench 2.0 are large for MiniMax-M3 and Qwen3.5-9B. The paper counts auxiliary compression-model tokens in the end-to-end numbers, which is the right accounting.

Critical weaknesses

W1: Small gains from single runs. Terminal-Bench 2.0 has 89 tasks, so the reported gains amount to 2–3 tasks per backbone. On DevEval and in the transfer study, each step is one task. All results are single-pass, with no variance or confidence intervals, so they do not support the claim of improved accuracy. W2:The ablation does not show that coordination helps. Full uses about 20M fewer tokens than History-only at the same accuracy. But the controller is a fixed threshold rule, and the thresholds and profile hyperparameters are not reported. The gap may simply come from the aggressive profile’s lower budget. Profile selection frequencies and the counts of compressions, rejections and fallbacks are missing. The low reduction on Qwen3.8-27B (16.1%) is not explained. W3Weak baselines and evaluation. There is no comparison with ACON, the closest method, and none with simple baselines such as sliding window, truncation or summarization alone. If the rule pool was evolved on the test tasks, there is a risk of leakage. Rewriting history breaks prefix/KV caching and adds sequential calls, so wall-clock time and cost should be reported, not just token counts.

Questions for authors

1.In Table 3, where does the 20M main-token reduction of Full over History-only and Dual-Level w/o Coordination come from? Please report the profile selection frequencies, compression, rejection and fallback counts, and all controller thresholds and profile hyperparameters. Please also add a sensitivity analysis.

2.Were the rule pool and history priors evolved on the same Terminal-Bench 2.0 tasks used for evaluation? If so, can you report results with memory built on a disjoint task split? Can you compare against ACON and a simple summarization or truncation baseline, and report wall-clock time and cost with prompt caching?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.