{"id":"91add1c6-47c5-4f2b-bd70-5360a47c7d30","arxiv_id":"2512.24780","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For log-sum-exp objectives, the gradient with respect to each distance is the negative softmax responsibility; the paper interprets this as implicit expectation-maximization in neural training.","lead":"This paper shows the gradient of a log-sum-exp loss with respect to each distance is the softmax responsibility, and argues this means gradient descent implicitly performs expectation-maximization. The identity is correct but textbook; the paper's stronger claims about optimization being inference are interpretations, not derivations.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 is a correct softmax derivative, but the paper's equation of gradient descent with EM conflates one gradient step on Q with EM's M-step maximization, so the central 'implicit EM' claim is unsupported.","rationale":"The identity in Theorem 1 is a textbook derivative and is not in dispute. The load-bearing step is the leap from 'responsibility-weighted gradients' to 'EM.' I checked this on the paper's own definitions: §3.2 says 'gradient descent on LSE objectives is exactly EM' while the footnote in §3.3 reduces implicit EM to responsibility-weighted updates. The stronger claim fails because a gradient step is not the M-step maximization; the weaker claim is true but reduces to the chain rule and does not support 'Bayesian structure is necessary.' The concrete test with a Gaussian mixture makes the mismatch explicit. This is the same concern the reader flagged in the second part of weakest_assumption, so I mark agreement. Since the core conceptual contribution is an overreach built on a correct algebraic identity, the reader's REJECT stands unchanged.","tokens_in":10579,"tokens_out":8069,"duration_ms":80551,"concrete_test":"Analytical check: fit a one-dimensional two-component Gaussian mixture by full-batch gradient descent on L=-Σ_i log Σ_j exp(-(x_i-μ_j)²/2σ²). At a current μ^t, compute responsibilities r_ij^t. One EM iteration updates μ_j^{EM}=Σ_i r_ij^t x_i / Σ_i r_ij^t. One gradient-descent step updates μ_j^{GD}=(1-η S_j/σ²)μ_j^t+(η/σ²)Σ_i r_ij^t x_i with S_j=Σ_i r_ij^t. Show these are unequal for any fixed η and unequal S_j; e.g., for two components with S_1≠S_2 no single η satisfies both. This settles that gradient descent is not EM's M-step.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Granting the distance semantics, the algebraic identity ∂L/∂d_j=-r_j is correct but only establishes that the score of the log marginal likelihood is the responsibility-weighted complete-data score (Fisher's identity). It does not establish that gradient descent 'is EM.' EM's M-step is θ^{t+1}=argmax_θ Q(θ|θ^t) with Q=Σ_i Σ_j r_ij^t log p_j(x_i|θ_j); a gradient step is θ^{t+1}=θ^t+η Σ_i Σ_j r_ij^t ∇_θ log p_j(x_i|θ^t). These coincide only when the step is the full maximizer, which for a Gaussian mixture would require η=σ²/Σ_i r_ij and is not the update produced by standard gradient descent. The paper's own §3.3 disclaimer ('not coordinate-ascent EM or guarantees about convergence') narrows the claim to 'responsibility-weighted parameter updates,' but §3.2, §4.1, and the abstract assert exact/implicit EM and 'optimization and inference are the same process.' Thus the central claim is not derived; it is an unsupported equivalence. The same problem breaks the 'Bayesian structure is a necessary consequence' conclusion: even if updates are responsibility-weighted, reaching a Bayesian posterior requires the M-step (or its infinitesimal limit with matching step sizes) and the missing volume/normalization terms, neither of which is provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims to explain EM-like and Bayesian behavior in neural networks by proving a gradient identity: for any objective L = log Σ_j exp(-d_j), the derivative with respect to the j-th distance is the negative softmax responsibility -r_j (Theorem 1, Eq. 5). On this basis it argues that gradient descent on such objectives is exactly/implicitly EM, that attention mechanisms and cross-entropy classification are the same mechanism under different constraints, and that Bayesian structure in trained transformers is a necessary consequence of the objective geometry. The proof of the identity is a correct application of the chain rule, but the paper's central interpretive claims are not supported by the derivation.","tokens_in":10915,"tokens_out":6932,"duration_ms":76801,"significance":"The algebraic identity in Theorem 1 is correct, and the paper is clearly written and candid about some of its limitations (e.g., §6.2, §6.4). If the paper had restricted itself to observing that LSE objectives yield responsibility-weighted gradients, it would be a modest, acceptable observation already present in the energy-based and mixture-modeling literature. As submitted, however, the paper makes much stronger claims: that gradient descent 'is EM,' that optimization and inference 'are the same process,' and that Bayesian structure in transformers is a 'necessary consequence.' These claims are not established. The manuscript therefore does not make a sound case for its headline contributions, despite the correctness of the elementary derivative.","major_comments":[{"comment":"The theorem is a textbook softmax derivative, but the inference from the identity to 'gradient descent ... is exactly EM' is not valid. Classical EM, as defined in §2.3, requires an M-step that maximizes the expected complete-data log-likelihood: θ^{t+1} = argmax_θ Σ_i Σ_j r_ij^t log p_j(x_i|θ_j). A gradient descent step is θ^{t+1} = θ^t + η Σ_i Σ_j r_ij^t ∇_θ log p_j(x_i|θ^t). These coincide only for specially chosen step sizes and parameterizations. Section 3.3 weakens 'implicit EM' to mean 'responsibility-weighted parameter updates,' which is already contained in Eq. (5). The abstract, §1.3, and §8 assert a stronger equivalence that is never derived.","section":"§3.2, §3.3, Eq. (5)"},{"comment":"Attention mechanisms do not instantiate Theorem 1 as stated. The training objective of a transformer is not L = log Σ_j exp(s_ij) over keys; softmax attention is a network layer, not a loss. The gradient of the task loss with respect to W_V or the attention scores is propagated through o_i = Σ_j α_ij v_j and downstream layers; it is not generally equal to a responsibility-weighted update of distance parameters. The claim in §4.2 that 'each backward pass updates those parameters in proportion to the responsibilities' requires specifying which objective L is being used and deriving the gradient. That derivation is missing, so the 'conditional regime' is not actually covered by Eq. (5).","section":"§4.2"},{"comment":"The conclusion that Bayesian structure in trained transformers is 'a necessary consequence of the objective geometry' is unsupported. The gradient identity constrains per-step dynamics; it does not by itself imply that fixed points are Bayesian posteriors, that attention weights converge to posterior probabilities, or that inference-time behavior matches the analytic posterior. The paper's own §6.2 notes that neural objectives omit the log-determinant volume term that is essential to Gaussian mixture likelihoods, implying the dynamics are not identical to full Bayesian mixture inference. No argument connects responsibility-weighted gradients to the sub-bit posterior agreement reported in Aggarwal et al. This is a central overclaim, not a minor caveat.","section":"Abstract, §5.2, §6.2"},{"comment":"There is a sign error in the unsupervised regime. For L = -log Σ_j exp(-d_j), the derivative is ∂L/∂d_j = +r_j, not -r_j. The text says the sign flip reflects switching from maximizing log-likelihood to minimizing its negative, but the displayed equation does not implement that flip: the derivative of -log Z is +exp(-d_j)/Z. As written, gradient descent would increase d_j for high-responsibility components, contradicting the following sentence that prototypes 'move closer.' The correct sign is needed for the claimed M-step dynamics.","section":"§4.1, Eq. (6)"}],"minor_comments":[{"comment":"Equation (6) is called the negative log marginal likelihood, but the LSE objective over distances without a volume term is not the likelihood of a full Gaussian mixture. The acknowledgment in §6.2 should be moved or previewed here to avoid implying a proper mixture model.","section":"§4.1"},{"comment":"The phrase 'This structure appears throughout deep learning... in attention mechanisms' is misleading in a paper whose theorem concerns objectives, not network layers. Attention weights are responsibilities appearing in a forward pass, not necessarily responsibilities arising as gradients of an LSE loss.","section":"§2.2"},{"comment":"The geometric substrate in §2.1 is imported from a self-citation (Oursland 2024) and is described as 'not a modeling choice' but 'a mathematical property.' Since the semantic interpretation is load-bearing for calling r_j a posterior responsibility, the prior work should be summarized sufficiently that this paper is understandable without accessing the unpublished preprint.","section":"§5.1"}],"recommendation":"reject","confidential_remarks":"The paper's central theorem is a correct but elementary derivative. The 'implicit EM' claim is made true by definition in §3.3, while the abstract and §1.3 assert a much stronger equivalence. The attention section applies the theorem to a setting where the theorem's premises do not hold, and the Bayesian-consequence claim is unsupported. These are not local presentation issues; they affect the paper's main contribution. The sign error in §4.1 is an additional correctness problem. I would not recommend revision without a fundamental reframing of the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is a clear, well-written essay around an elementary identity: for L = log Σ exp(-d_j), ∂L/∂d_j = -r_j where r is the softmax. The author knows it's a textbook derivative and says so. What's actually new is the framing — reading r_j as a posterior responsibility and using it to unify unsupervised mixture fitting, attention, and cross-entropy classification under one narrative. The prose is careful in places; Section 6.4 explicitly lists what the framework doesn't explain, and Section 3.2 defines \"implicit EM\" as responsibility-weighted gradient updates, explicitly disclaiming coordinate-ascent EM and convergence guarantees.\n\nThat last point is also the core problem. The abstract and conclusion say gradient descent \"performs expectation-maximization\" and \"optimization and inference are the same process.\" But the stress-test step is correct: EM's M-step is a maximization of the expected complete-data log-likelihood, not a gradient step. A single gradient step is not the argmax; they coincide only for special step sizes in special models. The paper's own definition of implicit EM sidesteps this, but then the title and abstract overclaim. The identity ∂L/∂d_j = -r_j is just the derivative of log-sum-exp; it does not imply that gradient descent is EM in any standard sense, nor that Bayesian structure in transformers is \"necessary.\" To get Bayesian posteriors you need volume terms and normalization, as the author admits in §6.2 and §6.4. So the strong claim collapses, and the weak claim is little more than a restatement of softmax gradient properties.\n\nThere's also the reliance on the distance-based reading from the author's own prior work. The algebra holds for any differentiable d_j, but calling r_j a \"posterior\" requires interpreting outputs as distances/energies to prototypes. That's a semantic import from an unreviewed preprint, and it does a lot of work in §2.1 and §5.1.\n\nThat said, the paper is honest about its limits, clearly written, and the taxonomy (unsupervised/conditional/constrained) is a decent organizational device. It doesn't present any new experiments or formal verification, and the contribution is essentially interpretive. For a perspective piece it might be fine, but as a research paper with \"Theorem 1\" and claims of necessity, it overreaches.\n\nWho gets value: someone looking for a gentle way to explain softmax gradients as responsibilities, or a discussion piece for a reading group on what counts as theory. It is not a result to build on.\n\nI would desk reject this. The algebra is correct but trivial, and the central equivalence is not derived. A serious referee would spend the review explaining the E-step/M-step distinction, which the author already half-acknowledges. If it resubmits, it should be reframed as an expository essay, not a theorem paper.","headline":"The algebra is right and the limits section is honest, but the paper's headline claim — that gradient descent on log-sum-exp objectives is EM — is a redefinition, not a derivation.","tokens_in":11355,"tokens_out":2424,"would_cite":false,"duration_ms":23733,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"For log-sum-exp objectives, the gradient with respect to each distance is exactly the negative posterior responsibility of that component—an algebraic identity, not an approximation. Consequently, gradient descent on such objectives perform","keywords":["log-sum-exp","expectation-maximization","gradient descent","softmax","posterior responsibilities","attention mechanisms","cross-entropy classification","mixture models"],"falsifier":"Train a transformer on a non-normalized objective (e.g., independent per-token sigmoid loss) in the same Bayesian wind tunnel setting; if it still reproduces the analytic posterior, then the claim that LSE geometry is necessary for Bayesian structure is refuted. Alternatively, compare the parameter-update trajectory of gradient descent on an LSE mixture objective with the explicit EM update trajectory; if they diverge in a way not explainable by learning rate, the 'implicit EM' equivalence is not exact.","tokens_in":10460,"feed_emoji":"🧮","tokens_out":2501,"duration_ms":27746,"temperature":0.7,"pith_summary":"This paper tries to establish that standard neural-network training objectives—when read as log-sum-exp functions over distances or energies—automatically perform expectation-maximization. The key is a simple identity: the gradient of the loss with respect to a component's distance equals the negative of that component's softmax responsibility. Because backpropagation delivers responsibility-weighted updates, the authors argue that any model trained with these objectives is implicitly running EM, with no separate inference step. A sympathetic reader would care because this would explain why attention heads specialize, why classification networks behave like mixture models, and why transformers can reproduce Bayesian posteriors—without invoking emergence or architectural accident. The paper presents this as a unification of three learning regimes under a single geometric mechanism.","feed_headline":"Softmax gradients equal posterior responsibilities","feed_subtitle":"One identity shows mixture learning, attention, and classification all run the same inference loop.","key_machinery":"The central object is the log-sum-exp (LSE) objective L = log Σ_j exp(−d_j), where d_j are distances or energies. The key identity is ∂L/∂d_j = −r_j, with r_j the softmax responsibility. This identity is the machinery: it converts a purely algebraic property of the softmax into a statement about learning dynamics. The paper interprets neural outputs as distances (prior work on Mahalanobis distance), so exp(−d_j) behaves as an unnormalized likelihood, and r_j becomes a posterior probability. The identity is then sufficient to claim that every gradient step is a responsibility-weighted update, which is the defining behavior of EM's M-step. The paper also relies on a semantic shift: probabiliti","core_discovery":"The paper's central claim is that for any objective of the form L = log Σ_j exp(−d_j), the gradient with respect to the j-th distance is exactly −r_j, where r_j = exp(−d_j) / Σ_k exp(−d_k) is the posterior responsibility of component j. This identity is derived by one application of the chain rule and holds whenever distances are differentiable. The authors then assert that gradient descent on such objectives is generalized expectation-maximization performed continuously: the forward pass computes unnormalized likelihoods and implicitly normalizes them (the E-step), while the backward pass propagates responsibility-weighted gradients to parameters (the M-step). They argue that this mechanism","pith_inferences":["A testable extension follows from the strong claim: if gradient descent on LSE objectives is truly EM, then the fixed points of training on a Gaussian mixture should match the maximum-likelihood estimates of EM; deviations would reveal where the analogy breaks.","The identity suggests that any softmax-normalized layer, including mixture-of-experts gates, is implicitly performing inference; this could lead to new architectures that explicitly control responsibilities rather than leaving them implicit.","The paper's own framing implies that the 'Bayesian geometry' seen in transformers could be deliberately turned on or off by choosing normalized versus unnormalized objectives, offering a design lever for uncertainty-aware models.","Because the derivation is algebraic, the same identity should hold for any differentiable distance function, including learned metrics or kernel distances—an extension the paper does not explore."],"forward_implications":["If correct, training any model with a softmax-based loss is doing probabilistic inference, which reframes interpretability: the assignments a network makes are literally present in the gradient signal.","The unification predicts that attention, mixture models, and classification heads share identical responsibility-weighted dynamics, so insights from one domain transfer directly to the others.","The framework explains why removing normalization (e.g., independent sigmoids) destroys soft assignment, and why objectives like maximum correntropy behave differently: they abandon the assignment structure entirely.","The missing volume term in neural objectives explains collapse as an inherent risk of the implicit-EM dynamics, pointing to where regularization or architectural constraints intervene.","The closed-world assumption of softmax—every input must be assigned—is a direct consequence of the objective geometry, not an arbitrary design choice."],"fun_headline_variants":["Gradient descent secretly runs EM on softmax losses","Softmax gradients are posterior responsibilities—training is implicit EM","Why neural nets cluster, attend, and classify: implicit EM","Log-sum-exp losses turn gradient descent into EM","The hidden EM algorithm inside every softmax layer"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that neural outputs should be read as distances or energies to learned prototypes, rather than as confidences or scores; if that semantic interpretation fails, the 'posterior responsibility' label is just a softmax weight and the EM claim loses its probabilistic force.","fun_headline_variants_meta":{"raw":{"variants":["Gradient descent secretly runs EM on softmax losses","Softmax gradients are posterior responsibilities—training is implicit EM","Why neural nets cluster, attend, and classify: implicit EM","Log-sum-exp losses turn gradient descent into EM","The hidden EM algorithm inside every softmax layer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1293,"prompt_tokens":758,"completion_tokens":535,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":458}},"tokens_in":502,"tokens_out":535,"duration_ms":5272,"temperature":1.0,"reasoning_tokens":458,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T13:14:10.101131+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a transformer on a non-normalized objective (e.g., independent per-token sigmoid loss) in the same Bayesian wind tunnel setting; if it still reproduces the analytic posterior, then the claim that LSE geometry is necessary for Bayesian structure is refuted. Alternatively, compare the parameter-update trajectory of gradient descent on an LSE mixture objective with the explicit EM update trajectory; if they diverge in a way not explainable by learning rate, the 'implicit EM' equivalence is not exact.","supporting_citations":[],"review_version":1}