Stable Geometry with Divergent Task Evidence for Efficient Long-Horizon Agent Compression
Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao, Guorun Yao, Yinglong Guo, Chao Ning, Hongyue Chen, et al.
cs.AI
Quick verdict
GEM cuts agent tokens 21.4% at equal reward; geometry-vs-evidence analysis is strong, but the 13-baseline comparison is unpaired and reward parity untested.
Summary
The paper argues that geometric redundancy in long-horizon LLM agent histories is an unreliable criterion for deciding what can be safely compressed. Using frozen embeddings of complete tool-call/result blocks (171 trajectories/4,701 blocks in Qwen3.5-9B space; 32 trajectories in the selector's Qwen3-Embedding-8B space), the authors show histories are strongly low-dimensional (median effective rank 8.39 vs 16.99 for a feature-shift null; Holm-adjusted p <= 1.86e-9), yet matched-budget experiments over 650 checkpoints show geometry-only selection preserves little next-action evidence: Top-3 retention 0.31 vs 0.69 for evidence-aware selection at identical counts, with centroid similarity unchanged at 0.98. A controlled single-block replacement removes action evidence (Top-3 -0.33) while global geometry metrics barely move. They then propose GEM, a training-free compressor that first protects recent, goal, state, error and source evidence, then greedily completes geometric coverage (tau=0.90, K=16) under a 5% per-request deletion guard. On WorkBuddyBench Full260 with DeepSeek-V4-Flash, combined tokens fall from 2.69M to 2.11M per task (-21.4%) with mean reward essentially unchanged (69.87 to 70.18). A fixed-40 panel compares 13 context-management baselines, but GEM's row comes from a different evaluation snapshot and is admittedly not paired. Main limitations: single benchmark and acting model, hand-tuned protection rules, no uncertainty quantification on end-to-end reward, and no code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 3
● low 0
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Agent histories are strongly low-dimensional (median effective rank 8.39 vs 16.99 for a feature-shift control; all 171 isolated-block and 32 selector-space comparisons agree in sign, Holm-adjusted p <= 1.86e-9), but this concentration does not imply blocks are safe to remove.
- At identical retained block counts, evidence-aware selection raises next-action Top-3 retention from 0.31 to 0.69 while centroid similarity stays 0.98 (650 checkpoints, 32 trajectories, Table 1).
- Controlled replacement of a single low-energy, action-aligned block cuts Top-3 retention by 0.33 while centroid similarity and captured energy change by ~0.00, showing action evidence lives largely outside dominant geometry (Table 2).
- GEM, a training-free protect-evidence-then-complete-geometry compressor, reduces mean combined tokens from 2.69M to 2.11M per task (-21.4%) on WorkBuddyBench Full260 with essentially unchanged mean reward (69.87 to 70.18).
Strengths
Well-controlled demonstration of the geometry-evidence gap
Feature-shift and Gaussian matched nulls preserve marginals/shape while disrupting alignment; 171/171 isolated-block and 32/32 selector-space sign consistency with Holm-adjusted p <= 1.86e-9 (Sec. 4.1, Table C.1).
Budget-controlled and interventional evidence, not just correlational
Table 1 (Top-3 0.31 vs 0.69 at identical retained counts, 650 checkpoints, centroid 0.98 both) and Table 2 (single-block replacement: Top-3 -0.33 with centroid/energy +0.00) separate selection quality from budget and establish dissociation beyond correlation.
Honest token accounting and deliberately scoped claims
Implementation-grade specification with correctness proofs
Multi-granularity ablations
Weaknesses
Headline reward parity lacks uncertainty quantification
13-baseline comparison is not paired
Much of the evidence signal may be simple recency
Many hand-set constants, single benchmark/model
Proxy metrics never linked to task reward
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets history compression for long-horizon LLM agents. Its central thesis: geometric redundancy (low effective rank, high coverage of dominant representation directions) is an unreliable proxy for what can be safely discarded, because records weakly expressed in the dominant subspace can still carry execution-critical evidence. Three claims are made: (1) agent histories are strongly low-dimensional relative to matched nulls (Sec. 4.1: median effective rank 8.39 vs 16.99 feature-shift control); (2) same-budget geometry-only selection loses next-action evidence that evidence-aware selection keeps, while global geometry metrics stay constant (Table 1: Top-3 0.31 vs 0.69, centroid 0.98 vs 0.98, 650 checkpoints); (3) GEM, a protect-evidence-then-complete-geometry compressor, cuts combined tokens 21.4% (2.69M to 2.11M per task) on WorkBuddyBench Full260 at essentially unchanged mean reward (69.87 to 70.18, Table 3). Claims are mostly scoped carefully in the text; the abstract's "maintaining comparable task reward" is the phrasing that most outruns the statistics.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What problem does GEM solve and how does it work?
LLM agents running long tasks accumulate huge interaction histories that inflate context and inference cost. GEM is a training-free external compressor that decides which complete history blocks (a tool call plus its results) to keep: it first protects execution-critical evidence (recent context, goal-related records, file-state writes, fresh errors, source observations), then fills remaining capacity by preserving geometrically uncovered embedding directions. The acting model is unchanged and receives the original, chronologically ordered messages.
How much does GEM save, and at what quality cost?
On the 260-task WorkBuddyBench with DeepSeek-V4-Flash, mean combined tokens per task dropped from 2.69M to 2.11M (-21.4%), while mean task score was essentially unchanged (69.87 vs 70.18 on a 0-100 scale). The paper's token decomposition shows savings come from reduced cached input (2.60M to 1.82M), partly offset by added embedding tokens and uncached input, all included in the accounting.
Why is keeping 'similar geometry' not enough for agent history compression?
Agent histories are strongly low-dimensional, so many blocks look geometrically redundant, yet blocks weakly expressed in the dominant directions can still carry information the next action depends on. At identical retained block counts, geometry-only selection kept only 0.31 Top-3 next-action retention versus 0.69 for evidence-aware selection, while the global centroid similarity stayed 0.98 in both cases - similar global geometry, very different preserved evidence.
Is code or data available for GEM?
The extracted text contains a reproducibility statement with detailed methodological specification (pseudocode, hyperparameter table, evaluation protocol) but no public code repository or dataset release. Evaluation depends on the WorkBuddyBench benchmark (itself a 2026 preprint) and the commercial DeepSeek-V4-Flash model, so independent reproduction currently requires substantial reimplementation.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-3 and 5 are available only as captions in the extraction; plotted distributions could not be inspected.
- Tables were flattened by text extraction; row/column alignment was re-derived from captions and prose.
- No code or data link appears in the extracted text; availability on the arXiv landing page could not be verified.
- Content beyond Appendix E.3 was not present in the extraction; any further appendices were not reviewed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.