{"id":"93ee53e6-9c3d-4f01-b94c-ea0bc5be233d","arxiv_id":"2608.06706","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Subtracting the action-marginal mean from a world model's action-conditioned predictions yields a vector action channel that exactly cancels common-mode distraction and works post hoc on frozen predictors.","lead":"A simple readout subtraction can expose the effect of an agent's own actions in latent world models, even when irrelevant scene motion dominates the prediction. The trick works on frozen, off-the-shelf world models and needs no reward, no decoder, and no distractor-specific loss.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The rejection guarantee is conditional on action-independent distractors being encoded action-independently in the offset head; nothing enforces that form, and the paper's leak probes, while supportive, are empirical rather than algebraic.","rationale":"The paper's strongest claim is that subtracting the action mean yields a clean, controllable action channel by exact common-mode cancellation. The algebraic identity in Proposition 1 is correct, but the practical guarantee depends on where the distractor lives in the learned offset. The reader identified this as the weakest assumption: an action-modulated interaction survives centering even when the environment's distractor is action-independent. My reading agrees, and I sharpen the point: the decomposition Delta = g + h is always possible by taking h as the action mean, so the real premise is that action-independent distractors are not encoded with action-dependent coefficients. The paper's own Appendix I states this limitation explicitly, and Section 5.1 concedes the boundary. The empirical evidence is strong — leak probes across synthetic, gridworld, DMC, and Atari settings show R2_w indistinguishable from zero — but it cannot turn an empirical regularity into an architectural guarantee. The proposed test directly inspects the offset head's dependency on the known distractor across actions, which would settle whether the common-mode premise holds at the level of the learned representation rather than only at the level of the final channel output. This concern does not invalidate the paper: the method works where measured, the negative results are disclosed, and the control claim is properly scoped to the gridworld. The conditional verdict remains appropriate, so I recommend no change.","tokens_in":24960,"tokens_out":12358,"duration_ms":116449,"concrete_test":"Take a trained centered checkpoint from the synthetic identifiability experiments with known controllable factor c and distractor w. For each action a, fit a ridge regression of Delta(z,a) on the true distractor state w and collect the coefficient vector beta_a. Test whether beta_a is identical across actions (e.g., with a permutation test or by checking overlap of confidence intervals). If beta_a varies across actions, the offset head encodes the action-independent distractor with an action interaction, and the common-mode premise fails in exactly the regime the paper's leak probes are designed to catch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 1 states that if Delta(z,a) = g(z,a) + h(z), then centering cancels h. Taken literally, the premise is vacuous: for any Delta one can take h(z) = mean_a Delta(z,a). The real assumption is a locality claim about the learned representation: the distractor's contribution to the learned offset must lie entirely in the action-independent component h. Because the offset is a neural network, an action-independent distractor can in principle be encoded as a zero-mean interaction s_dist(z) phi(a) (as Appendix I concedes), and such a term survives centering exactly. Nothing in the architecture forbids this; the only pressure toward the common-mode form is the offset-norm penalty L_off (lambda_off = 1e-4) and the capacity of the passive head B. The experiments show R2_w ≈ 0 across the tested settings, which is strong evidence for the premise, but it is not a guarantee. The central claim 'centering suffices for distractor rejection' therefore rests on an empirical regularity about trained offsets, not on the algebraic identity alone. This is the load-bearing assumption: if a future environment or model family produces action-modulated encodings of action-independent distractors, the advertised rejection fails silently, exactly where the paper's boundary (action-correlated distractors) does not cover it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a minimal restructuring of action-conditioned latent world models: replace the monolithic predictor P(z,a) with \\hat z' = B(z) + [\\Delta(z,a) - mean_{a'} \\Delta(z,a')], a dueling-style decomposition in which the action offset is centered by subtracting an action-marginal mean. The paper proves (Propositions 1 and 2) that any action-independent additive component h(z) of the offset cancels exactly, including when the action mean is estimated by a finite Monte-Carlo sample. It validates the mechanism on FourRooms with rolling distractors, synthetic generators with known controllable and nuisance factors, distracting DMC with composite occluders, and Atari Freeway, showing that the centered offset preserves decodability of the agent's own effect while nuisance leak is near zero. It also shows that post-hoc centering of frozen RePo and TIA models recovers an action channel, and reports a gridworld MPC success result. The stated boundary is action-correlated distractors; an inert learned gate is reported as a negative result, and a pre-registered DMC-scale control null is reported in full.","tokens_in":25205,"tokens_out":10862,"duration_ms":97827,"significance":"The algebraic core is simple and correct, and the empirical protocol is unusually transparent: three seeds with per-seed tables, pre-registered criteria with verbatim outcomes, explicit negative results, and a clearly stated failure boundary. The strongest asset is the triangular validation: an exact identity, known-factor generators that separate recovery from rejection, and a demonstration that the readout transfers to frozen third-party models. If the empirical premise concerning how trained offsets encode distractors holds, the method is a useful plug-in that requires no retraining, reward, or reconstruction. The paper's own limitations section and Appendix I are commendably honest, although the abstract and contribution bullets overstate what the identity alone guarantees. The central claim is defensible but needs a scope correction.","major_comments":[{"comment":"The central guarantee is conditional on a representational premise that the architecture does not enforce, and the abstract's wording is stronger than the theorem. For any offset head \\Delta, one can choose h(z) = mean_{a'} \\Delta(z,a') and g(z,a) = \\Delta(z,a) - h(z), so the decomposition \\Delta = g + h is always available; the proposition alone therefore does not certify any distractor rejection. Appendix I concedes that a zero-mean interaction s_dist(z) phi(a) survives centering exactly, and Sec. 5.1 correctly states that the identity removes only what the learned offsets represent identically across actions. The abstract and the first contribution bullet say that centering cancels \"the action-independent variation where distractors live\" and list only action-correlated distractors as the boundary. This should be corrected: the main text should state prominently that nuisance rejection is an empirical property of trained offsets, not an implication of the algebraic identity, and that action-independent distractors encoded as zero-mean interactions form a distinct failure mode from action-correlated distractors.","section":"Sec. 3.4, Eq. (1), Proposition 1; Sec. 5.1; Appendix I"},{"comment":"The rate claim in Proposition 2(ii) is stated for candidate actions drawn from a proposal \\nu and says the sampling error is O(1/sqrt(K)) under bounded second moments, but the proof requires the draws to be i.i.d. The deployed estimator in Appendix A pins the executed action as a_1 and mixes replay-buffer actions with Gaussian perturbations, so it is not an i.i.d. sample from a single proposal. The h-cancellation in part (i) is unaffected, but the theorem as stated does not cover the implemented estimator. The main text should either state the i.i.d. assumption explicitly in the proposition and describe the deployed structured proposal as a separate empirical check, or weaken the rate claim to match the measured decay. The empirical K-sweep with its fitted log-log slope of about -0.26 is a reasonable substitute, but it should not be presented as a direct instance of the proposition's rate.","section":"Sec. 3.4, Proposition 2(ii); Appendix B; Appendix A"}],"minor_comments":[{"comment":"The cross-references \"Secs. 4.2, 4.2\" appear twice in the limitations paragraph; these should be corrected to the intended sections (likely Sec. 4.2 and Sec. 4.3).","section":"Sec. 5.1"},{"comment":"The figure legends use the code name \"dueling\" for the \"centered\" variant while the text and tables use \"centered\"; the terminology should be unified.","section":"Fig. 2; Sec. 4.2"},{"comment":"The main text says the remaining K-1 candidate actions are \"split evenly\" between resampled and perturbed actions, while Appendix A specifies 7 and 8; the main text should say \"as evenly as possible\" to match the implementation.","section":"Sec. 3.2; Appendix A"},{"comment":"Table 11 reports standard deviations while other tables report s.e.m.; the caption or the protocol should clarify the convention.","section":"Appendix J, Table 11"},{"comment":"The phrase \"cancels whatever the actions share--the action-independent variation where distractors live\" should be reworded to match the theorem, for example \"cancels whatever enters the offset additively and identically across actions.\"","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest and unusually well documented; the pre-registered nulls and the explicitly reported negative results increase my confidence in the empirical core. The main risk is that the public abstract promises more than the algebra delivers, especially in light of the zero-mean interaction failure mode conceded in Appendix I. I do not see circularity or fabrication, but the abstract and contribution bullets need a scope correction before publication. The paper fits the journal's scope and the requested changes are feasible within a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a genuinely careful and honest contribution about a simple trick: for action-conditioned latent world models, subtract the mean prediction over actions at readout time. The algebra is elementary and correct; the novelty is modest (Dueling DQN and CAI arithmetic transplanted to vector latent dynamics), but the vector-valued common-mode-invariant action channel, and especially its post-hoc application to frozen third-party models, is new. What makes the paper worth your time is the reporting. Three seeds, per-seed tables, a pre-registered DMC gate that fails and is reported verbatim, a negative result on a learned gate, an explicit action-correlated boundary where everything collapses, and Appendix I stating plainly that the cancellation only removes what the learned offsets represent identically across actions. The frozen RePo/TIA result—raw action-delta probes near zero, centered readout of the same model at 0.19–0.52—is a nice demonstration that the channel is latent but unrouted.\n\nThe soft spot, in proportion: the rejection guarantee is conditional on an empirical premise. Proposition 1 cancels h if Δ(z,a)=g(z,a)+h(z), but nothing in the architecture forces an action-independent distractor to be encoded action-independently; a zero-mean interaction s_dist(z)φ(a) survives centering exactly, as the paper concedes. The leak probes (R2_w≈0 across synthetic, gridworld, Atari) are strong evidence that the premise holds in practice, but they are empirical, not a guarantee. So \"centering suffices for distractor rejection\" is only as strong as that premise. The paper scopes this correctly—it says \"common-mode rejection by construction, distractor rejection as measured\"—and the stress-test concern lands exactly where the paper already draws its boundary. I don't think that's a load-bearing flaw; it's a clearly labeled limit. For a reader who wants to deploy this, the check is: verify the common-mode premise in your own environment.\n\nThe other caveat is the control payoff. The preserved channel converts to goal-reaching in the gridworld, but the DMC-scale planner null is disclosed, and the trained-in centered model still shows action-separation collapse (channel preserved but unused). So the claim to defend is representation-level, not \"this fixes planning at scale.\"\n\nBottom line: this is a solid, honest paper. It deserves a serious referee; I'd send it out. The subfield (JEPA world models, model-based RL under distraction) will get a default diagnostic out of it, even if the guarantee is narrower than the title's phrasing suggests. I'd cite it and bring it to reading group.","headline":"A careful, honestly-scoped paper: action-mean centering is a real but modest contribution, and the cancellation guarantee is conditional on a premise the paper itself labels clearly.","tokens_in":25777,"tokens_out":2353,"would_cite":true,"duration_ms":22455,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Action-mean centering cancels action-independent distractors exactly in latent world models, restoring a clean action channel with no reward, reconstruction, or auxiliary losses.","keywords":["world models","distractor rejection","action-mean centering","dueling decomposition","latent dynamics","JEPA","common-mode cancellation","reinforcement learning"],"falsifier":"Train AD-JEPA on a controllable-nuisance generator where the nuisance's latent representation is gated by the executed action—e.g., the distractor feature is $+s$ for action 1 and $-s$ for action 2, while its dynamics remain action-independent. If the centered channel's leak to the nuisance state is measurably nonzero, the paper's empirical claim that action-independent distractors enter as common-mode would fail; the paper's own Appendix I shows the algebra permits exactly this failure.","tokens_in":24729,"feed_emoji":"🤖","tokens_out":15332,"duration_ms":103603,"temperature":0.7,"pith_summary":"Latent world models predict future states from an action, but when the frame contains motion the agent does not control, the predictions for different actions become indistinguishable—the model goes action-blind even as its training loss keeps improving. The paper claims that a minimal algebraic restructuring fixes this: decompose the predictor into a passive scene term and an action-offset term, and subtract from that offset its mean over actions. Whatever enters the offset identically across actions—the common-mode variation where action-independent distractors live—cancels exactly, with no reward, no reconstruction, and no distractor-specific loss. This is the dueling-decomposition identity $Q = V + (A - \\bar{A})$ transplanted from scalar value estimates to vector-valued latent dynamics, where centering becomes a distractor-rejection device rather than just an identifiability trick. Across a gridworld, synthetic generators with known factors, distracting continuous control, and natural-pixel Atari, the centered channel keeps the agent's own effect decodable with nuisance leak indistinguishable from zero, and the same subtraction applied post hoc to frozen pretrained world models surfaces an action channel their raw readouts miss.","feed_headline":"Subtracting the action mean cancels distractors in world models","feed_subtitle":"A dueling-style identity revives the action channel in world models, and the trick works on frozen pretrained ones.","key_machinery":"The central object is the action-decomposed predictor $\\hat{z}_{t+1} = B(z_t) + [\\Delta(z_t,a_t) - \\bar{\\Delta}(z_t)]$, with $\\bar{\\Delta}(z) = \\mathbb{E}_{a'\\sim\\nu}[\\Delta(z,a')]$, which the paper calls action-mean centering (the model is AD-JEPA). It is the dueling-decomposition identity $Q = V + (A - \\bar{A})$ of (Wang et al. 2016) transplanted from scalar values to vector latent dynamics, where the subtraction does double duty: it makes the channel well-defined despite the arbitrary-state-offset degeneracy, and it performs exact common-mode rejection. The load-bearing results are Proposition 1 (exact cancellation of any action-independent $h$ for discrete, enumerated action sets) and Proposition 2 (the same exact cancellation for any finite Monte-Carlo estimate of the action mean, with $O(1/\\sqrt{K})$ sampling error confined to an action-independent shift). Because the mechanism lives in the readout rather than the architecture, the same subtraction applies post hoc to frozen third-party world models, which is how the paper surfaces action channels in RePo and TIA without retraining.","core_discovery":"The central claim is that the effective action channel of any action-conditioned latent predictor is the action-offset minus its action-marginal mean: $u(z,a) = \\Delta(z,a) - \\bar{\\Delta}(z)$, and that this channel is exactly invariant to any component of the offset that is shared across actions. Concretely, if the offset head decomposes as $\\Delta(z,a) = g(z,a) + h(z)$ for an action-independent $h$, then $h$ cancels from $u$ for every value of the parameters (Proposition 1), and the same cancellation holds exactly when the mean is replaced by a Monte-Carlo average over any finite set of sampled actions (Proposition 2), with any sampling error confined to an action-independent shift. The model therefore is not blind to distractors—the passive head $B$ must track them to match the target—but the action channel has nowhere to hold them: rejection is routing, not blindness. Empirically the centered channel recovers the agent's own effect with nuisance leak statistically indistinguishable from zero across gridworld, synthetic, DMC, and Atari settings, and applied post hoc to frozen RePo and TIA hosts it surfaces action channels with probe $R^2$ between 0.19 and 0.52 where raw action-delta readouts read near zero. The paper is equally explicit about the boundary: a distractor whose motion is modulated by the agent's action is not common-mode and is not cancelled; the identity guarantees common-mode rejection, while distractor rejection is a measured empirical property, with the action-correlated failure reported alongside every positive result.","pith_inferences":["If the readout mechanism is as universal as the frozen-host results suggest, the same subtraction should act as a diagnostic for any action-conditioned predictive model beyond world models—for example video prediction or latent-action models—revealing whether an action channel exists at all; the paper does not test this.","The plug-in success on frozen RePo and TIA suggests that many world models may already contain the action information but route it through action-independent-looking offsets; we infer that centering could be applied to other entangled readouts, such as successor-feature heads or reward-prediction heads, to separate controllable from uncontrollable components.","Because the boundary is action-correlated distractors, any entity that reacts to the agent—pursuit, evasion, coordination—will defeat the method as stated; we infer that a multi-agent extension centering over joint actions or per-agent marginals is the natural next step, requiring observation or inference of other agents' actions that the current setting does not assume.","Centering is a readout change, so it should compose with any training-time distractor-suppressing objective; we infer that identity-level cancellation and incentive-based separation attack different failure modes and may be complementary rather than competing."],"forward_implications":["Any action-conditioned world model, including a frozen pretrained one, can expose an action channel by subtracting the mean prediction over sampled actions at readout time; the paper measures this on RePo and TIA, where raw action-delta probes read about zero but the centered readout reaches $R^2$ 0.19–0.52.","Under action-independent distraction the centered channel recovers the controllable factor and rejects the nuisance simultaneously: on the synthetic identifiability grid, recovery is $R^2_c = 0.87$–$0.93$ while leak is about $-$0.00, where the noncentered ablation recovers similarly but leaks at 0.74–0.97.","Because Proposition 2 makes the cancellation exact for any finite $K$, the method does not require enumerating actions; a $K=16$ Monte-Carlo mean preserves the channel in continuous control, with sampling error appearing only as an action-independent shift.","The algebra makes the failure mode legible: action-correlated (reafferent) distractors are not common-mode and are not cancelled, and the paper measures the collapse on both the probe and the control side (offset probes fall to chance, gridworld control degrades to 0.55–0.57) at exactly the configured stress point.","The preserved channel can convert into goal-reaching control in the gridworld (0.92–0.93 vs 0.57 for the standard predictor under distraction), though the paper reports a pre-registered null at DMC scale and limits the control claim to sufficiency, not superiority."],"supporting_citations":[{"why":"Supplies the dueling identity Q = V + (A - mean A) that the paper transplants from scalar values to vector latent dynamics; the centering subtraction originates here.","marker":"(Wang et al. 2016)"},{"why":"Introduces the sampled-action mean arithmetic (CAI) that the paper promotes from a post-hoc scalar influence score to a trained vector-valued channel.","marker":"(Seitzer, Schölkopf, and Martius 2021)"},{"why":"Provides the action-conditioned monolithic JEPA predictor (DINO-WM) that AD-JEPA restructures.","marker":"(Zhou et al. 2025)"},{"why":"Provides another action-conditioned JEPA host (V-JEPA 2-AC) of the same monolithic-predictor family.","marker":"(Assran et al. 2025)"},{"why":"TIA is one of the frozen host models whose raw action-delta readout reads near zero while post-hoc centering recovers a channel; the plug-in experiment needs these hosts.","marker":"(Fu et al. 2021)"},{"why":"RePo is the other frozen host model that demonstrates the inference-only plug-in property of centering.","marker":"(Zhu et al. 2023)"},{"why":"The exogenous-filtering line the paper contrasts with: centering claims only per-transition common-mode structure, not process-level exogeneity.","marker":"(Efroni et al. 2022)"},{"why":"The reafferent-distractor stress case that marks the measured boundary where common-mode cancellation fails.","marker":"(Hutson, Kauvar, and Haber 2024)"},{"why":"AD3's agent-like distractors delimit the same action-correlated boundary for the guarantee.","marker":"(Wang et al. 2024)"},{"why":"Concurrent work pursuing the same distractor-discarding outcome via an inverse-dynamics incentive; the paper positions its structural identity against this.","marker":"(Ivashkov, Balestriero, and Schölkopf 2026)"}],"fun_headline_variants":["Dueling subtraction gives world models a clean action channel","No reward, no reconstruction: just subtract the action mean","Common-mode distractors vanish with a single readout subtraction","Frozen world models gain action sight via dueling trick"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on action-independent distractors entering the predictor's action offsets the same way for every action; if a distractor's contribution varies with the chosen action, the subtraction leaves it in the channel.","fun_headline_variants_meta":{"raw":{"variants":["Dueling subtraction gives world models a clean action channel","No reward, no reconstruction: just subtract the action mean","Common-mode distractors vanish with a single readout subtraction","Frozen world models gain action sight via dueling trick"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1569,"prompt_tokens":1147,"completion_tokens":422,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":763,"completion_tokens_details":{"reasoning_tokens":355}},"tokens_in":763,"tokens_out":422,"duration_ms":4325,"temperature":1.0,"reasoning_tokens":355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:30:31.765427+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train AD-JEPA on a controllable-nuisance generator where the nuisance's latent representation is gated by the executed action—e.g., the distractor feature is $+s$ for action 1 and $-s$ for action 2, while its dynamics remain action-independent. If the centered channel's leak to the nuisance state is measurably nonzero, the paper's empirical claim that action-independent distractors enter as common-mode would fail; the paper's own Appendix I shows the algebra permits exactly this failure.","supporting_citations":[{"cited_title":"Causal Influence Detection for Improving Efficiency in Reinforcement Learning , booktitle =","cited_arxiv_id":null,"evidence_quote":"Introduces the sampled-action mean arithmetic (CAI) that the paper promotes from a post-hoc scalar influence score to a trained vector-valued channel."},{"cited_title":"Sensorimotor World Models: Perception for Action via Inverse Dynamics , journal =","cited_arxiv_id":null,"evidence_quote":"Concurrent work pursuing the same distractor-discarding outcome via an inverse-dynamics incentive; the paper positions its structural identity against this."}],"review_version":2}