{"id":"2b3fa7c8-5387-47ae-8ff1-dcc8959bc7cd","arxiv_id":"2412.07331","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"NeSyA integrates symbolic finite automata with neural perception under exact probabilistic semantics, achieving faster and more accurate neurosymbolic sequence classification than existing systems.","lead":"The paper introduces NeSyA, a hybrid model that combines symbolic automata with neural networks to classify and label sequences using both learned perception and hand-written temporal rules. It reports large speedups over comparable neurosymbolic systems and better generalization than purely neural models on a real-world event recognition task.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CAVIAR generalization claim is compromised because high-level labels are generated by the same SFA that NESYA uses; the claimed benefit over neural baselines may be an artifact of this circularity.","rationale":"The reader's weakest_assumption identifies the independence of per-symbol probabilities in Eq. (1) as the primary load-bearing concern. That is a legitimate modeling assumption, but it is not the most load-bearing for the central claim: a neural network can often compensate for independence misspecification when the task is classification, and the synthetic experiments are fully controlled. The CAVIAR label-generation circularity is more directly damaging because it undermines the paper's stated real-world generalization contribution. The paper openly states that the high-level labels are generated by the same SFA used in NESYA, making the symbolic component perfectly aligned with the target. The neural baselines lack this oracle, so the comparison is structurally biased. The reader did list this issue in the rationale but did not elevate it to the weakest assumption; I disagree with that prioritization. However, the overall verdict remains CONDITIONAL: the method and synthetic results are sound, but the real-world claim requires verification on human-annotated labels or at minimum a clear disclosure of the circularity as a limitation.","tokens_in":12295,"tokens_out":11538,"duration_ms":117075,"concrete_test":"Re-run the CAVIAR experiment using the original human-annotated high-level event labels instead of the SFA-generated ones as training targets, keeping all other protocol details identical (same features, architectures, learning rates, seeds, early stopping). If NESYA's macro-F1 advantage over CNN-LSTM and CNN-Transformer disappears or shrinks substantially, the claimed generalization benefit is an artifact of self-referential labels. Report mean and standard deviation over 10 seeds per learning rate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4.2, the paper states: 'Using the labels generated by the SFA, we assume perfect knowledge, i.e. that the symbolic component of NESYA can perfectly retrieve the high-level events, given the low-level activities.' This makes the CAVIAR evaluation circular: the target labels are produced by exactly the symbolic automaton that NESYA embeds, from ground-truth low-level activities. NESYA therefore receives its symbolic transition structure essentially for free, while the CNN-LSTM and CNN-Transformer baselines must learn the mapping from observations to high-level events entirely from data. The resulting claim that NESYA 'provides benefits in terms of generalization compared to purely neural systems' in a 'real-world event recognition task' is not evidence about generalizing to human-annotated events; it demonstrates only that injecting the exact label-generating automaton aids low-data learning. Because the central claim explicitly cites real-world generalization, this circularity is load-bearing and leaves that subclaim unverified. The original CAVIAR annotations are noisy and are discarded, so the experiment does not evaluate event recognition as annotated by humans.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NeSyA, a neurosymbolic model for sequence classification and tagging that combines symbolic automata with neural perception. It defines a probabilistic semantics via weighted model counting over transition formulas, uses knowledge compilation to make transition probabilities differentiable, and computes acceptance probabilities with a matrix-based forward recursion analogous to HMM inference. The authors prove Theorem 1 relating the forward probabilities to sums over trace probabilities, and evaluate on a synthetic driving domain against DeepStochLog, DeepProblog, and FuzzyA, and on the CAVIAR event recognition domain against CNN-LSTM and CNN-Transformer baselines.","tokens_in":12508,"tokens_out":6184,"duration_ms":62573,"significance":"The clean probabilistic semantics and exact, compilation-based inference for temporal neurosymbolic reasoning are valuable; Theorem 1 and the matrix-based recursion provide a solid formal foundation, and the code is released. The synthetic experiments show large scaling and accuracy margins over existing NeSy systems. However, the CAVIAR experiment does not currently support the claimed real-world generalization benefit because the high-level labels are generated by the same SFA that NeSyA uses, and the asserted equivalence with DeepStochLog is not demonstrated.","major_comments":[{"comment":"The CAVIAR evaluation is circular for the generalization claim: the paper states that 'Using the labels generated by the SFA, we assume perfect knowledge' and that the human-annotated labels are discarded because they are noisy. Since the same SFA both generates the training targets and performs temporal decoding in NeSyA, the comparison against CNN-LSTM and CNN-Transformer does not measure real-world event recognition; it measures the benefit of injecting the exact label-generating automaton into the model. The conclusion that NeSyA 'provides benefits in terms of generalization compared to purely neural systems in a real-world event recognition task' is therefore unsupported. Please either re-run with the original human annotations and a noise-handling procedure, or clearly reframe the experiment as a study of perfect-knowledge injection and remove the real-world generalization claim.","section":"Section 4.2 (Event Recognition)"},{"comment":"The paper asserts without derivation that NeSyA, DeepStochLog, and DeepProblog are 'equivalent in their computation and learning setup' and therefore have identical accuracy, and it omits accuracy comparisons with those systems. This equivalence is not obvious because DeepStochLog is based on stochastic logic programs with a different inference mechanism, and DeepProblog is atemporal. Without a formal statement or a reference establishing the equivalence, the reported runtime gaps cannot be attributed solely to the inference scheme, and the accuracy claim is unverified. Please add a derivation or proof, or report accuracy results for those systems.","section":"Section 4.1 (Synthetic Driving)"}],"minor_comments":[{"comment":"There is a typo in the running example: 'We is is α2 = ...' should read 'We then compute α2 = ...'.","section":"Section 3.1"},{"comment":"The word 'interptetaion' in the proof of Theorem 1 should be 'interpretation'.","section":"Appendix A"},{"comment":"The table header is ambiguous: the columns 'Pattern / Sequence Length / Method' with three numeric subcolumns are difficult to parse; please restructure the table so that the three sequence lengths are clearly separate.","section":"Table 1"},{"comment":"The independence assumption in Eq. (1), namely that symbol probabilities factor as a product of independent per-symbol probabilities, is stated but never discussed as a limitation; a sentence on its implications for correlated perception errors would be helpful.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The circularity in the CAVIAR experiment is the main obstacle to the paper's real-world generalization claim. The synthetic scaling and accuracy results are solid, and the formal development in Sections 2 and 3 is sound. The paper would be acceptable after reframing the CAVIAR claim or re-running with human annotations, and after supporting or softening the equivalence claim with DeepStochLog. The paper fits the scope of the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"NeSyA is a clean, useful piece of work. The new thing is the specific combination: symbolic finite automata as the symbolic layer, a neural network as perceiver, probabilistic semantics via independent per-symbol probabilities, and exact inference through knowledge compilation plus a matrix forward recursion. That combination is not in the cited FUZZYA, DeepStochLog, or DeepProbLog, and the α-recursion proof in the appendix is correct. The synthetic driving benchmark is the strongest part: NeSyA beats FUZZYA on accuracy with large margins as sequence length grows, and it is orders of magnitude faster than DeepStochLog; the code is public. Credit where due: the formal development is honest, the paper ships code, and the authors are transparent about discarding the noisy CAVIAR labels.\n\nThe soft spots are real. Most important is CAVIAR: the high-level event labels are generated by the same SFA that NeSyA embeds, from ground-truth low-level activities. NeSyA therefore gets the symbolic transition structure for free, while the CNN-LSTM and CNN-Transformer baselines must learn the mapping from observations to high-level events from data. The claimed 'generalization benefits' over purely neural systems in a real-world task is not established; the experiment shows that injecting the exact label-generating automaton helps in a low-data regime. The authors even say they assume perfect knowledge, so this is a conscious design choice, but the abstract's wording overstates what is measured. The equivalence claim with DeepStochLog (footnote 1: 'they generate the same results') is asserted without derivation; likely true given the probabilistic semantics and the automaton/grammar connection, but it should be shown or referenced. Figure 4 has no error bars; for runtime that is a minor issue. The independence assumption in Eq. (1) is standard for WMC and worth stating as a limitation; it could cause miscalibration with correlated perception errors, but it does not invalidate the method as a training objective.\n\nWho should read this: anyone working on temporal neurosymbolic learning or sequence models with background knowledge. It is a solid submission for a main conference or a good journal; a serious referee will find the algorithm sound and the claims mostly testable. I would not desk-reject it. My recommendation: send it to review, with a request that the CAVIAR claim be reframed or that the experiment be redone with human labels or a held-out SFA, and that the DeepStochLog equivalence be made precise. Those are major but addressable revisions.","headline":"Clean, useful algorithm for temporal neurosymbolic learning; the synthetic benchmark is solid, but the CAVIAR generalization claim rests on circular label generation.","tokens_in":13038,"tokens_out":1861,"would_cite":true,"duration_ms":19363,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68Q45","68T27"],"pacs":[],"model":"deepseek-v4-flash","headline":"NeSyA combines neural perception with symbolic automata under probabilistic semantics to give end-to-end differentiable sequence classification that scales and generalizes better than prior neurosymbolic systems.","keywords":["neurosymbolic AI","symbolic automata","sequence classification","sequence tagging","probabilistic semantics","knowledge compilation","weighted model counting","weak supervision"],"falsifier":"A direct test is to construct a synthetic perception stack whose symbol errors are correlated, for example where two symbols are misread together far more often than independently, while the ground-truth trace distribution is known, and compare NeSyA's predicted acceptance probabilities against empirically observed acceptance frequencies. If the predictions are systematically off whenever correlations are strong, the product-form independence assumption is the cause; a variant that models joint symbol probabilities should outperform it.","tokens_in":12098,"feed_emoji":"🔁","tokens_out":8900,"duration_ms":77314,"temperature":0.7,"pith_summary":"The paper proposes NeSyA, a neurosymbolic model for sequence classification and tagging that couples a neural network perceiving each time step with a symbolic automaton whose transitions are propositional logic formulas. The central claim is that giving the automaton clean probabilistic semantics, so that each transition probability is a weighted model count and the state distribution follows an HMM-style forward recursion, makes the whole system end-to-end differentiable and efficient to train from sequence-level labels only. The authors argue this combination scales better than prior neurosymbolic systems on temporal problems and is more accurate than a fuzzy-logic alternative on a synthetic driving benchmark, and that the injected symbolic knowledge improves generalization over purely neural sequence models on a real-world event-recognition task. A sympathetic reader would take the paper's contribution to be a reusable recipe: compile the temporal knowledge into a symbolic automaton, compile each logical transition into a tractable circuit, and run matrix-based forward inference.","feed_headline":"Symbolic automata make neurosymbolic sequence learning scale","feed_subtitle":"Probabilistic inference over logic-labeled transitions trains from sequence labels alone and beats prior baselines.","key_machinery":"The device carrying the argument is a deterministic symbolic automaton: a finite automaton whose transitions are labeled by propositional formulas over a fixed set of symbols, with exactly one outgoing transition true for every interpretation. NeSyA treats the neural net's output at each time step as independent symbol probabilities, turns each transition formula into an arithmetic circuit via knowledge compilation, computes a transition matrix of weighted model counts per time step, and updates the state distribution by vector-matrix multiplication. This is the same forward recursion as a hidden Markov model, but with transition probabilities derived from logical formulas rather than a fixed table. The recursion is exact, differentiable, parallelizable on GPUs, and polynomial in the number of compiled-circuit nodes and automaton states, which is what gives the scalability advantage over systems that reduce the whole temporal query to a flat propositional formula.","core_discovery":"The central discovery is that deterministic symbolic automata, whose transitions are propositional formulas over a set of variables, can be composed with a neural perception network under a probabilistic semantics such that the probability of an input sequence being accepted is computed exactly and differentiably. Given a neural probability vector $p_t$ over symbols for each observation, the probability of a transition formula is the weighted model count of that formula; compiling each formula once into a d-DNNF (decomposable negation normal form) arithmetic circuit makes every transition probability linear-time, and the standard forward recursion over automaton states then yields $\\alpha_t(q)$, the probability of being in state $q$ after $t$ observations. The paper proves that this recursion equals the sum over all traces that lead to state $q$, weighted by their product-of-per-step probabilities, so summing $\\alpha$ over accepting states gives exactly the probability that the symbolic automaton accepts the trace. Training uses weak supervision: only sequence-level binary labels (or frame-level high-level labels) are needed, and gradients flow through the arithmetic circuits and matrix products into the neural network.","pith_inferences":["Beyond the paper, the same construction suggests a modular route to temporal neurosymbolic AI: any high-level temporal language that compiles to a deterministic symbolic automaton inherits exact differentiable probabilistic inference, much as tractable circuit representations did for atemporal formulas.","Beyond the paper, the independence assumption is the obvious bottleneck; a natural testable extension is to let the neural network emit a richer joint distribution over symbols per time step, or to add a learned correlation term, and measure whether the accuracy gains persist.","Beyond the paper, because the automaton state distribution is computed per time step, NeSyA could be plugged into reinforcement-learning pipelines as a differentiable temporal constraint or reward-shaping signal, which the paper itself suggests as future work.","Beyond the paper, the event-recognition results use a three-state automaton and simple transition logic; a reader should not generalize to complex temporal knowledge until experiments cover larger automata and longer, noisier sequences."],"forward_implications":["NeSyA lets a temporal symbolic constraint be used as a differentiable loss: a convolutional network learns symbol grounding from sequence-level labels alone, without any per-symbol annotations.","For the tested synthetic patterns, NeSyA updates orders of magnitude faster than the stochastic logic programming and probabilistic logic programming baselines, making exact probabilistic neurosymbolic inference practical on sequences that are too long for those systems.","Against the fuzzy-logic alternative, NeSyA's probabilistic semantics give higher accuracy, with the gap growing as sequences lengthen.","On the CAVIAR event-recognition task, injecting the automaton's logical structure gives NeSyA competitive or better macro-F1 than recurrent and transformer sequence models while using roughly an order of magnitude fewer parameters, with smaller train-test gaps.","The alpha-semantics theorem makes the compiled symbolic automaton a tractable device for exact queries about uncertain symbolic sequences, so it can serve as a compilation target for temporal neurosymbolic programs."],"supporting_citations":[{"why":"Supplies the fuzzy-logic baseline and the experimental protocol that NeSyA is compared against for accuracy and runtime on the synthetic benchmark.","marker":"[Umili et al., 2023b]"},{"why":"Provides the stochastic logic programming baseline whose scalability NeSyA benchmarks against on temporal patterns.","marker":"[Winters et al., 2022]"},{"why":"Defines the weak-supervision learning setup and the flat probabilistic logic inference that NeSyA contrasts with for temporal domains.","marker":"[Manhaeve et al., 2018]"},{"why":"Supplies the knowledge compilation map that motivates compiling transition formulas into tractable circuits.","marker":"[Darwiche and Marquis, 2002]"},{"why":"Defines d-DNNF circuits, the tractable representation used to compute weighted model counts for transition probabilities.","marker":"[Darwiche, 2001]"},{"why":"Establishes the reduction of probabilistic logical inference to weighted model counting, which NeSyA uses for each transition.","marker":"[Chavira and Darwiche, 2008]"},{"why":"Provides the maximum-entropy Markov model formulation whose forward recursion NeSyA adopts for state updates.","marker":"[McCallum et al., 2000]"},{"why":"Supplies the event calculus definitions for CAVIAR that the paper encodes as the symbolic automaton for event recognition.","marker":"[Artikis et al., 2014]"}],"fun_headline_variants":["NeSyA merges neural nets and symbolic automata for sequence tasks","Symbolic automata learn sequences from labels alone","NeSyA: probabilistic automata for neural sequence learning","Weak supervision trains symbolic automata via exact probability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes the symbols at each time step are perceived independently: the probability of a full interpretation is the product of per-symbol probabilities, so correlated perception errors are not modeled. If real perception outputs have such correlations, the computed acceptance probabilities will be miscalibrated and training may chase the wrong signal.","fun_headline_variants_meta":{"raw":{"variants":["NeSyA merges neural nets and symbolic automata for sequence tasks","Symbolic automata learn sequences from labels alone","NeSyA: probabilistic automata for neural sequence learning","Weak supervision trains symbolic automata via exact probability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000698,"raw_usage":{"total_tokens":3129,"prompt_tokens":895,"completion_tokens":2234,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":2168}},"tokens_in":511,"tokens_out":2234,"duration_ms":62132,"temperature":1.0,"reasoning_tokens":2168,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:55:19.946086+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to construct a synthetic perception stack whose symbol errors are correlated, for example where two symbols are misread together far more often than independently, while the ground-truth trace distribution is known, and compare NeSyA's predicted acceptance probabilities against empirically observed acceptance frequencies. If the predictions are systematically off whenever correlations are strong, the product-form independence assumption is the cause; a variant that models joint symbol probabilities should outperform it.","supporting_citations":[{"cited_title":"Deepstochlog: Neural stochastic logic programming","cited_arxiv_id":null,"evidence_quote":"Provides the stochastic logic programming baseline whose scalability NeSyA benchmarks against on temporal patterns."},{"cited_title":"Deepproblog: Neural probabilistic logic programming","cited_arxiv_id":null,"evidence_quote":"Defines the weak-supervision learning setup and the flat probabilistic logic inference that NeSyA contrasts with for temporal domains."},{"cited_title":"Decomposable negation normal form","cited_arxiv_id":null,"evidence_quote":"Defines d-DNNF circuits, the tractable representation used to compute weighted model counts for transition probabilities."},{"cited_title":"Maximum entropy markov models for information extraction and segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the maximum-entropy Markov model formulation whose forward recursion NeSyA adopts for state updates."},{"cited_title":"An event calculus for event recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the event calculus definitions for CAVIAR that the paper encodes as the symbolic automaton for event recognition."}],"review_version":1}