acceptodds
Under review as a conference paper at ICLR 2027

LLM Agents Are Latent Context Managers: Eliciting Self-Managed Context via State Proprioception

Abstract

Long-horizon tool agents are bottlenecked by how their context grows toward the limits of the context window. Existing systems delegate context management to runtime rules or train agents to compress their working memory, often leaving runtime state implicit or discarding evidence through summarization. We argue that a basic obstacle is context-state blindness: agents must decide what to keep or archive without an explicit view of block costs and remaining capacity. We introduce VISTA (Visible Internal State for Tool Agents), a training-free, model-agnostic layer that represents working memory as typed, addressable blocks, exposes token usage, creation age, archive status, and remaining budget through a runtime dashboard, and archives blocks as recoverable full-fidelity payloads. Across LOCA-Bench, BrowseComp-Plus, and GAIA, the same untrained interface transfers across trajectory scales. On LOCA-Bench, it raises Gemini-3-Flash accuracy from 22.7% to 50.7%; it reaches 58.0% on BrowseComp-Plus and remains competitive on GAIA. Pressure sweeps characterize its operating range, gains transfer across four backbones, and ablations show that the dashboard contributes beyond archive and recovery tools.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, context management, long-horizon reasoning, working memory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_x8HiOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VISTA, a training-free context-management framework for long-horizon LLM agents. It exposes explicit runtime state through a dashboard and allows agents to archive and recover context blocks. The method is evaluated on LOCA-Bench, BrowseComp-Plus, and GAIA, with additional ablations and a small RL adaptation study.

Critical strengths
  1. The paper studies an important and practical problem for long-horizon LLM agents: how to manage growing context under limited context budgets. The motivation is clear and easy to understand.

  2. VISTA is simple, training-free, and model-agnostic. The explicit context dashboard together with recoverable archiving provides a practical interface that can be integrated into existing agent systems without modifying the underlying model.

Critical weaknesses
  1. Both BrowseComp-Plus and GAIA are evaluated with DeepSeek-V4-Pro under a restrictive 12K active context window. For BrowseComp-Plus, the paper explicitly states that this setting is deliberately chosen to induce evidence eviction. While this can serve as a useful stress test, it may also disproportionately favor a method specifically designed for context management. The ReAct baselines on both BrowseComp-Plus and GAIA also appear relatively weak. Evaluation under more standard long-context settings, such as 128K or 256K, would make the results more convincing.

  2. The cross-backbone evaluation is only conducted on LOCA-Bench. BrowseComp-Plus and GAIA are evaluated only with DeepSeek-V4-Pro, which makes the claim of model-agnostic generalization less convincing. It would be useful to verify whether the gains on these benchmarks also hold across different model families.

  3. The RL objective mainly uses a trajectory-level outcome reward together with local heuristic penalties. However, context-management actions can have delayed effects across multiple workspace states: whether an archive or recovery action is beneficial may only become clear several segments later. The current formulation does not explicitly model such cross-segment reward propagation, making the credit assignment relatively coarse. Related work such as IterResearch [1] provides a more explicit treatment of multi-round state transitions and reward assignment, which may be relevant here.

  4. The RL experiments are relatively limited, with only a small Qwen3-8B adaptation study. Important training details, including the number of training examples, rollout budget, group size, total training tokens, and update steps, are not clearly reported. Several of these are even listed as pending implementation checks in the appendix, making the RL results difficult to assess and reproduce.

  5. The theoretical section formalizes several intuitive properties of the proposed design under highly stylized assumptions, such as the benefit of lossless recovery and explicit block-size information. However, these results appear only loosely connected to the actual empirical setting and do not seem necessary for explaining the main experimental findings. In several places, the assumptions already encode much of the behavior that the theory is intended to justify, while the corresponding theoretical quantities are not directly evaluated in the experiments. As a result, the extensive theoretical development feels somewhat disproportionate to the practical contribution of the paper. A more concise analysis, or a tighter connection between the theoretical results and measurable behavior in the actual benchmarks, would make this part more convincing.

Reference

[1] Guoxin Chen et al. IterResearch: Rethinking Long-Horizon Agents via Markovian State Reconstruction. ICLR 2026.

Questions for authors

As far as I know, the GAIA validation split contains 165 questions in total. Why does the paper refer to a “random 165-question subset”? Is this actually the full validation set, or was some other sampling procedure used?

1 comment
AuthorsOpenReview ↗

Response to Reviewer x8Hi

Thank you. Two of your concerns are answered by the setting of our main benchmark, and we have corrected the paper on the others.

Standard long-context settings. The headline LOCA-Bench results already use a standard 128K active window (max-context-size = 128,000) with Gemini-3-Flash. There, VISTA solves 38/75 against ReAct's 17/75, paired b/c = 1/22, exact McNemar p = 5.7×10⁻⁶. The LOCA sweep keeps that 128K window and raises how much context each task accumulates. Where accumulation is heavy, the gain is large and stable: VISTA 57.3 vs ReAct 36.0 at 96K, 50.7 vs 22.7 at 128K, and 32.0 vs 12.0 at 256K (Table 8). So the effect does not depend on an artificially small window; it appears when the task outgrows the window, which is the regime the method is for.

The 12K window is used only for BrowseComp-Plus and GAIA, whose trajectories are around 100K and 10K tokens. With a 128K window those runs would rarely need to evict anything, and the context managers in Table 1 would behave much like ReAct. We agree that 12K is a stress setting chosen to create retention pressure. The revised §4.1 says so and no longer cites the gold-retrieval-depth statistic, because we cannot show it was computed on development data.

The ReAct baselines appear weak. ReAct is the harness-native append-until-truncation loop, with the same tools, prompt, budget and scoring as every other row; it is not tuned down. On the same 75 LOCA tasks at the 8K growth setting, where little management is needed, it solves 84.0%. It falls to 22.7% at 128K because the context fills, and that is the failure the paper studies.

Cross-backbone evaluation. We agree that the backbone evidence is LOCA-only, and Appendix J now says so. On LOCA at 128K, the same interface runs unchanged on four backbones (Figure 6). "Model-agnostic" in the paper means that the layer wraps a backbone without per-backbone tuning. We also report a fifth backbone we had run but not plotted, GPT-5.2, where both methods stay near the floor (VISTA 9.3%, ReAct 8.0%). We report it because a transfer claim should show its misses.

GAIA "random 165-question subset". It is the full 165-question validation split (Levels 1/2/3: 53/86/26). "Random subset" was a wording error, corrected in §4.1 and Appendix I.

RL credit assignment and details. We agree the objective is coarse: a trajectory outcome reward plus local penalties, with no cross-segment propagation. IterResearch's Markovian state reconstruction is the relevant treatment, and we will discuss it in related work. The details you asked for are now in Appendix M.1: group size G = 8, one training seed, PPO ratio clip 0.2/0.2 with dual-clip 3.0, gradient-norm clip 1.0, and maximum prompt and response lengths of 8,192 and 32,768 tokens. Per-variant update counts, training-set sizes and total training tokens are listed as open in M.1. With one seed, §4.4 is now labelled a single-seed pilot, and none of the frozen-model results depend on it.

Theory. We partly agree. Theorem 1's operative quantity is the rate of the size channel available to the agent: a low-rate interface forces make-room errors. The one place we measure that rate is the size-estimation probe. Without the ledger, four backbones misestimate total context size by a median relative error of 0.43–0.84 (Table 12), which is the low-rate regime the theorem describes. We will tighten §3.3 around this link and move the stylized constructions to the appendix.

Informativeness. The result we would point to as new is a dissociation: keeping every archive and recovery tool but removing only the per-block display costs about ten tasks on the matched launches, and every launch with the display is above every launch without it (34–39 solved versus 26–29; paired p = 0.013 on the submitted launches). Prior context managers ship tools and display together, so as far as we know this separation has not been measured before.

Reviewer_NsdvOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a training-free mechanism for context management by dividing the context into blocks (blocks are single-turns). As per compression needs, a block is summarized and removed form context while the exact content remains discoverable as a stored file. The VISTA method provides metadata such as per-block token cost, age, etc. to allow better decision making by the agent to manage its context. Experiments are performed across popular benchmarks and model backbones to establish superior performance over existing methods.

Critical strengths

S1: VISTA provides a simple and useful systems idea for context management

S2: The evaluation is broad and covers multiple domains and models to show the generality of the approach

Critical weaknesses

The most critical point I would like to highlight is that the work is unfinished as Appendix M contains a detailed list of To-Dos and weaknesses. The To-Dos substantially undermine the presented results, and may indicate the authors have not read their own submission in its entirety. Below, I provide my own notes before I read Appendix M:

W1: The presentation is extremely dense.

W2: Limited novelty and comparison to existing compression methods (acknowledged by the text that PACE, HiAgent, learned managers are missing)

W3: Block granularity is underexplored. Large blocks can mix relevant/irrelevant information and operations at block-level may add additional burden to the agent. The impact of summary quality is also not isolated

W4: The approach invalidates KV cache, leading to higher costs

Questions for authors

Q1: Were all items listed in Appendix M unresolved at submission time?

Q2: Can the authors provide baseline comparisons for context management approaches (PACE, Mem1)?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Nsdv

Thank you. You identified the most damaging artefact in the submission, so we answer it first and directly.

Q1. Were the Appendix M items unresolved at submission time? Yes. Appendix M was our internal reproduction checklist, and it should not have been submitted in that form. Its opening sentence says the records "remain to be verified against the evaluated code and run logs", and that was accurate.

The revised Appendix M is a status register. Each submitted item is restated and marked resolved, with its value, or open, with the claim it bears on:

  • M.1 RL configuration: values reported (G = 8, ε = 10⁻⁶, clip 0.2/0.2 with dual-clip 3.0, gradient-norm clip 1.0, one training seed). Update counts, seed stability and boundary-case frequencies are open, so §4.4 is now labelled a single-seed pilot.
  • M.2 Training state and events: open; bears on the RL pilot only.
  • M.3 Budget contract and guard: dashboard placement verified (below); guard counts and a guard-off control are open.
  • M.4 Auto-Archive configuration: resolved from request logs. It renders a status display every turn, so Table 1's State cell is corrected to partial. This changes the reading of the fixed-rule comparison (general response, §4.6).
  • M.5 Statistics and accounting: paired outcomes for the component arms (Appendix I.1) and repeated launches for three arms (§4.6) are reported; the BrowseComp-Plus 160K vs 247K/423K accounting is explained; the window statistic is withdrawn. Per-baseline cost breakdown and latency are open.
  • M.6 Mechanism controls: a fixed-rule controller with the status display is reported; a no-display fixed rule, a ledger-ranked fixed rule, and field-wise ablations are open.

What the completed records support. On LOCA at 128K, every VISTA launch we hold (34–39 solved) is above every no-dashboard launch (26–29). On the submitted launches the paired difference is 10 tasks (b/c = 2/12, [4.0, 22.7] points, exact McNemar p = 0.013, unadjusted). That is the abstract's ablation claim, and it holds on repeated launches.

W4. KV cache and cost: the cost premise does not hold on our measurements.

  1. The harness never stores the dashboard in the history. In all 290 sampled requests of a full-VISTA launch (75 tasks), it is the final message, appended per request, so re-rendering it leaves the cached prefix intact.
  2. The prefix is invalidated only when the context is edited. Table 6 counts the edits on the same 75 tasks: VISTA's agent makes 69 archive calls over roughly 2,700 requests (36.4 steps × 75), ReAct makes 636 trims, and tool-result clearing makes 1,968 clears. Guard offloads add to VISTA's count; their reconciled count is open (M.3).
  3. Measured total tokens per task: 2.86M for VISTA, 3.51M for ReAct, 6.72M for Claude Code (Table 5; VISTA's mean includes its 20 timed-out runs).

What we cannot report is a cache-weighted dollar cost. The provider reported cache reads for VISTA runs but returned none for ReAct runs, so a cache-adjusted comparison would be asymmetric. We would rather say so than quote a hit rate.

W3. Block granularity and summary quality. Archiving acts on whole blocks (one per observation, §3.1), but recovery does not: the agent can read an exact range of an archived payload, search it, or follow a nested handle (§3.2, "Archive and recovery" and "Hierarchical access"). A large tool result that mixes decisive and irrelevant content can therefore leave the active context whole and later return only the needed row. Figure 9 traces an inbox export archived and read back for the final action. On the burden of block-level operations: full VISTA makes 69 archive calls across 75 LOCA tasks, under one per task, and averages 36.4 steps per task against ReAct's 54.6 (Table 5). Summary quality is not isolated experimentally; Appendix J lists it as open.

W2 / Q2. Comparison set and novelty. We did not run PACE, Mem1 or HiAgent; Appendix J lists PACE, HiAgent and learned context managers as missing comparisons. The comparison set covers clearing, masking, summarization, agent self-compression, recoverable external storage (Auto-Archive) and a complete production agent (Claude Code), all in one harness. On novelty: MemGPT-style systems give agents paging tools and pressure warnings. Our no-dashboard ablation keeps every tool and removes only the per-block display. It costs about ten tasks on the matched launches, and every launch with the display is above every launch without it. That dissociation of display from tools is the contribution, and it is measured directly.

W1. Density. We will condense §3 and Figure 3 and move theory details to the appendix for the camera-ready.

The rebuttal does not address any of my concerns and I maintain my score.

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

VISTA helps language-model agents manage limited context without retraining. It exposes named context blocks, their sizes, and remaining space, and lets agents archive blocks while retaining access to their exact contents. Evaluation spans three online benchmarks, four model backbones, component tests, and context-pressure sweeps. Reported success is 50.7% on LOCA-Bench, 58.0% on BrowseComp-Plus, and 73.3% on GAIA. The paper also provides theoretical analysis, offline memory tests, and a reinforcement-learning extension.

Critical strengths
  1. Practical, training-free context management that transfers across backbones. The dashboard supports targeted context edits, while recoverable archives preserve exact evidence that summaries may lose. The method works without retraining on four backbones. This is important because it can be added to existing agents.

  2. Controlled same-harness evidence. Harness-native baselines fix the agent loop, tools, budget, backbone, and scoring. Only the context-management method changes. On LOCA, VISTA solves 38/75 tasks, compared with 33 for Auto-Archive + Recover, 27 for Active Compression, 22 for SLIM, and 17 for ReAct. Paired tests show significant gains over stale masking (Holm p = 0.002) and SLIM (Holm p = 0.007). Removing the dashboard lowers success on all three benchmarks: 50.7→37.3, 58.0→50.0, and 73.3→68.5. VISTA is also cheaper than Claude Code on LOCA: 2.86M tokens and 36.4 steps, versus 6.72M tokens and 171.5 steps. This is the paper’s strongest support for its context-management claim.

Critical weaknesses
  1. The gain is not shown to come from agent self-management, the paper's central claim.

    • Missing controls: There is no fixed controller using the same ledger. There is no guard-off run, although the guard offloads into the same store. The dashboard ablation removes all fields at once (J; M.3).
    • Without the dashboard, agent control trails a fixed policy: On LOCA, agent-controlled archiving without the dashboard scores 37.3%, below Auto-Archive's 44.0%. The rows may differ in more than the controller, and Auto-Archive's dashboard status is unresolved (M.4). Still, this points to the dashboard, not self-management, as the main driver.
    • No significant win over strong baselines: VISTA has significant gains only over stale masking and SLIM. Against Active Compression, Holm p = 0.087 (unadjusted 0.043). Against Claude Code, which uses a different harness, the 95% CI is [−4, 20]. Against Auto-Archive, there is no paired test. On GAIA, VISTA scores 73.3, below the best baseline's 73.9.
    • Close methods and novelty: There is no matched comparison with PACE or HiAgent. Context-Folding appears to have been run without its trained policy (p. 22). MemGPT already provides paging tools and memory-pressure warnings (p. 3), so the novelty is the per-block ledger, which is narrower than the title claims.
  2. **Key results are requires more strenghting **

    • RL setup. The RL stabilization settings are pending. Safe guard objectives are not clear: the objective also misbehaves at boundaries: penalties are divided by ε when every sample succeeds, and clipping erases differences between failures (L.1).
    • Evidence size. The 4-point RL gain over context-tool GRPO is about 6 of 150 questions. No uncertainty across seeds is reported. Most cells use one trajectory per task. Token totals are "pending verification," and guard counts are not reported.
    • Possible test-informed setting. The 12K BrowseComp-Plus window was chosen because it falls below the median first retrieval of a gold document. Whether that statistic came from development or test data is open (M.5).
Questions for authors
  1. Did Auto-Archive + Recover retain the dashboard and runtime guard? Please report its paired outcomes against VISTA. Do results exist for a fixed controller that uses the same ledger? If agent selection does not beat this controller, the contribution is the ledger, not self-management. How often did the guard offload content in VISTA runs, and how does a guard-off run perform at the same cap?

  2. How was the Context-Folding row implemented? Was it a trained policy or a prompt-level adaptation?

  3. What were the RL settings actually used: ε, G, the PPO clip values, the advantage cap, the gradient-norm rule, and the handling of all-success and all-failure groups? How often did each boundary case occur? Is the 4-point gain over context-tool GRPO stable across seeds?

Minor clarification: Was the 12K window chosen using development or test gold-document statistics? BrowseComp-Plus declares a 160K budget, yet Table 1 reports 247K and 423K trajectory totals. What does each figure count? (Appendices G.3 and M.5.) Also, the GAIA validation split has 165 questions, so a "random 165-question subset" (p. 25) appears to be the full split. Please confirm.

Suggestion (future work): Run matched comparisons with PACE, HiAgent, and MemGPT-style memory in the same harness. The main idea is understandable, but Figure 3 is cluttered and makes the context-management workflow harder to follow. A simpler layout with a clearer sequence of steps would improve readability.

3 comments
AuthorsOpenReview ↗

Response to Reviewer CqNu (1/2): the fixed-controller question

Thank you for a review that went straight to the question that matters. Our answer begins with a correction to Table 1, because it changes the premise of your central comparison.

Q1a. Did Auto-Archive + Recover retain the dashboard and the runtime guard? It kept a status display and the guard. It did not keep agent archive control. Its launch configuration keeps strict-budget mode (the guard) and the block store, disables the agent's archive and delete tools, and sets a fixed occupancy rule (target ratio 0.98) that externalizes old, large tool transcripts when a request would exceed the limit. The agent sees the default status renderer (budget bar, large visible blocks, recent blocks with token sizes and status) rather than the full factual ledger that VISTA renders. In 289 of 289 sampled requests across 75 tasks, that block is the final message. Those payloads come from the 37-solved launch; the 33-solved launch in Table 1 uses the same launch configuration. Table 1's State ✗ was therefore wrong, as your M.4 point anticipated. The cell is now ∼ (partial), and §4.6 and Appendix M.4 state the configuration.

What this does to your central inference. You compared the dashboard-free agent (37.3%) with Auto-Archive (44.0%) and concluded that the dashboard, not self-management, drives the gain. With the corrected label, that comparison sets an agent without state against a fixed rule whose agent still sees state. It shows that state matters, which is the paper's thesis; it does not show a state-blind rule beating the agent. The contrast you then asked for, agent control against a fixed controller with state shown in both, is VISTA against Auto-Archive:

LOCA, 128Ksubmitted + two relaunchesother launches
VISTA38, 38, 3934–38 (three with a development build)
Auto-Archive33, 32, 3237
VISTA w/o dashboard28, 26, 29none

Paired on the submitted launches (per-task records, Appendix I.1 procedure): VISTA vs w/o dashboard b/c = 2/12, +13.3 points, [4.0, 22.7], exact p = 0.013; VISTA vs Auto-Archive b/c = 6/11, +6.7, [−4.0, 17.3], p = 0.33.

Your test, answered at the strength the data supports. You wrote: "If agent selection does not beat this controller, the contribution is the ledger, not self-management." Agent selection leads on matched launch means (38.3 vs 32.3) and in the paired point estimate, but it is not separable at n = 75 per launch, and the 37-solved Auto-Archive launch sits one task below VISTA's 38. We therefore do not claim that agent control beats a fixed rule, and §4.6 says so.

We would push back on the dichotomy, though. In VISTA the ledger is not a rival explanation to self-management; it is the means of it. The title claims that state proprioception elicits context management from an untrained agent, and the abstract's ablation claim is that "the dashboard contributes beyond archive and recovery tools". Both hold across launches: every launch without the display (26–29 solved) is below every launch with it (34–39) and below every fixed-rule launch (32–37); with the display, the agent matches a tuned rule without needing one. We do not claim more than that, for example that agent choice helps only when state is visible, because that would need a fixed-rule-without-display cell we have not run.

Q1b–c. Auto-Archive's paired outcomes are above and in Appendix I.1. It is a fixed controller on the same block store and recovery path, but its rule does not rank blocks by the ledger's fields. A ledger-ranked fixed controller and a fixed rule with no display are not run (Appendix J, M.6).

Q1d. Guard frequency and guard-off. We do not yet have a reconciled per-trigger guard count, so we do not report one. The guard-off control is not run (M.3).

AuthorsOpenReview ↗

Response to Reviewer CqNu (2/2): baselines, novelty, RL, and setup questions

W2. Strong baselines. From the per-task records of the submitted launches: VISTA vs ReAct b/c = 1/22 (+28.0, [17.3, 38.7], p = 5.7×10⁻⁶); vs tool-result clearing 2/20 (+24.0, p = 1.2×10⁻⁴). Against the two strongest comparators the difference is not significant, and we say so: Auto-Archive (part 1) and Claude Code (+8.0, [−4.0, 20.0]). Claude Code is a complete system in its own harness; VISTA's point estimate is 8 points higher at 2.86M tokens and 36.4 steps per task, against 6.72M and 171.5 (Table 5; VISTA's means include its 20 timed-out runs).

W3. Novelty. We agree that the novelty is the per-block ledger together with recoverable payloads; that is what "state proprioception" in the title names. MemGPT provides paging tools and pressure warnings. Our no-dashboard ablation keeps every tool and removes only the per-block display; it costs about ten tasks on the matched launches, and every launch with the display is above every launch without it. Matched PACE and HiAgent runs are not done (Appendix J).

Q2. Context-Folding. It is not a trained policy. No checkpoint was available, so the row follows the published inference-time design at the prompt level, without a trained folding policy (Appendix H).

Q3. RL settings. Group size G = 8; ε = 10⁻⁶; PPO ratio clip 0.2/0.2 with dual-clip 3.0; gradient-norm clip 1.0; learning rate 10⁻⁶, weight decay 0.01; one PPO epoch; no KL loss; one training seed (Appendix M.1). Your boundary concern is correct, and sharper than the paper admitted. With ε = 10⁻⁶, a non-zero process penalty in an all-success group gives a shaped advantage of order 10⁵; ratio clipping does not bound it, and gradient-norm clipping bounds the resulting update. Pilot screening lowers the share of such groups but does not remove them, and we did not log boundary-case frequencies. With one seed, the 4-point gain (6 of 150 questions) carries no stability claim; §4.4 now labels the RL study a single-seed pilot. None of the frozen-model results depend on it.

Q4. The 12K window. We cannot show that the gold-retrieval-depth statistic was computed on development data, so we withdraw it as a justification. 12K is a retention stress setting applied identically to every method, not tuned per method, and §4.1 now says so. The headline LOCA results use a standard 128K window.

160K vs 247K/423K. These are two meters. B = 160K caps newly admitted tokens. The Traj column sums provider-reported tokens over all calls, so context re-sent on each call is counted each time. Totals can therefore exceed B without violating it (Appendix M.5).

GAIA "random 165-question subset". You are right: it is the full 165-question validation split (Levels 1/2/3: 53/86/26). "Random subset" was a wording error, corrected in §4.1 and Appendix I.

Figure 3. Agreed. We will simplify the workflow figure for the camera-ready.

I have read all the comments carefully. However, the responses do not address the concerns I raised, and I therefore maintain my original score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision, and one correction that changes the central comparison

We thank all three reviewers. The revised PDF marks every change in blue, and the main text remains within nine pages. We begin with a correction to Table 1, because it changes how the central comparison should be read.

1. Auto-Archive kept a status display. Table 1 said it did not. The submission marked Auto-Archive's State as ✗ and left its dashboard status as an open item (Appendix M.4). The request logs resolve it: in all 289 sampled requests of an Auto-Archive launch (75 tasks), a <context_workspace_status> block with the budget bar and per-block token sizes is the final message. The arm disables the agent's archive and delete tools and lets a fixed occupancy rule externalize old, large tool transcripts; the agent still sees a status display and can recover payloads. The State cell is now partial (∼), and §4.6, Appendix J and Appendix M state what the arm contains.

2. What the repeated launches show (LOCA-Bench, 128K, §4.6).

armsubmitted + two relaunchesother launches
VISTA (agent + ledger)38, 38, 3934–38 (three with a development build of the policy)
Auto-Archive (fixed rule + status display)33, 32, 3237
VISTA w/o dashboard (agent, no display)28, 26, 29none
  • The display. Every VISTA launch is above every no-dashboard launch. Paired on the submitted launches with the paper's own Appendix I.1 procedure, the gap is 10 tasks: b/c = 2/12, task-pair 95% interval [4.0, 22.7] points, exact McNemar p = 0.013 (unadjusted).
  • Agent versus fixed rule, both with a display. The agent leads on matched launch means (38.3 vs 32.3 solved), but the paired test on the submitted launches does not separate them (b/c = 6/11, [−4.0, 17.3], p = 0.33), and the other Auto-Archive launch (37) is one task below VISTA's submitted 38.

Our reading: removing the display costs the same agent about ten tasks and puts it below a fixed rule; with the display, the agent matches that rule, with point estimates in its favour, and needs no hand-set archive rule. That is what the title and abstract claim (explicit state elicits context management from an untrained agent; "the dashboard contributes beyond archive and recovery tools"), and we claim no more. In particular, we do not claim that agent choice pays off only when state is visible: that would need a fixed-rule-without-display cell, which we have not run.

3. Other changes.

  • Appendix M is now a status register: each submitted item is marked resolved, with its value, or open, with the claim it bears on. M.1 gives the RL configuration.
  • The 12K BrowseComp-Plus window is described as a retention stress test. We withdraw the gold-retrieval-depth statistic as its justification, because we cannot show it was computed on development data.
  • GAIA uses the full 165-question validation split; "random subset" was a wording error.
  • The RL study is labelled a single-seed pilot.
  • Appendix I.1 gives realized paired outcomes for the component arms and Auto-Archive.
  • Appendix J reports a fifth backbone we had run but not plotted (GPT-5.2: VISTA 9.3%, ReAct 8.0%).

4. Not done. A guard-off control, a fixed rule without any display, a ledger-ranked fixed rule, matched PACE/HiAgent/learned-manager comparisons, and a second RL seed. Appendices J and M list each as open.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.