{"id":"bacee20f-efb9-4a50-8ca0-e60893469815","arxiv_id":"2608.04396","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"By subtracting a vision-only policy's action from the full policy, CofactVLA reduces vision-override and improves VLA robustness and out-of-distribution generalization in simulation and real robots.","lead":"CofactVLA adds a counterfactual branch to vision-language-action models: it runs the policy once without the language instruction and pushes the main action away from that vision-only habit. The method reports large success-rate gains in simulation and real-world tests, including a 52.3% absolute improvement in out-of-distribution robot scenes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) adds the biased factual velocity back, so OPG cannot 'cleanly eradicate' visual bias; the deconfounding claim is contradicted by the paper's own algebra.","rationale":"The reader's CONDITIONAL verdict remains appropriate, but the condition should be sharpened. The reader's weakest assumption concerns whether the language-masked counterfactual branch captures exactly the visual confounder and whether the orthogonality/spectral assumptions hold. That is a valid concern. However, a more direct and load-bearing problem is internal to the algebra: even granting the counterfactual branch captures v_bias exactly, Eq. (5) does not remove v_bias from the final velocity. It adds the orthogonal semantic component to the original biased velocity, leaving the bias term with coefficient 1. The paper's own terminology calls v_perp the 'pure, deconfounded' velocity, yet the final formula reintroduces the biased v_cond. This is an internal inconsistency, not merely an unvalidated assumption. If the implemented method actually uses a different formula (e.g., subtraction), the paper needs to state it and release code; if it uses Eq. (5), the causal deconfounding interpretation is unsupported. The empirical success-rate gains could still arise from CFG-like semantic amplification rather than from deconfounding. Since this is a fixable issue but currently invalidates the central theoretical claim, the CONDITIONAL verdict stands, with the condition being a corrected derivation or code confirming the implemented intervention.","tokens_in":25917,"tokens_out":8275,"duration_ms":89542,"concrete_test":"Independently re-derive Eqs. (3)-(5) under the paper's own assumption that the counterfactual branch equals the visual confounder. With v_cond = v_sem + v_bias and v_uncond = v_bias, verify that Eq. (5) yields v_causal = (1+gamma) v_sem + v_bias, so the bias component is not canceled. Ask the authors either to correct Eq. (5) to v_causal = v_perp (or v_cond - gamma * v_proj) and re-run the LIBERO and real-robot ablations, or to provide a revised derivation showing how the bias term is eliminated. If the implemented formula differs from the paper's Eq. (5), code release would resolve the discrepancy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2's OPG does not implement the claimed deconfounding. Under the paper's own assumption that the counterfactual branch v_uncond captures exactly the visual confounder v_bias, decompose v_cond = v_sem + v_bias with v_sem orthogonal to v_bias. Eq. (3) gives v_proj = v_bias, and Eq. (4) gives v_perp = v_sem, which the text calls 'the pure, deconfounded language semantic intent.' But Eq. (5) sets v_causal = v_cond + gamma * v_perp = (1+gamma) v_sem + v_bias. The spurious bias term v_bias remains with coefficient 1; it is not eradicated, only relatively downweighted. Thus the central claim that OPG 'cleanly eradicates visual bias' and 'exclusively re-weights the causal mode odds along the semantic direction' is false as written. To actually remove the bias, the formula would need v_causal = v_perp or v_cond - gamma * v_proj. Relatedly, the densities in Section 3.2 are mislabeled: p_cond is written as p(A|O) and p_uncond as p(A|T), which is the reverse of the factual/counterfactual branches defined in Section 3.1. This suggests a sign/definition error in the core intervention. The empirical gains may still be real, but they would not support the causal deconfounding mechanism claimed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CofactVLA, a two-branch causal intervention framework for Vision-Language-Action (VLA) models that aims to remove 'vision-override' bias caused by spurious visual confounders. A factual branch conditioned on image and language and a counterfactual branch conditioned only on the image are generated in a single forward pass; the paper then applies two interventions: Action-Level Orthogonal Projection Guidance (OPG) to the flow-matching velocity fields, and Feature-Level Counterfactual Covariance Reduction (CCR) to latent attention features. Experiments on LIBERO, LIBERO-Plus, and a real 6-DoF robot arm report state-of-the-art results, including a 52.3% absolute success-rate gain in real-world out-of-distribution scenarios. The central claim is that these interventions 'cleanly eradicate' visual bias and preserve the pure language semantic intent.","tokens_in":26364,"tokens_out":4048,"duration_ms":48234,"significance":"If the deconfounding mechanism were correct as stated, this would be an important contribution: a general counterfactual intervention method for suppressing visual shortcuts in VLA policies, with strong evidence of OOD generalization. The proposed framework is simple and could be applied to any flow-matching-based VLA. However, the theoretical foundation is the principal added value of the paper, and it is not currently supported by the manuscript's own equations. The empirical improvements may be real, but as written they do not establish the causal deconfounding mechanism claimed. The paper does not ship code or machine-checked proofs, so the credibility of the contribution rests on the derivation and the experiments.","major_comments":[{"comment":"The central OPG formula does not perform the claimed deconfounding. Under the paper's own assumption that the counterfactual branch v_uncond captures exactly the visual bias v_bias, decompose v_cond = v_sem + v_bias with v_sem orthogonal to v_bias. Eq. (3) gives v_proj = v_bias and Eq. (4) gives v_perp = v_sem. But Eq. (5) then yields v_causal = v_cond + gamma * v_perp = (1 + gamma) * v_sem + v_bias. The spurious bias term v_bias remains with coefficient 1; it is not eradicated, only relatively downweighted. The claims in Section 3.2 that OPG 'cleanly eradicates visual bias' and 'exclusively re-weights the causal mode odds along the semantic direction' are therefore false as written. To actually remove the bias, the formula would need v_causal = v_perp or v_causal = v_cond - gamma * v_proj. Because OPG is one of the two core contributions, this is a load-bearing error.","section":"Section 3.2, Eq. (5)"},{"comment":"The marginal densities are mislabeled. The manuscript defines 'p_cond := p(A_tau | O)' and 'p_uncond := p(A_tau | T)', but the factual branch defined in Section 3.1 is conditioned on both O and T, and the counterfactual branch is conditioned on O only. The correct labels should be p(A_tau | O, T) and p(A_tau | O), respectively. This reversal is more than a typo: it indicates a sign or definitional inconsistency in the derivation of the intervention, and it should be corrected before the score/velocity composition argument can be evaluated.","section":"Section 3.2, density definitions"},{"comment":"The proof of Theorem 1 does not support the implemented CCR algorithm. The appendix establishes identifiability for the top generalized eigenspace of the pair (Sigma_Delta, Sigma_0) and uses the Sigma_0-orthogonal projector Pi_C = B(B^T Sigma_0 B)^{-1} B^T Sigma_0. In contrast, the main text Eq. (6)-(9) performs an eigendecomposition of Delta_Sigma = Sigma_cf - Sigma_f and subtracts the plain projection (F U_bias) U_bias^T using ordinary eigenvectors with eigenvalues above epsilon. Assumption 2 is stated for the whitened matrix M = Sigma_0^{-1/2} Sigma_Delta Sigma_0^{-1/2}, not for Delta_Sigma. The proof therefore does not show that the top-k eigenvectors of Delta_Sigma span S_C, nor that the plain projection in Eq. (9) removes the nuisance component while preserving S_O. The theoretical guarantee for CCR is unproved for the implemented method.","section":"Section 3.3 and Appendix A.1, CCR proof"},{"comment":"The 'strict' guarantees of CCR rest entirely on Assumptions 1 and 2, which are asserted without empirical validation or a concrete geometric argument linking the latent feature space of a VLA backbone to the assumed orthogonal decomposition. In particular, Assumption 2 requires that the eigenvalues of the whitened covariance difference on the confounder subspace are strictly larger than those on the intent subspace; the paper provides no experiment, visualization, or diagnostic that checks this spectral separation. Since the method's causal interpretation depends on this assumption holding in practice, the current manuscript does not establish that CCR actually deconfounds rather than merely applying a heuristic feature-space penalty.","section":"Section 3.3, Assumptions 1-2"}],"minor_comments":[{"comment":"Eq. (1) is the definition of conditional mutual information, not a substantive theorem; the text's phrasing 'This theorem explicitly dictates that the visual confounder intrinsically corrupts' overstates what the identity establishes.","section":"Section 3.2, Eq. (1)"},{"comment":"The caption refers to 'Contrastive Covariance Reduction' while the method is called 'Counterfactual Covariance Reduction' in the text and abstract; the terminology should be unified.","section":"Figure 2 caption"},{"comment":"Several entries in Table 2 appear to run together (e.g., '65.121.6' and 'Openvla-OFT_m: 55.6 21.781.0 92.791.0 78.668.7 67.9'), suggesting a formatting or column-alignment error that should be fixed.","section":"Table 2"},{"comment":"There are stray '/uni000...' character sequences in the text around Figure 12 and several figure captions, indicating a corrupted source-rendering artifact that should be removed.","section":"Appendix A.5"},{"comment":"The wording 'strictly deconfounded action trajectories' and 'strictly and consistently grounded in causal semantic intents' is stronger than what the current equations support; after correcting OPG, these claims should be reworded to match the implemented formula.","section":"Contributions and Section 5"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the empirical package is real, but the paper's central causal claim is undone by its own algebra. Equation (5) sets v_causal = v_cond + gamma * v_perp. Under the paper's assumptions (v_cond = v_sem + v_bias, v_uncond collinear with v_bias), v_proj equals v_bias and v_perp equals v_sem. Then v_causal = (1+gamma) v_sem + v_bias. The visual bias is not eradicated, only relatively downweighted. To actually remove it you'd need v_causal = v_perp or v_cond - gamma*v_proj. The density definitions in Section 3.2 are also swapped: p_cond is written as p(A|O) and p_uncond as p(A|T), reversed relative to the branches defined in Section 3.1. This suggests the core intervention has a sign/definition error.\n\nWhat's genuinely new: the language-masked counterfactual branch within a single forward pass, and the two mechanisms—OPG's orthogonal projection and CCR's covariance-difference suppression—are real additions beyond CAG and CAST. The experiments are extensive: LIBERO, LIBERO-Plus, and real-world tasks with ~100 trials each, plus ablations and sensitivity sweeps. The reported gains, especially the 52.3% OOD improvement, are non-trivial.\n\nSoft spots in proportion: The OPG issue is serious and load-bearing. The CCR assumptions (gain-bias decomposition, contrastive eigengap) are stated without validation, so the 'strictly preserves semantic features' guarantee is unearned. The counterfactual branch is assumed to capture exactly the visual confounder; if it contains any useful semantic signal, the subtraction throws away legitimate information. Gamma and beta are tuned on the target benchmark, and there are no error bars, code, or data. That combination means the precise magnitudes should be treated with caution.\n\nBottom line: the method may well work as an empirical recipe, but the causal interpretation as stated is not supported. This deserves a serious referee—the framework and experiments are worth engaging—but it needs major revision: fix or reframe the OPG derivation, validate the assumptions, release artifacts, and report variance. I'd send it to review, not desk reject.","headline":"A useful empirical recipe with a load-bearing theoretical flaw: Equation (5) doesn't deconfound, but the experiments deserve referee time.","tokens_in":26750,"tokens_out":3711,"would_cite":false,"duration_ms":36461,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CofactVLA deconfounds vision-language-action policies by subtracting a language-masked counterfactual branch from both latent features and generated actions, and reports a 52.3-point absolute gain in real-world out-of-distribution success.","keywords":["vision-language-action models","counterfactual intervention","vision-override phenomenon","orthogonal projection guidance","counterfactual covariance reduction","flow matching","out-of-distribution generalization","robotic manipulation"],"falsifier":"The decisive experiment is to feed the counterfactual branch a different, semantically meaningful instruction instead of a mask: if the direction of the supposed visual bias changes with the content of that wrong instruction, the branch is not language-free and the subtraction is mis-specified; alternatively, on a dataset where the base policy already follows language with no measurable vision-override, CofactVLA should leave success rates unchanged, and any drop would show the intervention deletes useful signal.","tokens_in":25670,"feed_emoji":"🤖","tokens_out":8783,"duration_ms":96177,"temperature":0.7,"pith_summary":"Vision-language-action models often execute the most visually salient action in a scene instead of following the spoken instruction; the paper names this failure the vision-override phenomenon and traces it to a spurious visual-confounder path in the policy's causal graph. CofactVLA responds by adding a language-masked counterfactual branch that predicts what the policy would do from the image alone, and then subtracting that visual-only tendency from both the latent features and the generated velocity field. The paper reports that this dual deconfounding raises the average success rate on a standard simulation benchmark to 98.5 percent, and in real-world tests under unseen conditions raises the base policy's 23.5 percent success rate to 75.8 percent, a 52.3 percentage-point absolute gain. The intended upshot is a general way to make VLA policies follow the language instruction rather than their visual habits, without requiring a new architecture.","feed_headline":"Counterfactual branch lifts robot success in unseen scenes by 52 points","feed_subtitle":"By subtracting a language-free visual prior from actions and features, the policy stops ignoring instructions.","key_machinery":"The central object is the language-masked counterfactual branch: the same VLA run in the same visual scene with the language masked out, yielding a visual-only velocity field and visual-only attention features. OPG exploits the exact score-velocity equivalence of the Gaussian flow-matching forward process so that orthogonal projection of the factual velocity field away from the counterfactual one is a legitimate re-weighting of action-mode probabilities rather than a heuristic interpolation. CCR makes the same separation at the representation level by taking the positive part of the covariance difference between the two branches, which the paper proves spans exactly the spurious visual subspace under its gain-bias and contrastive eigengap assumptions.","core_discovery":"On the paper's own terms, action generation in a VLA is a Dual-path Deconfounding Graph: the language instruction has a causal path to the action, but the dense image also opens a backdoor image-to-confounder-to-action path that overrides the instruction. To sever the backdoor, the paper constructs a counterfactual branch in which the language token is masked, treating its output as the pure visual instinct. Action-level OPG projects the factual flow velocity onto that instinct and subtracts the projection, leaving an orthogonal component that is then re-weighted as the deconfounded semantic intent; feature-level CCR takes the covariance difference between counterfactual and factual attention features, extracts its positive eigenspace as the nuisance basis, and projects the features away from it. The paper asserts that this dual intervention cleanly eradicates visual bias while preserving the valid action manifold, and supports the claim with state-of-the-art simulation scores and a 52.3-point absolute real-world OOD improvement over the base policy.","pith_inferences":["The paper leaves implicit that the same orthogonal-subtraction recipe applies to any two-modality policy where one dense modality overrides a sparse command, so the method could transfer beyond robotics to vision-and-language generation or audio-visual control.","The key untested boundary is the assumption that the image-only branch encodes no task-relevant semantics; a stronger design would estimate the confounder subspace from data instead of identifying it with the language-masked branch.","The reported sensitivity sweeps (guidance scale gamma and intervention strength beta) suggest the method's benefit depends on calibrating both strengths, so one testable extension is an automatic scheduler that adapts them per scene rather than using fixed values.","If the deconfounding truly isolates a causal language direction, then interventions on the orthogonal component should be composable: the same projection machinery could in principle damp multiple confounders (color, layout, camera angle) one at a time."],"forward_implications":["Policies trained with CofactVLA should keep following the instruction when a salient distractor object competes for the robot's attention, which is the exact failure mode the paper diagnoses.","The intervention operates inside a single forward pass, so it can be layered onto existing flow-matching VLA checkpoints instead of requiring a redesigned policy network.","The reported zero-shot results imply that the method transfers across simulation and physical hardware: the real-world OOD success rate climbs from 23.5 percent to 75.8 percent while the standard-environment rate stays at 90.8 percent.","Because the covariance and projection penalties are applied at two different levels of the stack, the two mechanisms can be ablated, tuned, or combined independently, with the paper reporting combined performance above either alone."],"supporting_citations":[{"why":"supplies the base flow-matching VLA checkpoint that CofactVLA initializes from and fine-tunes","marker":"[17]"},{"why":"provides the prior flow-matching VLA whose continuous action space this method operates in","marker":"[34]"},{"why":"is the scalar counterfactual action guidance method whose linear extrapolation OPG replaces","marker":"[40]"},{"why":"gives the classifier-free guidance reference point for composing conditional and unconditional generative distributions","marker":"[13]"},{"why":"establishes the flow-matching formulation that yields the score-velocity linear equivalence OPG relies on","marker":"[30]"},{"why":"is the standard simulation benchmark on which the main success-rate tables are measured","marker":"[41]"},{"why":"provides the perturbation benchmark used to test zero-shot out-of-distribution generalization","marker":"[10]"},{"why":"serves as a baseline VLA compared in the real-world deployment experiments","marker":"[14]"}],"fun_headline_variants":["CofactVLA: Counterfactual deconfounding boosts OOD robot success by 52%","Counterfactual intervention stops VLAs from ignoring language, lifts OOD success 52%","Deconfounding VLAs with counterfactual branches: 52% OOD gain","CofactVLA: Causal fix for vision-override, +52% OOD success","Subtracting visual confounders from VLAs yields 52-point OOD gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the language-masked, image-only branch captures exactly the visual confounder effect, with that effect collinear with the no-language velocity field or confined to the positive eigenspace of the covariance difference; if the image-only branch still carries useful semantic or affordance information, subtracting it removes legitimate signal and the deconfounded action is not guaranteed to preserve the language intent.","fun_headline_variants_meta":{"raw":{"variants":["CofactVLA: Counterfactual deconfounding boosts OOD robot success by 52%","Counterfactual intervention stops VLAs from ignoring language, lifts OOD success 52%","Deconfounding VLAs with counterfactual branches: 52% OOD gain","CofactVLA: Causal fix for vision-override, +52% OOD success","Subtracting visual confounders from VLAs yields 52-point OOD gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000499,"raw_usage":{"total_tokens":2485,"prompt_tokens":1031,"completion_tokens":1454,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":1346}},"tokens_in":647,"tokens_out":1454,"duration_ms":11528,"temperature":1.0,"reasoning_tokens":1346,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:43:26.722900+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The decisive experiment is to feed the counterfactual branch a different, semantically meaningful instruction instead of a mask: if the direction of the supposed visual bias changes with the content of that wrong instruction, the branch is not language-free and the subtraction is mis-specified; alternatively, on a dataset where the base policy already follows language with no measurable vision-override, CofactVLA should leave success rates unchanged, and any drop would show the intervention deletes useful signal.","supporting_citations":[{"cited_title":"Libero: Benchmarking knowledge transfer for lifelong robot learning","cited_arxiv_id":null,"evidence_quote":"is the standard simulation benchmark on which the main success-rate tables are measured"}],"review_version":1}