PaperlayerView exact revision on arXiv

Source-linked AI summary

FocusMem: Factorizing Content, Readout, and Trust in Latent GUI Memory

Zhuoran Zhang, Bowen Li, Jingcheng Ju, Yang Shi, Qixun Wang, Haotian Wang, Wei Chen, Tengjiao Wang

arXiv:2608.04530v1cs.CV

TL;DR

GUI agents need latent memories that preserve role-specific evidence, expose decision-relevant information, and avoid letting irrelevant trajectories mislead actions. FocusMem factorizes these responsibilities and consistently improves performance across five GUI-agent benchmarks, including a 13.5-point gain over Action-only Fixed.

  • Problem

    Existing latent GUI memory maps trajectories to fixed, state-independent blocks, leaving content recoverability, decision-specific exposure, and evidence relevance unresolved.

  • Method

    FocusMem learns role-aware content bases, state-conditioned readouts, and a trust gate while keeping the GUI policy frozen.

  • Results

    Across five GUI-agent benchmarks, FocusMem consistently improves task success, outperforming Action-only Fixed by 13.5 points.

  • Takeaways & Limitations

    Compact latent memory benefits from separating what is retained, what is exposed, and what is allowed to influence the policy.

  • Takeaways & Limitations

    Evaluation primarily uses a frozen Qwen3-VL-8B policy on web benchmarks, leaving transfer to other backbones and mobile or desktop environments open.

Abstract

from arXiv · show

GUI agents must remember both useful experience from earlier tasks and unfinished progress in the current interaction. Latent memory offers a compact solution by compressing multimodal trajectories into a few continuous tokens. Existing methods, however, usually map each trajectory to one fixed memory block and train it mainly through next-action supervision. This creates three practical problems: important details may be lost during compression, the same memory block must serve different decision stages, and irrelevant retrieved trajectories may still mislead the agent. We introduce FocusMem, which separates these responsibilities within a compact latent-memory interface. A role-aware content basis encourages episodic memory to retain reusable experience and working memory to retain task progress. A state-conditioned readout generates a decision-specific view of the same stored evidence, while a lightweight trust gate can suppress memory blocks that appear irrelevant to the current step. All components are trained while the GUI policy remains frozen. Across five GUI-agent benchmarks, FocusMem consistently outperforms a fully matched action-only fixed-memory baseline and prior latent memory adaptations. Further analysis shows that semantic and functional supervision preserve complementary information, state-conditioned readout is more robust as surrounding trajectory context grows, and the trust gate reduces the harm caused by injected irrelevant episodic evidence. These results show that effective latent memory depends not only on compressing past interaction, but also on what is retained, what is exposed, and what is allowed.

1. Introduction

FocusMem factorizes latent GUI memory into three responsibilities: retaining role-appropriate content, exposing decision-relevant evidence, and controlling whether retrieved evidence influences the policy. Across five benchmarks, it improves task success while keeping memory compact, with analyses attributing gains to memory formation, state adaptation, and trust control.

  • Motivation: GUI agents need episodic memory for reusable completed-task evidence and working memory for current constraints and progress, but replaying full trajectories is costly and text can lose visual or procedural details.Repeated screenshots and action traces expand policy context, while verbalization may omit information that is difficult to describe textually.
  • Motivation: Existing latent-memory methods typically compress each selected item into a state-independent block trained mainly through downstream action supervision, conflating recoverability, exposure, and trust.The introduction identifies these as distinct responsibilities: what survives compression, what is exposed under the current state, and whether selected evidence should enter the policy.
  • Method: FocusMem replaces fixed latent blocks with a factorized interface comprising a role-aware content basis, state-conditioned readout, and independent trust gate.Episodic content targets reusable experience, working content targets current constraints and progress, readout adapts exposure to the decision state, and the gate controls policy influence.
  • Evaluation: Across five GUI-agent benchmarks, FocusMem consistently improves task success over no-memory agents, raw-trajectory replay, and prior memory methods while keeping policy-facing memory compact.The evaluation uses a common experimental setup and also tests recoverability, state-conditioned readout under growing context, and suppression of misleading irrelevant evidence.
  • Analysis: The reported analyses indicate that FocusMem’s benefit comes from how latent memory is formed, adapted to the current decision, and controlled before use rather than simply from providing more past interaction.They examine information recoverability after compression, decision-relevant readout under a fixed latent budget, and protection against irrelevant retrieved experience.

2. Related Work

Prior GUI-agent memory spans textual, structured, history-aware, executable, and latent representations, while related compression work improves conditional extraction or bottleneck recoverability. FocusMem builds on these directions by separating retained content, state-conditioned readout, and memory influence within a compact latent interface.

  • GUI-agent memory: GUI-agent memory includes textual workflows and lessons [2] [19] [27], structured multimodal records [22] [32] [37], history-aware reasoning, executable routine graphs [11], and latent trajectory embeddings [28] [34].CoMEM [28] encodes retrieved experience, whereas Mem-W [34] shares a fixed compressor across episodic and working memory.
  • Conditional and recoverable latent compression: Conditional compression extracts context-relevant information [5] [21], while reconstruction or distillation improves bottleneck recoverability [17] [3].VisMem [31] and MemGen [33] extend adaptive latent-memory study beyond GUI trajectories; FocusMem combines recoverable content supervision with state-conditioned generation.
  • Learning and analyzing GUI memory: Recent GUI-memory studies examine what and when agents memorize, how records become task-driving states, and failures of naive visual memory.FocusMem addresses these concerns through content probes, evidence expansion, and contamination analysis within a compact latent interface.

3. Method

FocusMem builds compact GUI memory from retrieved episodic trajectories and segmented expired history, then separates what memory retains, exposes, and admits to policy attention. It uses role-aware content supervision, state-conditioned readout, and block-level trust gating while keeping the GUI policy frozen.

  • Evidence Construction: Memory evidence combines top-M complete trajectories retrieved by cosine similarity with nonoverlapping chunks of the expired history, giving each item the same K-token latent budget.The retriever is fixed across matched variants and is not optimized through the memory pathway; both roles share an evidence encoder and Q-Former-style compressor projected to the frozen policy dimension d.
  • Role-Aware Content Basis: FocusMem assigns role-specific query bases so episodic memory preserves reusable decisions and constraints, while working memory preserves progress, state changes, and remaining constraints.Semantic supervision transfers broadly verbalizable content, while functional supervision emphasizes role-specific control-relevant information; only the memory pathway is updated.
  • Optimization: Training proceeds in two stages: Stage A learns fixed role-aware content, then Stage B enables conditional readout and trust while updating the memory pathway and keeping the GUI policy frozen.Accepted dynamic blocks are prepended to ordinary context embeddings, while rejected blocks are omitted; Stage B also updates the shared compressor and tests content recoverability from fixed blocks.
  • State-Conditioned Readout: A state–item adapter conditions each latent block jointly on the current decision state and encoded evidence, allowing the same memory to expose different decision-facing information across states.The adapter adds a residual over role-specific base queries and is zero-initialized to preserve the learned fixed readout at optimization start.
  • Evidence Trust: A block-level trust module scores decision-facing memory blocks and removes rejected blocks from policy attention, with threshold γ selected on validation data.Training uses a straight-through HardConcrete mask and injects high-confidence irrelevant episodic items from unrelated tasks; gate gradients do not directly reshape block content.

4. Experiments

Across five GUI-agent benchmarks, FocusMem achieves the best overall interactive task success while using a frozen policy and a compact latent-memory interface. Ablations and diagnostics show that role-aware content supervision, state-conditioned readout, and evidence gating each contribute distinct benefits, especially under growing or contaminated context.

  • Main results: FocusMem achieves the highest Overall success rate across all five benchmarks and consistently outperforms the fully matched Action-only Fixed baseline and other memory baselines.The comparison uses a common 15-step protocol with matched policy, training data, memory bank, retrieval, latent-token and action budgets, grounding, decoding, and evaluator.
  • Main results: Relative to system-level references, FocusMem improves over No Memory by nearly 20 points on average, Retrieved Full Trajectory by about 15 points overall, and Mem-W-style [34] by 4.5–8.0 points.It also surpasses CoMEM-style [28], HyMEM-style [37], and Mem-W-style [34] on every benchmark.
  • Component ablation: Cumulative ablations show successive SR gains of +4.5, +5.0, +4.0, and +4.5 points from fixed memory, Content Basis, Dynamic Readout, and Evidence Gate.The components add nontrivial value beyond fixed action-only compression, with full FocusMem 18.0 points above No Memory.
  • Content-basis analysis: Combining semantic and functional supervision produces the best content-basis recoverability on every metric for both episodic and working memory.Semantic supervision mainly improves general trajectory reconstruction, whereas functional supervision yields larger gains on role-aligned and unseen transfer queries.
  • Dynamic readout: Dynamic Readout is substantially more robust than Fixed memory as surrounding trajectory context grows, with drops of only 12, 9, and 6 points versus Fixed drops of 22, 25, and 25 at B = 4, 8, 16.The methods perform nearly identically when given only the concise Oracle Core, so the advantage emerges under expanded evidence.
  • Evidence gating: Under maximum injected episodic contamination, full FocusMem remains 9–11 points stronger than the no-gate variant, while Dynamic Readout benefits from retrieved values with intact key–value correspondence.Removing evidence or breaking key–value correspondence largely eliminates the readout gain.

5. Limitation

FocusMem’s evaluation is limited to a frozen Qwen3-VL-8B policy on web benchmarks, while transfer to other backbones and mobile or desktop environments remains untested. Its trust and scoring diagnostics also leave subtler memory mismatches and judge bias unresolved.

  • 5. Limitation: Evaluation remains limited to a frozen Qwen3-VL-8B policy on web benchmarks, leaving transfer to other backbones and mobile or desktop environments open.
  • 5. Limitation: The trust diagnostic tests injected irrelevant episodic trajectories but does not cover subtler mismatches or stale working memory.
  • 5. Limitation: All scores rely on Gemini-3.1-Pro; fixed, method-blind prompts reduce but do not eliminate judge bias.

6. Conclusion

FocusMem factorizes latent GUI memory into content recoverability, state-conditioned readout, and evidence trust. Diagnostics and matched interactive evaluation support these roles and show consistent gains over fixed action-only compression and prior memory adaptations.

  • 6. Conclusion: FocusMem factorizes latent GUI memory into content recoverability, state-conditioned readout, and evidence trust.This separates what is retained, what is exposed, and what evidence is trusted.
  • 6. Conclusion: Content probes, oracle-centered expansion, and controlled evidence and contamination diagnostics support the intended roles of these components.
  • 6. Conclusion: Matched interactive evaluation shows consistent gains over fixed action-only compression and prior memory adaptations.

GUI Memory Supplementary Material … C. Data, Memory Bank, Splits, and Leakage Prevention

The supplementary material specifies FocusMem’s role-aware latent-memory pipeline, including fixed compression, content formation, state-conditioned readout, and evidence trust. It also documents supervision prompts, trainable pathways, and separate action-training and episodic-memory data sources with online working memory.

  • A.1. Evidence Encoding and Fixed Readout: Each episodic trajectory or expired working-memory chunk is compressed into role-specific latent queries, with masking and role embeddings preserving valid-token and memory-role information.The shared compressor processes multimodal trajectory features; padding is excluded from cross-attention, and role embeddings mark resulting latent blocks.
  • A.2. Stage A: Role-Aware Content Formation: Stage A trains a role-aware content basis using semantic distribution matching and functional supervision while keeping the GUI policy frozen.Semantic supervision isolates information lost through the latent bottleneck, while functional targets are role-specific and combined as Lbasis = λsemLsem + λfuncLfunc.
  • A.3. Stage B: State-Conditioned Readout and Evidence Trust: Stage B conditions latent queries on the decision state and role, then applies a block-level trust gate to admit or reject evidence before policy input.The dynamic readout is initialized to match Stage A, while the gate uses validation-selected thresholding and discards rejected, padding, or empty blocks.
  • A.4. Trainable Pathway Configurations: The trainable pathway adapts only the compression backbone with LoRA, while the independent policy copy remains frozen and carries no adapters.Table 4 excludes the vision tower, token embeddings, language-model head, Q-Former, and frozen policy copy from LoRA adaptation.
  • B.1. Raw Evidence Serialization: Memory items are serialized chronologically as screenshot-action turns, with complete retrieved trajectories forming episodic items and expired-prefix chunks forming working-memory items.Headers identify the task when applicable, memory role, and number of steps; teacher prompts use raw serialization, whereas student prompts use compressed representations.
  • B.2. Semantic Distribution Prompt: Semantic supervision matches the frozen policy’s full raw next-token distribution, while functional supervision covers four role-specific QA types spanning episodic experience and working-memory state changes.The four functional subtypes are high-level experience, anchor recovery, latest-state recovery, and observed-change recovery.
  • B.3. Teacher–Student Separation: Teacher–student separation generates all pseudo-targets from raw evidence with a frozen teacher, while latent students receive the optimization signal without updating the teacher.Action imitation scores only demonstrated assistant actions, whereas semantic supervision retains the complete raw-policy distribution at every description position.
  • C. Data, Memory Bank, Splits, and Leakage Prevention: FocusMem uses an action-supervision corpus for training memory components and a separate episodic retrieval bank, while working memory is constructed online from the current episode’s expired prefix.The episodic bank serves retrieval during both training and evaluation; working-memory evidence is not stored in that bank.

C.1. Evaluation Benchmarks and Accessible Subsets · C.2. Trajectory Data Construction · C.3. High-Quality Selection

The evaluation uses full official subsets for locally hosted MMInA suites and screened accessible subsets for live-web benchmarks, with standardized trajectory judging. Training and episodic data come from disjoint successful in-domain trajectories built by rewriting existing data or generating fresh rollouts, then filtered for quality.

  • C.1. Evaluation Benchmarks and Accessible Subsets: All benchmarks use fixed task subsets and benchmark-specific success judges instantiated with Gemini-3.1-Pro, evaluating only trajectories that emit an explicit final answer within 15 steps.Trajectories reaching the step cap without producing an answer are scored separately from answer-emitting trajectories.
  • C.1. Evaluation Benchmarks and Accessible Subsets: Evaluation covers full official MMInA subsets and screens live-web tasks for external obstructions such as unreachable sites, login walls, CAPTCHAs, WAF blocks, redirects, and modal dialogs.An LLM auditor distinguishes website-imposed obstruction from agent failure using fixed categories and a deterministic output format.
  • C.2. Trajectory Data Construction: Training and episodic-bank trajectories are successful, in-domain, website-sharing, and disjoint from evaluation tasks, ensuring the stored experience does not duplicate evaluated tasks.The data are constructed through two routes: rewriting existing trajectories or generating fresh rollouts.
  • C.2. Trajectory Data Construction: The rewrite route adapts WebVoyager and Online-Mind2Web trajectories from MolmoWeb, quality-filtering them before regenerating concise per-step state descriptions while preserving actions and screenshots.The rewrite prompt conditions descriptions on the task, prior action history, and current context, and forbids external hints.
  • C.2. Trajectory Data Construction: The generate-and-rollout route synthesizes MMInA-Shopping, MMInA-Wikipedia, and DeepShop data from task seeds followed by fresh trajectories.Shopping seeds use scraped real products and task types, while Wikipedia seeds target self-contained single-article factual or visual questions.
  • C.3. High-Quality Selection: High-quality selection applies a three-stage pipeline combining rule-based cleanup, LLM rubric scoring, and a final selection stage.The rule layer removes unparseable actions, missing explicit stops, adjacent duplicate actions, and other degenerate patterns; the rubric assesses correctness, necessity, coherence, monotone shape, and termination discipline with hard rejection caps.

C.4. Dataset Splits … D. Evaluation and Annotation Protocols

The paper uses disjoint, carefully represented datasets and leakage controls, then standardizes automated evaluation and human annotation protocols across compared variants. These procedures define the data boundaries, serialization, judging, and annotation conditions for the reported experiments.

  • C.4. Dataset Splits: Table 6 defines the principal dataset splits, their purposes, and disjointness relationships with evaluation tasks under a three-level match.The three levels are specified as task identifier, annotated instance, and slot-normalized instruction template.
  • C.5. Leakage Prevention: The memory sources and evaluation tasks are kept non-overlapping by task identifier, annotated instance, and slot-normalized instruction template.The paper applies these exclusions during training-set construction and again during training and evaluation.
  • C.5. Leakage Prevention: Training-set construction removes trajectories matching an episodic-bank or evaluation task under any of the three leakage keys.This is the first stated point at which overlap is prevented.
  • C.6. Data Representation: A task seed is defined as a start-URL and task pair, while a trajectory records the task, observed screenshots, and actions at each step.This representation distinguishes episode initialization from the stepwise interaction record.
  • C.6. Data Representation: Actions are serialized as JSON objects containing only arguments relevant to each action type, as illustrated in Figure 13.Screenshots are stored as PNG files at the policy input resolution.
  • C.6. Data Representation: The data representation removes URLs, account identifiers, and personal information from stored records.This condition is stated alongside the screenshot and action-storage conventions.
  • D. Evaluation and Annotation Protocols: Evaluation reports metrics, task-completion and diagnostic judgments, diagnostic controls, run and parsing protocols, and human annotations for oracle-centered and contamination experiments.Automated judgments use Gemini-3.1-Pro at temperature 0 with fixed prompts, and judge inputs omit method, checkpoint, and intervention identifiers.

D.1. Metrics · D.2. Automated Judges

The evaluation defines task, action, content, rollout, timing, and gate metrics, with success judgments automated by specified judges. Episodic and working-memory panels use separate sampled manifests while maintaining fixed items within each role.

  • D.2. Automated Judges: Task success and one-step action validity are determined by their corresponding automated judges.
  • D.2. Automated Judges: Content-probe scores are assigned by a diagnostic content judge.
  • D.1. Metrics: Hit-Max, executed steps, wall-clock time, and gate passage are computed directly from rollout and gate logs.
  • D.1. Metrics: The metric suite covers task success, one-step action validity, content-probe scores, Hit-Max, executed steps, wall-clock time, and gate passage.These metrics are defined in Table 7 and used across the reported evaluations.
  • D.2. Automated Judges: All success judgments use Gemini-3.1-Pro.Table 7 specifies Gemini-3.1-Pro (§D.2) as the judge for success evaluations.
  • D.1. Metrics: The three metric categories use the same fixed item manifest within each role, while episodic- and working-memory panels use separate sampled manifests and are reported independently.

. Task-Completion Judges … . Evidence-Dependence Interventions

The paper specifies benchmark-specific task-completion judges and method-blind diagnostics for evaluating latent-memory content, action validity, oracle-centered evidence use, and evidence dependence. These evaluations isolate role-relevant information, valid next-step behavior, and the effects of replacing or mismatching retrieved episodic evidence.

  • . Task-Completion Judges: Task-completion evaluation uses fixed Gemini-3.1-Pro judges, combining text-only fuzzy matching, screenshot-grounded assessment, and Online-Mind2Web’s staged WebJudge.MMInA compares final answers with gold answers from text alone; DeepShop and WebVoyager use the final answer plus five screenshots; Online-Mind2Web extracts key points, scores screenshots, and issues a verdict without reading the final answer.
  • . Content-Basis Probe Generation: Content-basis probes expose one fixed latent block with its role and instruction, while withholding raw trajectories, current state, dynamic readout, trust gating, and teacher targets.Held-out aligned probes use training categories with new wording and formats, whereas transfer probes query categories absent from training.
  • . Content-Basis Probe Judge: A method-blind content judge scores candidate answers from 0 to 100 against raw chronological evidence, emphasizing factual grounding and role-specific recovery for semantic, aligned, and transfer probes.Aligned probes target episodic procedure and anchors or working-memory latest state and observed change; transfer probes assess remaining subgoals or harmful repetition without unsupported advice.
  • . One-Step Action-Validity Judge: The one-step action-validity judge accepts any well-formed, visible, state-consistent, task-constrained action that reasonably advances completion, rather than requiring exact reference-action agreement.Malformed actions, absent-element targets, unsupported stop answers, and repetitions of failed actions are invalid.
  • . Content-Basis Probe Sets: Probe manifests sample held-out WebVoyager tasks without replacement and retain at most one evidence item per task, preventing longer trajectories or extra decision points from dominating evaluation.The episodic manifest contains 200 successful trajectories, while the working-memory manifest is constructed separately under the same sampling principle.
  • . Evidence-Dependence Interventions: The evidence-dependence intervention keeps the task, policy context, checkpoint, and decoding fixed while changing only the three episodic memory slots’ supplied evidence.It compares original retrieval, no evidence, coherent same-domain donor retrieval, and original-key/shuffled-value mismatches; the latter breaks correspondence between query conditioning summaries and token-level evidence.
  • . Evidence-Dependence Interventions: The intervention distinguishes Dynamic Readout’s pooled conditioning key from its token-level evidence value, testing whether state-conditioned queries use trajectory content coherently or are misled by task-mismatched evidence.Shuffled Retrieval preserves coherent key–value pairs from another same-domain task, whereas Original Keys + Shuffled Values deliberately mismatches the summary and evidence components.

. Irrelevant-Trajectory Contamination … F. Training Details and Hyperparameter Search

The evaluation tests FocusMem under nested injection of manually verified irrelevant trajectories, compares trust-gating configurations, and reports benchmark-specific success and gate-passage measures. The supplied passages also specify single-run reporting, human verification of negatives, and the software/hardware environment, but provide no training-details or hyperparameter-search content.

  • . Irrelevant-Trajectory Contamination: The contamination study uses independent 100-task subsets with full interactive rollouts under the standard 15-step action budget and benchmark-specific success judges.These subsets are excluded from the oracle-centered and one-step evidence-dependence diagnostics.
  • . Irrelevant-Trajectory Contamination: Contamination replaces exactly q ∈ {0, 1, 2, 3} retrieved candidates, producing nested contamination ratios of 0%, 33%, 66%, and 100%.Higher levels retain lower-level replacements and progressively remove higher-ranked original candidates; replacements are fixed in a contamination manifest and are designed to remain visually and procedurally plausible where possible.
  • . Irrelevant-Trajectory Contamination: FocusMem is evaluated against No Gate and Working-Only under progressively higher contamination, with the full model using the validation-selected threshold γ = 0.3.No Gate admits every non-empty episodic and working block, while Working-Only removes episodic blocks but preserves working memory.
  • . Irrelevant-Trajectory Contamination: Interactive success rate is computed separately for WebVoyager and Online-Mind2Web, while gate passage is reported only for full FocusMem as a pooled micro-average over non-empty episodic candidates.Padding slots and missing candidates are excluded, and the clean q = 0 passage rate is well-defined.
  • D.4. Run Protocol: Main results and diagnostics are reported from a single interactive rollout unless otherwise stated, without implying replication for unreplicated numbers.This applies to Table 1 main results, Table 3 diagnostics, and Figs. 3–4.
  • D.5. Human Annotation Protocol: The contamination diagnostic uses high-confidence irrelevant negatives from unrelated tasks, selected and human-checked as clearly inapplicable to the current query.This protocol is intended to test whether a trust gate rejects injected trajectories.
  • E. Software and Hardware Environment: All experiments ran on a single server, with the software stack and hardware configuration summarized in Tables 10 and 11.Table 9 also summarizes configurations for the main evaluation and selected diagnostic experiments.

F.1. Training Details … G.3. Case 3: State-Conditioned Readout (Fixed vs. Dynamic Query)

FocusMem is trained by optimizing only its memory pathway while keeping the GUI policy and grounding model frozen, with settings selected through held-out validation. Qualitative cases show that its factorized interface recovers reusable content, preserves working progress, and uses state-conditioned readout to avoid unproductive re-searching.

  • F.1. Training Details: FocusMem trains only the memory pathway, using 23,438 successful 8–15-step trajectories, staged learning rates, and a global batch size of 32 while the policy and grounding model remain frozen.Stages A and B use learning rates of 1 × 10−5 and 5 × 10−6, respectively.
  • F.2. Hyperparameter Search: Final hyperparameters were selected on a held-out validation split after varying learning rates, batch sizes, and trust thresholds.The evaluated learning-rate candidates were 1×10−6, 5 × 10−6, 1 × 10−5, and 5 × 10−5; batch sizes were 16, 32, and 64; and γ ranged from 0.1 to 0.5.
  • G. Qualitative Case Studies: The qualitative analysis uses genuine runs of the same frozen Qwen3-VL-8B policy under two memory conditions, showing actual screenshots and per-step actions without staged or edited trajectories.For clicks, red dots mark grounding-resolved pixels and coordinates are printed beneath the relevant steps.
  • G.1. Case 1: Content Recoverability (Episodic Reusable Experience): Figure 21 shows FocusMem answering the coffee-mug color question in 5 steps, whereas No Memory times out after 15 steps without an answer.The No-Memory run repeatedly scrolls the listing, while FocusMem opens the product on its third step.
  • G.2. Case 2: Working Memory Retains Progress: Figure 22 shows FocusMem correctly retaining the first product’s attribute and answering the comparison in 15 steps, while No Memory times out before reaching either product.The case requires preserving product A while navigating to inspect product B.
  • G.3. Case 3: State-Conditioned Readout (Fixed vs. Dynamic Query): Figure 23 shows dynamic-query readout answering the Nanning question correctly in 15 steps, while fixed-query readout repeatedly re-searches and produces no answer.The matched checkpoints differ only in query mechanism, and the divergent action is re-typing versus scrolling without grounding involvement.
Loading 2608.04530v1…