{"id":"2de3fb0e-993d-4a46-82d7-cd432e1b00a8","arxiv_id":"2602.02239","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Interpretability for time series models should require semantic alignment — human-meaningful variables and mechanisms whose alignment persists under temporal evolution.","lead":"This position paper argues that interpretability in deep time series models should mean semantic alignment: predictions expressed in terms of human-meaningful variables and mechanisms, preserved over time. It formalizes this requirement and offers a blueprint for building such models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Definition 1 presumes dynamic concepts are deterministic functions of the observed past (σ(X≤t)-measurable); for stochastic concepts the equality in Eq. (5) is unsatisfiable, so the formal core of the paper's central claim is not operational as stated.","rationale":"The reader identified the stability of human concepts as the weakest assumption, but the more precise fragility is the σ(X≤t)-measurability built into Eq. (5). Even granting stable, shared concepts, the definition is satisfiable only if future concept values are deterministic functions of the past. This is not a minor technicality: the paper's motivating examples (industrial monitoring, clinical reasoning) involve stochastic latent processes, and the blueprint's concept propagation also explicitly treats concepts as random variables. The central novel claim — that temporal alignment must be preserved — becomes either impossible or trivial under this definition. A position paper can propose ideals, but this one presents Definition 1 as the formal target; if the target is empty for the intended applications, the paper's practical relevance is undermined. I recommend CONDITIONAL: the argument is worth publishing, but condition (ii) should be relaxed (e.g., alignment of conditional distributions P(Z_{t+1}|X≤t)=P(C^Z_{t+1}|X≤t)) or scoped explicitly to deterministic concepts. The proposed counterexample would settle which reading is intended.","tokens_in":17997,"tokens_out":7691,"duration_ms":72969,"concrete_test":"Construct the minimal linear-Gaussian counterexample: let C_{t+1}=0.9C_t+η_{t+1}, X_t=C_t+ε_t, with η,ε iid Gaussian centered. Take any model from Eqs. (1)-(3) with Z_{t+1}=f(X≤t) (F_t-measurable). Compute P(Z_{t+1}=C_{t+1}|X≤t); since C_{t+1} has positive conditional variance given X≤t, this probability is 0 for continuous variables (or <1 for discrete). If the check confirms, Definition 1 is vacuous for stochastic dynamic concepts, and the paper must either add a distributional version of Eq. (5) or explicitly restrict its scope.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The weakest point is in Definition 1 (Eq. 4–5, Sec. 3.2). The paper first sets up time series as realizations of a stationary stochastic process and the predictive task as modeling P(Y|x≤t) (Sec. 2.1), and concepts as stochastic processes (Sec. 3.1). It then requires C^Z_{t+1} to be σ(X≤t)-measurable and demands P(Z_{t+1}=C^Z_{t+1}|X≤t)=1 a.s. Since Z_{t+1} is produced by propagation from X≤t, this forces the future dynamic concept to be a.s. equal to a function of the observed past. Any dynamic concept with irreducible aleatoric uncertainty — e.g., thermal stress affected by unobserved load, vital signs with measurement noise, or any stochastic system — violates condition (ii): the conditional probability is <1 for every model in the considered class. The paper does not flag this determinism assumption; Observation 3.1 discusses only temporal stability of semantics, not σ(X≤t)-measurability. Consequently, the formal definition is not operational for the very class of probabilistic forecasting problems that motivate the paper. Either alignment must be stated distributionally (e.g., posterior over concepts matching model posterior), or the paper must explicitly restrict SA to deterministic concepts and acknowledge that the central temporal claim does not apply to stochastic dynamics.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a position paper arguing that interpretability of deep time series models should be defined by semantic alignment (SA): internal variables and mechanisms should correspond to human-meaningful concepts, and this correspondence must be preserved under temporal evolution. It formalizes SA in Definition 1 (concepts) and Definition 2 (mechanisms), proposes a blueprint extending concept bottleneck models to time series via source concepts, temporal and spatio-temporal propagation mechanisms, and a three-term training objective, and discusses implications for trust, intervention, counterfactuals, and interaction with symbolic knowledge. An extended appendix reviews existing post-hoc and intrinsic interpretability paradigms and classifies them against SA.","tokens_in":18339,"tokens_out":8248,"duration_ms":72755,"significance":"If the central formal notion were fully operational, the paper would make a useful contribution: it isolates semantic from structural opacity in a time-series context, introduces a temporal invariance requirement absent from static concept-based models, and offers a concrete architectural template. The writing is clear and the literature review is broad; Table 1 is a useful synthesis. As a position paper, it does not provide empirical demonstrations, and several statements are explicitly speculative; that is appropriate for the genre. The value of the paper currently hinges on whether Definition 1 can be made to apply to the probabilistic settings the paper itself motivates. With that repair, the paper would be a good anchor for future work.","major_comments":[{"comment":"The definition requires C^Z_{t+1} to be σ(X≤t)-measurable and P(Z_{t+1}=C^Z_{t+1}|X≤t)=1 a.s. Combined with the model class in Eq. (2), where Z_{t+1} is computed from X≤t, this forces every dynamic concept to be an a.s. deterministic function of the observed past. But Sec. 3.1 defines concepts as general stochastic processes, and the motivating applications (physiological or financial signals with noise or unobserved drivers) have dynamic concepts with irreducible aleatoric uncertainty given X≤t. For such concepts, P(Z_{t+1}=C^Z_{t+1}|X≤t)<1 for every model in the class, so Definition 1 is not operational for the very probabilistic forecasting problems the paper sets up in Sec. 2.1. Observation 3.1 addresses only temporal stability of semantics, not σ(X≤t)-measurability. Please either restate alignment distributionally (e.g., P(Z_{t+1}|X≤t)=P(C^Z_{t+1}|X≤t)) or explicitly restrict Defini","section":"Sec. 3.2, Definition 1, Eq. (5)"},{"comment":"Definition 2 is a placeholder: it asserts P(V|V') ∈ M^h_{V|V'}, but M^h_{V|V'} is not constructed, verified, or exemplified. As written, any mechanism can be declared aligned for a suitable choice of M^h, so the definition carries no formal content and cannot be falsified. The surrounding text says human knowledge often specifies only constraints, but it does not say how those constraints become M^h. The paper should either specify a concrete construction (e.g., M^h as the set of distributions satisfying a list of monotonicity, sparsity, or physics constraints) or explicitly state that Definition 2 is a schema to be instantiated per domain. As it stands, the claimed separation of variable alignment and mechanism alignment is not yet a definition.","section":"Sec. 3.2, Definition 2"}],"minor_comments":[{"comment":"The loss equation uses L_spatial and L_temporal, but the bullets define L_concept and L_prop. The notation should be harmonized.","section":"Sec. 4.3"},{"comment":"The sentence 'This is not a limitation' is too strong; stable semantics is a substantive assumption. It should be phrased as an explicit assumption, and the relaxation mentioned should be described as a limitation.","section":"Observation 3.1"},{"comment":"The text says U_t and Z_t are realizations of stochastic processes, but Eqs. (1)-(2) present them as deterministic functions of x≤t. Please clarify whether stochasticity comes only from the input process or also from model-internal randomness.","section":"Sec. 2.2"},{"comment":"The assertion that omitting Eq. (5) 'could lead the SA to decay exponentially over time' is unsubstantiated. If it is a conjecture, mark it as such; if it is a theorem, provide a statement and proof sketch.","section":"Sec. 3.2"},{"comment":"The legend uses ✓/∼/✗ but the criteria for 'partially satisfies' are not defined. Add a sentence or footnote explaining how partial satisfaction is judged.","section":"Table 1"},{"comment":"The phrase 'aligned with C^U_t and C^Z_t, respectively' could be ambiguous because condition (ii) uses C^Z_{t+1}; consider aligning the notation consistently.","section":"Definition 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid position piece, but the formal definition needs substantive repair before publication. The stochastic-concept measurability issue is the main blocker; Definition 2 should also be tightened or labeled as a schema. I see no reason to reject: the conceptual contribution is valuable and the fixes are within scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a better-than-usual position paper. It gives time series interpretability something it lacked — a formal definition of semantic alignment that includes a temporal persistence condition (Eq. 5), not just static concept matching. The distinction between instantaneous and dynamic concepts is real, and the blueprint with the extra temporal loss term is concrete enough to build on. The taxonomy in the appendix and Table 1 are handy. The authors are honest that the blueprint is speculative; that is fine for the genre.\n\nThe main soft spot is the formal core. In Definition 1, they require C^Z_{t+1} to be σ(X≤t)-measurable and demand P(Z_{t+1}=C^Z_{t+1}|X≤t)=1. Since Z_{t+1} is a function of X≤t through the model class, this forces the future dynamic concept to be deterministically determined by the observed past. Any concept with irreducible aleatoric uncertainty — vital signs with measurement noise, financial series with unobserved drivers, any latent-state process — violates the condition for every model in the class. The paper never flags this. Observation 3.1 discusses the temporal stability of concept semantics, which is a different issue. So the headline claim that alignment must be preserved under temporal evolution is stated much more broadly than the definition actually supports. The stress-test note lands.\n\nThat said, the flaw is fixable and not fatal to the paper's purpose. A distributional version of alignment, or an explicit scoping to deterministic concepts, would patch it. But as written, the central definition does not cover the probabilistic forecasting problems the paper uses as motivation.\n\nOther soft spots are minor. Definition 2 is a placeholder: a mechanism is aligned if it lies in a set of human-admissible distributions, but how that set is built or verified is left open. The accuracy-interpretability rebuttal leans on prior CBM results, which is acceptable in a position piece but not evidence. No experiments, but none are promised.\n\nWho this is for: people working on concept bottleneck models, interpretable time series, and anyone tired of saliency maps as the default answer. It is worth a serious referee. I would send it out, with a request that the authors either relax Eq. (5) or explicitly restrict the definition to deterministic concepts and say what that costs them.","headline":"A genuinely useful position paper on semantic alignment for time series, but Eq. (5) is secretly a determinism assumption that the authors never flag.","tokens_in":18776,"tokens_out":1819,"would_cite":true,"duration_ms":20544,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Time series AI is interpretable only when it thinks in human concepts","keywords":["semantic alignment","time series interpretability","concept-based models","temporal propagation","mechanism alignment","concept drift","trustworthy AI","explainability"],"falsifier":"In a real forecasting application, ask domain experts to label the intended concept (say, 'thermal stress') at many timesteps and test whether experts agree with each other and with their own past labels, and whether a model whose states align at time t continues to align at t+1. If experts' labels drift or agree poorly, the probabilistic equality in Definition 1 cannot be satisfied; if a model that fails the equality is still judged interpretable by its users, the definition is too strong.","tokens_in":17934,"feed_emoji":"🕰️","tokens_out":6295,"duration_ms":60039,"temperature":0.7,"pith_summary":"Deep time series models are accurate but opaque, and this position paper argues that the dominant interpretability toolkit—saliency maps, attention weights, surrogates, counterfactuals—answers a different question than the one practitioners care about. The authors claim that interpretability in this setting is semantic alignment: a model's internal variables must equal the concepts a domain expert actually reasons with (for example, 'thermal stress accumulated' instead of 'hidden unit 47'), and this equality must survive the model's own temporal propagation. They formalize this as two probabilistic equalities, one for instantaneous concepts and one for dynamic concepts, plus a separate requirement that the mechanisms relating those concepts honor human constraints such as monotonicity or known physical laws. If accepted, the definition turns interpretability into a structural, verifiable property rather than a post-hoc story, and gives model designers a concrete blueprint: predict through a bottleneck of expert-meaningful concepts whose evolution is explicitly supervised. The reader should care because the paper is setting a new evaluation target for a field where 'explainability' currently means many incompatible things.","feed_headline":"Time series AI is interpretable only when it thinks in human concepts","feed_subtitle":"A formal definition requires model states to match expert concepts and to keep matching as predictions unfold.","key_machinery":"Definition 1 (semantic alignment of concepts) is the load-bearing object: it requires P(Ut=CUt|X≤t)=1 and P(Zt+1=CZt+1|X≤t)=1, turning 'interpretable' into a probabilistic equality between model variables and human concepts rather than a human judgment about explanations. The temporal clause—that alignment must be preserved one step ahead under propagation—is what distinguishes the time-series setting from static concept-bottleneck models. Definition 2 (semantic alignment of mechanisms) makes mechanism interpretability a constraint-satisfaction problem: a mechanism is aligned if its conditional distribution belongs to the user-specified admissible set. Together these definitions provide the","core_discovery":"The paper's central claim is a definition. A deep time series model is semantically aligned when (i) the encoder's representations Ut equal instantaneous human concepts CUt with probability one given the observed past, and (ii) the propagated state Zt+1 equals the dynamic concept CZt+1 with probability one given the observed past. The second condition—preservation of alignment under temporal evolution—is the constraint with no analogue in static settings; without it, a representation that matches the intended concept at time t can silently become an unrelated concept at time t+1. The paper pairs this with a definition of mechanism alignment: a mechanism is aligned when the conditional distri","pith_inferences":["A practical corollary the paper does not spell out: alignment can be measured continuously, for example as the expected distance between model states and expert-labelled concept trajectories, so the definition could be turned into a benchmark metric rather than a yes/no criterion.","The assumption of temporally stable concept semantics is likely the practical bottleneck; in domains with regime shifts, concepts like 'recession' or 'overheating' change meaning, so Definition 1 would require explicitly modelling concept drift, not just model dynamics.","The intervention story (correct a concept value and roll the dynamics forward) presupposes that learned propagation mechanisms are causal, not merely predictive; enforcing alignment of conditional distributions may not be enough for genuine counterfactual reasoning.","By making interpretability a property of the model rather than the explanation, the definition shifts research effort from post-hoc analysis to annotation: scalable collection of expert concept labels over time becomes the enabling resource, possibly via automated or language-model-assisted annotation workflows."],"forward_implications":["Users can intervene on aligned models in concept space: correcting a concept value at time t updates all future concept estimates, enabling temporally consistent counterfactuals and targeted mechanism repair without retraining.","Aligned concepts enable formal verification (model checking) and fairness analysis, because internal states correspond to explicit domain-relevant properties that can be checked and traced over time.","Omitting the propagation supervision term allows concept drift: per-step concept accuracy is not enough to keep alignment from decaying as the model steps forward.","Residual pathways that process unconstrained representations in parallel with the concept space let semantically aligned models retain the expressivity of black boxes, so interpretability and accuracy need not trade off.","By the paper's own comparison table, existing paradigms—attention, saliency, surrogates, counterfactuals, prototypes, linearized dynamics, and most physics-informed or symbolic models—fail at least one of the two alignment conditions, so none of them qualifies as making time series models semantically interpretable."],"fun_headline_variants":["AI time series models need to think in human concepts to be interpretable","Interpretable time series AI must keep its states matched to human concepts","Deep time series interpretability hinges on sustained semantic alignment","Time series AI is interpretable only if its reasoning stays in human terms"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that human reasoning about a domain can be modeled as random variables whose semantics remain stable over time, so that 'does the model's state equal the human concept?' is a well-defined yes/no question at every step.","fun_headline_variants_meta":{"raw":{"variants":["AI time series models need to think in human concepts to be interpretable","Interpretable time series AI must keep its states matched to human concepts","Deep time series interpretability hinges on sustained semantic alignment","Time series AI is interpretable only if its reasoning stays in human terms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1136,"prompt_tokens":654,"completion_tokens":482,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":398,"completion_tokens_details":{"reasoning_tokens":408}},"tokens_in":398,"tokens_out":482,"duration_ms":4700,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:24:05.226104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a real forecasting application, ask domain experts to label the intended concept (say, 'thermal stress') at many timesteps and test whether experts agree with each other and with their own past labels, and whether a model whose states align at time t continues to align at t+1. If experts' labels drift or agree poorly, the probabilistic equality in Definition 1 cannot be satisfied; if a model that fails the equality is still judged interpretable by its users, the definition is too strong.","supporting_citations":[],"review_version":1}