{"id":"ae7a4327-3443-4bb8-b60b-c12298be9ab1","arxiv_id":"2610.30478","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"In a pre-registered embodied-agent reach task, inference-time retrieval of past experience added +34 percentage points to success over loading the same experience into the policy weights via distillation, which added only +1 point.","lead":"This study compares two ways for an embodied robot to use its past experience: baking it into the policy's weights or keeping it outside and retrieving it at decision time.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"No significant objection identified.","rationale":"The reader's weakest_assumption identifies the same concern I have: the marginal/conditional mechanism is supported primarily by exploratory analyses. The paper is transparent about this limitation, scopes its claim accordingly, and does not present the mechanism as proven. The core empirical finding—retrieval on fixed weights adds +34 pp—is backed by confirmatory tests with high significance and consistent sensitivity checks. The L2 result is n.s. but the paper honestly reports it and limits the cross-task generality claim. I see no additional load-bearing concern that would change the verdict. The paper sets a high standard for pre-registration and transparent reporting.","tokens_in":31752,"tokens_out":3274,"duration_ms":33778,"concrete_test":"Re-run the L0a experiment with n=400 episodes per arm to increase power for the mechanism probe; if the R1∆X variance ratio (C_retrieval σ ≈14.4 vs distilled σ ≈9) replicates and the R1' slope remains positive (r>0.2), the conditional-correction interpretation gains independent support. If the variance difference attenuates, retrieval-noise may dominate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central interpretive claim (marginal knowledge transfers into weights; conditional knowledge belongs in context) rests on the exploratory R1∆X probe and variance analysis, which the paper transparently labels as exploratory and not pre-registered for the specific variance comparison. The R1' slope analysis (§4.2b) partially addresses the concern that retrieval noise could produce the variance; it shows with-buffer corrections track the required correction (r=+0.29 to +0.49), consistent with principled conditional structure rather than pure noise. The L2 σ test was null (F=1.25), but the paper acknowledges the cross-task replication is incomplete and does not overclaim. The reader's weakest_assumption correctly identifies this as the weakest link, but the paper's honesty about the evidence level and the robustness of the confirmatory retrieval effect (+34 pp, z=5.15, p≈2.6e-7) support the ACCEPT verdict.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper presents a pre-registered experimental study comparing two ways of supplying an embodied agent's experience to its policy: baked into weights via distillation, or kept external and retrieved at inference. Using a memory-augmented VLM teacher on a vision-conditioned single-arm reach task (L0a) and a harder dual-arm position-goal reach task (L2), the authors train LoRA student policies from identical teacher trajectories. The primary confirmatory test (T4) shows that attaching inference-time retrieval to fixed student weights significantly outperforms the baked-in student (+23 pp, p=7.8e-4); an exploratory identical-weights contrast gives +34 pp (p≈2.6e-7). The baked-in student shows no significant gain over an action-only control (T1, +1 pp, n.s.), leading the authors to withdraw the memory-in-weights claim while noting the test was underpowered. An exploratory behavioural probe (R1∆X) suggests that distillation transfers only the marginal (average) correction, while retrieval supplies the conditional (situation-specific) correction. A second experiment (L2) shows a direction-consistent but non-significant retrieval advantage (+6 pp), with exploratory diagnostics attributing the attenuation to buffer coverage and start-distance gating.","tokens_in":31991,"tokens_out":3732,"duration_ms":38411,"significance":"If the results hold, they provide a clean, controlled demonstration that for this distillation recipe (SFT+composable-KTO, LoRA rank-16) on these tasks, the conditional component of retrieved experience does not consolidate into weights, while the marginal component does. The marginal/conditional distinction offers a design principle for where to route different types of knowledge in embodied agents and connects to complementary-learning-systems theory. The paper's strength lies in its rigorous pre-registration, transparent reporting of limitations, and the fact that even the exploratory mechanism probe was pre-registered with per-arm predictions. The confirmatory retrieval effect is robust and survives the pre-registered alpha. The cross-task evidence is weaker but honestly scoped. The work is likely to influence how the community thinks about memory substrates in robotics and VLA post-training.","major_comments":[{"comment":"The sentence 'This is the criterion D11 was ultimately built to test, and it is the transferable claim we carry out of a single-recipe study across two tasks' could be read as claiming cross-task validation. However, the L2 contrast (C_retrieval vs A_ctrl_rat, +6 pp, n.s.) did not meet the pre-registered replication criterion (Amendment 2), and the mechanism evidence on L2 is post-hoc with a null σ test. The paper's own §5.9.1 and §5.9.3 carefully scope the L2 result to 'mechanism-and-direction reading, not effect-size replication.' To avoid overstatement, please revise the above sentence to explicitly note that the cross-task support is only directional and exploratory, consistent with the scoping already in §5.9.","section":"§6.1, second paragraph","authors_acknowledge":"The paper’s claim is scoped more narrowly, pre-registered as aconditional: Amendment 2 pre-specified thata significant, same-direction contrast would constitute cross-task replication of the identical-weights retrieval effect(not of the full 2×2).","authors_acknowledge_p":0.78},{"comment":"The variance contrast (σ=14.4 for C_retrieval vs ~9 for distilled arms) is the primary visual evidence for the conditional-correction claim. This analysis was not pre-registered (the probe itself was exploratory per Amendment 14 §14.5) and the paper transparently labels it as such. However, given that the mechanistic claim is central to the paper's transferable takeaway, the exploratory label should be kept prominent in the summary and discussion. The current discussion (§6.1) repeats the claim without restating that it rests on exploratory, non-pre-registered evidence. Please add an explicit caveat (e.g., 'This mechanism reading is supported only by exploratory analyses that were not pre-registered for the variance comparison; it requires confirmatory replication.') in §6.1 where the criterion is introduced.","section":"§4.3, R1∆X variance analysis","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"Line 8: 'the registered test proves underpowered for its own minimum effect' – awkward phrasing. Suggest: 'the registered test is underpowered for its own minimum effect threshold, so we withdraw rather than refute the memory-in-weights claim.'","section":"Abstract","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence 'The recap prompt received an unformatted {target_color} instruction template (disclosed in §4.5)' is ambiguous. Clarify whether the unformatted template appears only in the buffer entries or also in the prompts. The instrument-defect item in §4.5 helps, but consider adding one sentence in §3.2 explaining the effect.","section":"§3.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Table 4 uses the term 'D10-ext buffer' for C_retrieval's inference retrieval, but the Methods (§3.4) refer to 'frozen teacher buffer snapshot.' Harmonize the terminology for consistency.","section":"Table 4, arm definitions","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The correlation across two tasks (L0a vs L2) is reported as 'Spearman ρ = -0.58' (implied from the text). This is a sample of n=2 tasks, which cannot support a reliable correlation estimate. The paper already says 'n=2 tasks' and calls it exploratory, but the quantitative ρ may give a false impression of precision. Consider presenting only the medians and the effect direction, without claiming a correlation coefficient.","section":"§5.9.3, diagnostic (a)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The p-values for the round-3 correlations are permutation-based. State the number of permutations (2000) in the table caption or footnote, as is done in the text. Also note that the Bonferroni-corrected threshold for 12 tests (4 arms × 3 rounds) would be 0.0042; B_main's p=0.001 survives, but A_ctrl_rat's p=0.013 does not. The text does not discuss multiplicity for these post-hoc correlations.","section":"§4.4b, Table 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The asymmetry in recap generation (reasoning-on for L0a vs reasoning-off for L2) is disclosed, but the potential impact on recap quality is not discussed. Since recaps are the content of the conditional correction, a difference in reasoning mode could affect the nature of the retrieved context. Please add a brief note on whether and how this might influence the comparative interpretation.","section":"Appendix A (Teacher configuration)","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"The headline result here is real and worth taking seriously: on fixed student weights, adding retrieval boosted success by +34 pp (p≈3×10⁻⁷). The core empirical claim—that for this recipe and task, memory lives in context, not in weights—is well-supported.\n\nWhat this paper does really well is its methodological discipline. Full pre-registration with amendment chain, honest power analysis on the null, clear separation of confirmatory versus exploratory, and careful scope limitations. The paper sets a standard for how to report a mixed-result study without overclaiming.\n\nThe conceptual contribution—the marginal/conditional distinction—is interesting as a hypothesis, but the evidence for it is exploratory: the variance analysis and R1∆X probe were not pre-registered for the mechanism claim, and the cross-task replication (L2) was direction-consistent but non-significant. The authors are upfront about this, but readers should take the mechanism as a plausible post-hoc interpretation, not a confirmed law.\n\nSoft spots: (1) The mechanism claim rests on exploratory analyses with moderate sample sizes. (2) The L2 null limits generality and was predicted by the authors' own power analysis. (3) The distillation null is underpowered, so the claim that weights 'can't hold conditional knowledge' is less certain than the retrieval positive result. The paper acknowledges all of these.\n\nWho is it for? Anyone building memory-augmented embodied agents, and anyone interested in rigorous pre-registered experimentation in ML. It deserves peer review: the empirical core is solid, the interpretation is honestly scoped, and even if the mechanism doesn't generalize, the finding that for this setting retrieval trumps distillation is a contribution. I'd recommend acceptance with minor revision, mainly to keep the exploratory status of the mechanism explicit in the final version.","headline":"Methodologically rigorous empirical study showing that for this embodied task, retrieval vastly outperforms weight consolidation; the proposed marginal/conditional mechanism is a plausible but unconfirmed interpretation.","tokens_in":32483,"tokens_out":1526,"duration_ms":16016},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"The paper establishes a design criterion for where an embodied agent's experience should live: marginal knowledge (the correction that holds across situations) transfers into weights, while conditional knowledge (the situation-specific…","keywords":["memory externalisation","conditional knowledge","pre-registered experiment","embodied agent","distillation","retrieval-augmented generation","LoRA","vision-language-action model"],"falsifier":"Run the same experimental design (L0a reach task, LoRA student, frozen retrieval buffer) but replace the flat distillation with a surprise-gated consolidation method (e.g., Titans-style writing that writes only unexpected experiences) and check whether the conditional trace reappears in the student weights at round 1. If the conditional σ approaches 14.4 (retrieval level) rather than 8–10 (static prior), the marginal/conditional split would be refuted as a necessary architectural constraint.","tokens_in":33189,"feed_emoji":"💡","tokens_out":4122,"duration_ms":43503,"temperature":0.7,"pith_summary":"The paper asks a controlled pre-registered question about an embodied reaching agent: when a memory-augmented teacher produces trajectories, should the student bake that experience into its LoRA weights during distillation, or should it keep the teacher's retrieval buffer external and consult it at inference? The answer is both, but the distinction is which part of the knowledge. Distillation moved the average correction (a static prior) into weights, but the student gained only +1 percentage point over an action-only baseline. By contrast, attaching the same frozen retrieval buffer to a fixed set of student weights added +34 points, matching the teacher's own success rate. An exploratory behavioural probe suggests the mechanism: the student's weights captured the constant marginal shift but not the per-episode conditional adjustment that retrieval supplies. The paper therefore recommends externalising memory for situational corrections and relying on weights only for the average behaviour.","feed_headline":"Marginal in weights, conditional in context","feed_subtitle":"A pre-registered embodied study finds that distillation transfers only the average correction; retrieval supplies the situation-specific…","key_machinery":"The central mechanism is the marginal/conditional distinction, operationalised through the R1∆X behavioural probe: the round-1 lateral displacement toward the target. The four distilled arms all cluster at a static prior of about -16.7 grid units (close to the memory-teacher's marginal fingerprint), while retrieval produces a per-episode variable correction with mean near zero and large variance (σ=14.4). This split provides the evidence that distillation transfers the average but retrieval supplies the situation-specific adjustment.","core_discovery":"The core discovery is a split in what kind of knowledge each substrate holds when an embodied agent learns from a memory-augmented teacher. Under a pre-registered design that isolates the variable, the paper shows that distilling the teacher's trajectories into LoRA weights transfers the marginal correction—the average distribution shift that applies across all episodes—but fails to transfer the conditional, situation-specific correction that varies per episode. When the same retrieval buffer is attached at inference time to identical student weights, the student recovers the full teacher-level success by supplying that conditional correction per-episode. The result is summarised as: 'Marginal knowledge (the average correction that holds across situations) transfers into weights; conditional knowledge (the correction that depends on the situation) belongs in context, because weights approximate it with a constant and context conditions on it for free.'","pith_inferences":["If the marginal/conditional split generalises beyond reaching tasks, then lifelong learning agents should architecturally separate a slow weights channel (for average corrections) from a fast retrieval channel (for situational adjustments), with the retrieval channel indexed by state similarity to the current situation.","The exploratory round-3 correlations (B_main r=+0.40) hint that a conditional trace may exist in later rounds; a round-resolved analysis with more statistical power or a round-by-round distillation target could determine whether the conditional structure is genuinely absent from weights or simply under-learned at round 1.","The buffer-provenance caveat (teacher's buffer vs student's own buffer) suggests that a student retrieving its own self-generated lessons might produce different conditional corrections; this is an open question that could be tested by running a separate arm where the student builds its own buffer through self-exploration.","The paper's pre-registration stance—with a full amendment chain and a locked analysis rule before p-values—offers a methodological template for reporting null results in embodied learning, which may encourage more honest disclosure of non-significant findings in the field."],"forward_implications":["System builders should externalise the memory: keep a retrieval buffer of episodic experience for situation-specific corrections, while using weights only for the average behavioural prior; the adapter adds nothing detectable once the buffer is attached.","The recipe for parametric memory consolidation should target the conditional component, not just the marginal; flat distillation (SFT + KTO) is insufficient for transferring situation-specific corrections.","In the L2 harder task the retrieval effect was directionally consistent but non-significant (+6 pp), and the attenuation is linked to thinner memory coverage and a distance-gated conversion threshold, suggesting the effect scales with the amount of retrievable success experience.","The pre-registered null on parametric memory is scoped to this recipe, data scale, and adapter capacity; stronger methods (surprise-gated writing, multi-round training) remain untested and may capture the conditional structure."],"supporting_citations":[{"why":"Provides the LoRA adapter architecture used for all student policies; the capacity limitation of rank-16 LoRA is directly discussed as a sibling explanation for the null on parametric memory.","marker":"[Hu et al., 2022]"},{"why":"Supplies the KTO alignment method used in the two-stage student training (SFT + composable-KTO); the training recipe is fixed across all arms.","marker":"[Ethayarajh et al., 2024]"},{"why":"Defines the VLM teacher backbone (Gemma-4-31B-it) and the base model for all students; the memory-augmented teacher's behaviour and the retrieval context depend on this model.","marker":"[Gemma Team, 2024]"},{"why":"DINOv2 provides the visual embedding used for similarity-based retrieval of past episodes; the retrieval quality metric relies on this embedding.","marker":"[Oquab et al., 2024]"},{"why":"Establishes the retrieval-augmented generation (RAG) paradigm that motivates the externalised-memory contrast in the paper.","marker":"[Lewis et al., 2020]"},{"why":"Voyager is an external-memory agent that the paper cites as supporting evidence for the retrieval-augmented pole, which its results align with.","marker":"[Wang et al., 2024]"},{"why":"Their state-conditionality rule for GUI agents parallels the paper's marginal/conditional split; the paper discusses why their 'lessons in weights' result differs from its own.","marker":"[Wu et al., 2026a]"}],"fun_headline_variants":["Memory belongs in context, not in weights","Distillation transfers average; retrieval supplies specifics","Externalise the memory: weights get marginal, context conditional","Keep memory out: weights store averages, context stores specifics"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The claim that distillation transfers only the marginal correction while retrieval supplies the conditional correction rests heavily on an exploratory behavioural probe (R1∆X) whose units were corrected via an erratum after the fact, and whose variance analysis (the σ contrast) was not pre-registered; without this probe the mechanism reading would lack direct evidence.","fun_headline_variants_meta":{"raw":{"variants":["Memory belongs in context, not in weights","Distillation transfers average; retrieval supplies specifics","Externalise the memory: weights get marginal, context conditional","Keep memory out: weights store averages, context stores specifics"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000132,"raw_usage":{"cost":0.00013189722,"is_byok":false,"cost_details":{"upstream_inference_cost":0.00013189722,"upstream_inference_prompt_cost":5.442822e-05,"upstream_inference_completions_cost":7.7469e-05},"total_tokens":1864,"prompt_tokens":1089,"completion_tokens":775,"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":713}},"tokens_in":1089,"tokens_out":775,"cost_source":"openrouter_usage","duration_ms":9552,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":713,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-23T15:07:21.168204+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":"Run the same experimental design (L0a reach task, LoRA student, frozen retrieval buffer) but replace the flat distillation with a surprise-gated consolidation method (e.g., Titans-style writing that writes only unexpected experiences) and check whether the conditional trace reappears in the student weights at round 1. If the conditional σ approaches 14.4 (retrieval level) rather than 8–10 (static prior), the marginal/conditional split would be refuted as a necessary architectural constraint.","supporting_citations":[{"cited_title":"Gemma: Open Models Based on Gemini Research and Technology","cited_arxiv_id":"2403.08295","evidence_quote":"Defines the VLM teacher backbone (Gemma-4-31B-it) and the base model for all students; the memory-augmented teacher's behaviour and the retrieval context depend on this model."},{"cited_title":"DINOv2: Learning Robust Visual Features without Supervision","cited_arxiv_id":"2304.07193","evidence_quote":"DINOv2 provides the visual embedding used for similarity-based retrieval of past episodes; the retrieval quality metric relies on this embedding."},{"cited_title":"Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks","cited_arxiv_id":"2005.11401","evidence_quote":"Establishes the retrieval-augmented generation (RAG) paradigm that motivates the externalised-memory contrast in the paper."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}