{"id":"dca5a583-0a72-49d8-b304-4a09f5929129","arxiv_id":"2610.01787","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Component routing sends each of four experience types to the weights or the prompt by recurrence and state-conditionality, measured before training, beating whole-trajectory reuse across three backbones and two environments.","lead":"GUI agents that reuse their own past tasks store experience in different places for different kinds of knowledge: navigation hints and correction lessons pay off inside the model's weights, while step procedures and screen facts are better served in the prompt. The authors show two simple pre-training features, how often an item recurs and how much it depends on the current screen, decide the destination and improve held-out success over reusing whole trajectories.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The rule's held-out recovery rests on a single near-boundary split (state facts vs lessons) whose weights-side signs are soft (three of six lesson intervals cross zero, S=21.5/24), and kappa0 — only Spearman 0.56 against online state-conditionality — may contribute nothing beyond recurrence…","rationale":"Three candidate concerns were weighed and rejected before settling on this one. (1) The routing advantage being a measured artifact: the routed arm is actually trained and evaluated with its assignment fixed before any adapter is trained, evaluations are paired over (template, instance, seed), and the paper reports per-cell intervals; the +3.5-point average over the better single destination is a measured fact, softened only by the one tie cell (AndroidWorld/GUI-Owl routing 33.3 vs all-to-context 33.6, [−2.6,+2.0]), which the paper honestly reports as an 'on average' claim. (2) Budget inequality between the notes tokens and adapter training: the paper explicitly reports both accounts in Appendix H without equalizing them (sentence b374), so this is an acknowledged scope condition, not a hidden flaw. (3) The near-boundary weak point: this is the genuine soft spot, and it is where the reader's CONDITIONAL already gestures; my contribution is to decompose it. The decisive pair is state facts vs lessons, the weights-side lesson signs are within noise of zero in half the cells, and kappa0 may be redundant with recurrence, which alone reproduces the full 24-of-24 assignment. That decomposition changes the character of the risk. If recurrence alone suffices, the weak proxy is irrelevant and the rule is more robust than the paper argues (one property, directly measured), but then the two-property explanation and the boundary figure overstate mechanism. If kappa0 is required and load-bearing, then the boundary placement at the only genuinely decided cells rests on a proxy with 0.56 rank agreement and on near-zero lesson gains, making the 24-of-24 recovery a semi-tautological re-assertion of cross-family sign consistency with real predictive content concentrated on a near-degenerate pair. Both resolutions are informative and both are reachable by one cheap refit rather than by new training runs, which is why the recurrence-only decomposition (and, on artifact release, a fourth family or third environment) should be a condition before the verdict is lifted. Weighing all this, the reader's CONDITIONAL is appropriate and my concern does not move it, so the verdict remains UNCHANGED in substance while being expressed as CONDITIONAL.","tokens_in":40755,"tokens_out":11485,"duration_ms":70169,"concrete_test":"Refit the three-fold held-out procedure of Eq. 5 with kappa0 removed, i.e., score s(c) = alpha·r(c) − gamma only, using the same logistic objective, ridge penalty, and fold structure (Appendix D.1–D.2), and compare: (i) the 24-cell sign recovery and weighted score S; (ii) the routed-arm held-out success versus the better single destination on the same 1106/232 instances with the same paired bootstrap. If recurrence-only attains the same 24/24 (S) and the same routed advantage within interval noise, then kappa0 is not load-bearing: the near-boundary concern is resolved, but the paper's two-property explanation should be restated as a one-property rule and the Spearman-0.56 proxy weakness is moot.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The routing gain of the measured routed arm is real (the arm was actually trained and evaluated with its assignment fixed to disk before training, and all differences are paired over instances and seeds), so my concern is not with the +3.5-point average advantage but with the strength of the explanation that accompanies it: that a rule in two pre-training properties, recurrence r and state-conditionality kappa0, recovers the destination of a held-out backbone family in 24 of 24 cells (Eq. 5, Table D.2). Decompose where that recovery actually decides. Locators (full-fit score ≈ +2.5) and procedures (≈ −4.4) sit two to four score units from the boundary and carry the identical sign in all three families; the fit adds no information for them. The rule's only decision content is placing lessons (score +0.74) on the weights side and state facts (−0.89) on the context side, a margin of ≈1.6 units versus 2.5+ elsewhere. On exactly that decisive pair the weights-side observation is softest: three of the six lesson cells have 95% intervals that include zero — GUI-Owl MobileGym +0.7 [−0.3,+1.6], GUI-Owl AndroidWorld +0.9 [−1.0,+2.7], ScaleCUA AndroidWorld +0.6 [−1.7,+2.7] — each counted at half weight, which is why the held-out score is S=21.5/24 rather than 24. Second, the two-property account is unverified in its second dimension. Recurrence alone orders all four components correctly (procedures 0.05, state facts 0.31, locators 0.39, lessons 0.44), so a recurrence-only threshold near r≈0.35 would reproduce the identical 24-of-24 assignment; kappa0's only online confirmation is Spearman 0.56 (Eq. 3 vs Eq. 4), and its near-boundary contribution is never isolated. Two readings follow. If kappa0 is non-load-bearing, the weak-proxy worry is moot but the claimed 'two properties' mechanism is not established and the paper effectively reports one property.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proposes 'component routing' for self-improving GUI agents: instead of treating whole trajectories as the unit of experience, it splits each trajectory into four component types (locators G, procedures P, state facts F, lessons L), measures two pool-level properties of each component (recurrence r and state-conditionality κ0, both available before any training), fits a linear rule in these two properties to the observed signs of route gains (weights-minus-context) with one backbone family held out, and reports: a leave-one-family-out sign recovery of 24/24 cells (weighted S = 21.5/24 because five cells have intervals crossing zero); routing by the rule beating every whole-trajectory baseline and, on average, the better single destination by +3.5 points; an eightfold training dose moving procedures toward the weights; a state-conditioning ablation costing 3.4–3.9 points of context gain; and analyses of note readout after training and of producer–consumer information/policy gaps. The measurements are paired over instances and seeds, with the analysis fixed before the runs and a prospective half-split test whose destinations were fixed before training.","tokens_in":41265,"tokens_out":9879,"duration_ms":60839,"significance":"The empirical protocol is a strength: route gains are paired per instance, the analysis was fixed before the runs, the routed arm's assignment was written to disk before training, and the authors transparently report that five of twenty-four cells have intervals crossing zero (S = 21.5/24). If the routing comparison holds, the paper makes a concrete contribution by showing that the destination of experience depends on component type rather than only on pipeline choice, and the +3.5-point average advantage over the better single destination is a reproducible, falsifiable result. The main caveat is the strength of the explanatory rule: the leave-one-family-out recovery is dominated by components far from the boundary, and the only borderline pair (state facts vs lessons) is decided almost entirely by recurrence, with κ0 contributing little discriminative information. The core routing measurement is therefore credible, but the headline 'rule in two properties' claim needs reframing or additional evidence before it can be taken at face value.","major_comments":[{"comment":"The held-out 24/24 sign recovery is not uniformly discriminating: locators (full-fit score ≈ +2.5) and procedures (≈ −4.4) sit several score units from the boundary and carry identical signs across all three families, so twelve of the twenty-four cells are recovered by any rule that places G on the weights side and P on the context side. The only genuinely informative pair is state facts (score ≈ −0.89) versus lessons (≈ +0.74), and there the weights-side evidence is the softest in the paper: three of the six lesson intervals include zero, which is exactly why the authors' own weighted held-out score is S = 21.5/24 rather than 24. The headline 'recovers the destination in 24 of 24 cells' should therefore be accompanied by a decomposition of how many cells the boundary actually decides, e.g., a per-pair held-out score or a statement that the decisive pair carries a margin of only ~1.6 units with the weakest observation.","section":"§5.1, Eq. (5), Table D.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Recurrence alone orders the four components as P (0.05) < F (0.31) < G (0.39) < L (0.44), so a horizontal threshold on r between 0.31 and 0.39 yields the same assignment (F, P → context; G, L → weights) for every family, and the decisive F-versus-L pair differs chiefly in recurrence (r ≈ 0.31 vs 0.45) while its κ0 values nearly coincide (≈ 0.53–0.56). The paper never reports a recurrence-only held-out fit, so the claimed 'rule in two properties' is not established: κ0's incremental contribution to any routing assignment is unquantified, and its online confirmation carries only Spearman 0.56, which itself compares near-identical F and L values. Please add a recurrence-only baseline to the leave-one-family-out comparison, and report how many of the 24 held-out assignments change when κ0 is dropped or perturbed within its measurement error.","section":"§4.1 (Eqs. 2–3), Table A.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The prospective half-split test is a genuine strength and is properly pre-registered, but it splits lessons and state facts on recurrence only (thresholds r = 0.377 and r = 0.394), so it supplies no evidence for κ0's role in routing. Taken together with the near-coincident κ0 of F and L, the two-property explanation currently rests on a single soft boundary whose discrimination is essentially one-dimensional. Either demonstrate κ0's routing value explicitly (for instance by a perturbed-κ0 sensitivity of the assignment) or soften the abstract's 'rule in two properties' formulation to 'a rule in recurrence, with state-conditionality as a secondary descriptor.'","section":"Table D.4","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The main text says routing 'exceeds the better single destination on every MobileGym backbone' and 'matches it with GUI-Owl-8B' on AndroidWorld; the AndroidWorld GUI-Owl cell is in fact −0.3 points below all-to-context with interval [−2.6, +2.0]. Please report the paired-bootstrap interval for the +3.5-point average gain over the better single destination so that the one non-positive cell is visible in the headline number.","section":"§5.4, Table 2, Table E.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The online state-conditionality κ(c) is defined as a ratio whose denominator is U(C_E^c) − U0; please state how the quantity is treated when the denominator is near zero or negative (e.g., when the exact-only route does not beat the base agent), since the rank correlation of 0.56 with κ0 depends on those cells.","section":"§3, Eq. (4)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The figure reports Spearman −0.92 for the after-to-before readout ratio against recurrence; please state explicitly over how many points this is computed (twelve cells of three backbones and four components, or item-level pools) and add the per-component r-coordinates to the figure caption.","section":"§4.3, Figure 6b","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The reference list is dominated by 2026 arXiv preprints, several of which are the authors' own works; please verify that the characterizations of the contrasted studies (Alakuijala et al., Li et al., Agrawal et al.) are accurate and complete, and consider trimming self-citations that do not bear on the comparison.","section":"Related Work and References","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The abstract promises 'Code and data will be released'; since several claims depend on the rule-based extractors and the exact retrieval templates, please provide a release link or supplement with the extractor code, pool statistics, and routing assignments so that the held-out protocol can be audited.","section":"Abstract / Reproducibility","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"The routing result is real and the reframing is worth having. Splitting experience into four components and asking where each belongs settles the fine-tune-versus-retrieve disagreement more cleanly than the whole-trajectory literature has managed, and the measurement is careful: paired instances, three seeds, analysis fixed before the runs, assignments written to disk before training, bootstrap intervals, interventions. The +3.5-point average edge over the better single destination is a genuinely trained-and-evaluated arm, not an extrapolation.\n\nThe soft spot is exactly what the stress-test note hits. The 24-of-24 held-out recovery is carried almost entirely by the lessons-versus-state-facts boundary. Locators and procedures sit two to four score units from the line and carry identical signs in every family, so the fit adds nothing for them. The one pair that decides is the pair whose signs are softest — three of six lesson intervals cross zero. The authors disclose S=21.5/24 rather than 24, which helps, but the headline number still overstates what the rule demonstrably discriminates.\n\nThe second dimension is the bigger problem. Recurrence alone orders all four components correctly; a threshold near r≈0.35 reproduces the identical assignment. kappa0 is confirmed only at Spearman 0.56, and its near-boundary contribution is never isolated. So 'two properties' is not established; the paper effectively reports one, recurrence — though the state-condition interventions do real work (removing the condition costs 3.4–3.9 points of context gain), so kappa0 is not dead. It is simply not shown to move the boundary.\n\nOther proportionate nits: code and data are promised, not shipped; on AndroidWorld the routed arm matches rather than beats the better single destination for GUI-Owl-8B; the budget inequality between routes is disclosed but makes the comparison less than fully symmetric. The self-citation cluster is heavy but the papers are mostly external and recent; not a real problem.\n\nThis deserves a serious referee. The measured advantage stands; what needs revision is the causal language around the two-property rule and an honest statement about how little the fit discriminates. A recurrence-only ablation would settle it cleanly, and asking for it would materially improve the paper.","headline":"Solid measurement, overstated law: the +3.5-point routing advantage is real, but the 24/24 held-out rule is effectively a one-dimensional recurrence threshold dressed up as a two-property account.","tokens_in":41908,"tokens_out":2104,"duration_ms":15339},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"GUI agents store their own experience in two places at once: a linear rule in two properties measured before any training, recurrence and state-conditionality, decides which of four component types belongs in the weights and which in the…","keywords":["GUI agents","self-improvement","experience reuse","component routing","in-context learning","fine-tuning","recurrence","state-conditionality"],"falsifier":"Take a new GUI environment or backbone family where procedures or state facts lie at high recurrence and low state-conditionality, the locator corner of the plane; if routing by the fitted rule sends them to the weights but the measured route gain on the same held-out items is negative, the boundary is misrouted.","tokens_in":42182,"feed_emoji":"📱","tokens_out":5415,"duration_ms":35715,"temperature":0.7,"pith_summary":"Self-improving GUI agents collect their own trajectories and return them to the agent either by fine-tuning into the weights or by retrieval into the context, and studies comparing these two destinations disagree. The paper attributes the disagreement to bundling: a trajectory mixes items with different properties, so a conclusion about the whole bundle depends on its mix. Component routing splits the experience into locators, procedures, state facts and lessons, and compares the two destinations per component on the same held-out items across three backbone families and two environments. The finding is that one pool has two destinations: locators and lessons win in the weights, procedures and state facts in the context. A linear rule in recurrence and state-conditionality, both measurable before any training, recovers the winning destination of a held-out backbone family in 24 of 24 cells, and routing by the rule beats every whole-trajectory baseline and, on average by +3.5 points, the better single destination.","feed_headline":"Routing beats dumping: two properties decide where experience goes","feed_subtitle":"On held-out tasks the rule wins by +3.5 points against the better single destination and beats every whole-trajectory baseline.","key_machinery":"The central object is component routing: an extractor carves a trajectory pool into four item sets—locators (where a named element sits), procedures (a recurring run of action targets), state facts (preconditions that must hold before an action) and lessons (failed-action / succeeded-action pairs)—each item carrying a key, a note for the context route, and a sample for the weights route. Two properties, both measured on the pool before any training, characterise each component: recurrence r, the share of unrelated tasks where the item's key reappears, and state-conditionality κ0, how much the item's usefulness depends on the current state matching its source state. The comparison is the route gain Δ(c)=U(Wc)−U(Cc), held-out success with the component in the weights minus the same items in the context. A linear rule sθ(c)=αr(c)−βκ0(c)−γ, fitted by weighted logistic loss with a ridge penalty while holding out one backbone family, predicts the sign of Δ(c), and the online quantity κ, read from exact-only versus relaxed-only retrieval, confirms the pre-training proxy. This rule is what carries the argument: it recovers the winning destination in 24 of 24 held-out cells and splits lessons and state facts at the boundary by item recurrence, with the ten-times learning rate leaving the route gains essentially unchanged.","core_discovery":"On the same items and the same held-out tasks, the route gain Δ(c)=U(Wc)−U(Cc) is positive for locators (+4.9 points pooled) and lessons (+1.5) and negative for procedures (−4.2) and state facts (−4.1), and this sign pattern repeats across three backbone families and two environments. The paper's central claim is that experience type, not pipeline choice, determines where experience should be stored, and that two properties measurable on the pool before any training—recurrence, how often a key reappears in unrelated tasks, and state-conditionality, how much usefulness depends on the state matching its source—predict the destination. A rule fitted with one backbone family held out recovers the sign in 24 of 24 cells, splits lessons and state facts at the boundary by item recurrence so that both halves land on the predicted side, and routing each component by the rule exceeds the better single destination of each backbone by +3.5 points on average. Two interventions support the mechanism: copying procedure items eightfold moves their route gain toward the weights by +3.5 points, and stripping the state condition from state facts costs 3.4 points. Training on a component makes its notes redundant, most for the items that recur most (after-training readout falls to 11 percent for the highest-recurrence locators and lessons), context gains increase with the information gap, and weights gains decrease with the policy gap.","pith_inferences":["If the experience-type principle holds, it generalises beyond GUI agents: self-improving systems of any kind could subdivide their trajectory data by functional role and measure per-item properties before committing to a storage medium, rather than treating whole trajectories as an atomic unit.","The near-degenerate state-fact / lesson pair, with the pre-training proxy κ0 confirming online state-conditionality at only Spearman 0.56, suggests that a finer-grained property or a non-linear boundary may be needed to separate components that sit close to the line.","The readout drop after internalisation points to a concrete token-saving strategy: items that recur most should be internalised once and then dropped from the context, without loss of accuracy.","A testable extension is a blind routing experiment on a fresh environment: measure recurrence and state-conditionality before training, predict each component's destination, and compare the predicted routing to the empirically optimal one on the same held-out items."],"forward_implications":["No single destination is right for all experience: locators and lessons win in the weights, procedures and state facts in the context, on the same items.","A rule in recurrence and state-conditionality, both measurable before any training, predicts the winning destination on a held-out backbone family and splits components internally by item recurrence.","Routing each component by the rule exceeds the better single destination by +3.5 points on average, while the reverse assignment loses 7.8 to 8.9 points against routing.","Fine-tuning a component makes its notes redundant, most for the items that recur most, so coexistence adds little where the weights already win.","Context gains grow with the information gap and weights gains shrink with the policy gap, so notes carry across producers and consumers while fine-tuning largely buys practice on the consumer's own trajectories."],"supporting_citations":[{"why":"Provides MobileGym, the simulated-phone environment with readable application state whose task templates and held-out split the experiments use.","marker":"Wu et al., 2026j"},{"why":"Provides AndroidWorld, the second environment with an accessibility tree but no exposed application state, used to test the route gains and the rule.","marker":"Rawles et al., 2025"},{"why":"Supplies LoRA, the mechanism the weights route uses to fine-tune each backbone on component samples.","marker":"Hu et al., 2022"},{"why":"Supplies direct preference optimization, the training method for the lessons component's preference pairs.","marker":"Rafailov et al., 2023"},{"why":"Provides OpenApps with its two application versions, used to test what happens when the weights hold one version and the notes come from another.","marker":"Ullrich et al., 2026"},{"why":"Supplies Qwen3-VL-8B, one of the three backbone families whose pools and route gains the rule is tested across.","marker":"Bai et al., 2025"},{"why":"Supplies ScaleCUA-7B, another backbone family used in the three-family comparison.","marker":"Liu et al., 2026a"}],"fun_headline_variants":["Two properties pick where GUI experience goes, beats baselines","Component routing: one rule sends each skill to weights or context","Rule routes locators to weights, procedures to context, +3.5 points","Experience type, not pipeline, decides storage: 24 of 24 cells","Recurrence and state-conditionality predict component destination"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The load-bearing premise is that the two pre-training properties, recurrence and state-conditionality, really measure why a component belongs in the weights or the context, so the boundary fitted on two backbone families can be trusted on a new family; the near-degenerate state-fact / lesson pair and the only moderate agreement between the pre-training proxy and the online measure mark where that premise is thinnest.","fun_headline_variants_meta":{"raw":{"variants":["Two properties pick where GUI experience goes, beats baselines","Component routing: one rule sends each skill to weights or context","Rule routes locators to weights, procedures to context, +3.5 points","Experience type, not pipeline, decides storage: 24 of 24 cells","Recurrence and state-conditionality predict component destination"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"cost":0.00019222,"is_byok":false,"cost_details":{"upstream_inference_cost":0.00019222,"upstream_inference_prompt_cost":0.00015554,"upstream_inference_completions_cost":3.668e-05},"total_tokens":1242,"prompt_tokens":1111,"completion_tokens":131,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":38}},"tokens_in":1111,"tokens_out":131,"cost_source":"openrouter_usage","duration_ms":2147,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":38,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-02T09:43:58.603217+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Take a new GUI environment or backbone family where procedures or state facts lie at high recurrence and low state-conditionality, the locator corner of the plane; if routing by the fitted rule sends them to the weights but the measured route gain on the same held-out items is negative, the boundary is misrouted.","supporting_citations":[{"cited_title":"AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents","cited_arxiv_id":null,"evidence_quote":"Provides AndroidWorld, the second environment with an accessibility tree but no exposed application state, used to test the route gains and the rule."},{"cited_title":"Manning, and Chelsea Finn","cited_arxiv_id":null,"evidence_quote":"Supplies direct preference optimization, the training method for the lessons component's preference pairs."},{"cited_title":"OpenApps: Simulating Environment Variations to Measure UI-Agent Reliability","cited_arxiv_id":null,"evidence_quote":"Provides OpenApps with its two application versions, used to test what happens when the weights hold one version and the notes come from another."},{"cited_title":"Qwen3-VL Technical Report","cited_arxiv_id":"2511.21631","evidence_quote":"Supplies Qwen3-VL-8B, one of the three backbone families whose pools and route gains the rule is tested across."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}