{"id":"931db738-7f31-4e03-806a-075a7000e3a0","arxiv_id":"2607.08136","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ASPEn integrates ASP stable-model semantics with energy-based models for joint discrete-continuous optimisation and end-to-end training on visual reasoning and multi-object tracking.","lead":"The paper presents ASPEn, a framework that couples Answer Set Programming with energy-based models so that logical constraints and non-monotonic reasoning shape continuous learning. It aims at end-to-end neurosymbolic systems for perception and tracking tasks such as CLEVR and MOT.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The end-to-end claim rests on an untested assumption that ASP-derived free energies produce usable gradients for the neural energy factors under realistic combinatorial load.","rationale":"The reader correctly isolates the weakest assumption: that the ASP-derived free-energy contrast supplies usable gradients and remains tractable. My stress-test simply makes that assumption more precise by noting that none of the three empirical sections actually exercises the claimed joint training loop on a non-trivial combinatorial space. MNIST is too small; CLEVR and MOT train the neural components with ordinary supervised or pairwise losses and use ASP only for inference or assignment. The formal construction itself is sound and the circularity burden is low, so the paper remains a useful platform contribution; the broader “robust end-to-end training for dynamic domains” claim is simply not yet evidenced. Hence the verdict stays CONDITIONAL and confidence remains moderate. The concrete test above would settle the issue with a single, well-defined experiment.","tokens_in":14359,"tokens_out":639,"duration_ms":7614,"concrete_test":"Re-implement the CLEVR attribute heads so that they are trained end-to-end with the free-energy objective of Sec. 3.2 (positive free energy under the functional-program constraints, negative free energy over unconstrained stable models). Measure (i) whether attribute and VQA accuracies remain within 5 % of the reported supervised figures and (ii) wall-clock time and memory for the free-energy solves on scenes with ≥ 10 objects. If either accuracy collapses or the free-energy step exceeds a few seconds per image, the end-to-end claim is unsupported for the combinatorial regimes the paper advertises.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that the contrastive free-energy objective L = F+ − F− (Sec. 3.2) yields informative, stable gradients for the neural energy factors fθ, while Clingo’s exact MAP optimisation over the grounded program remains tractable for the combinatorial spaces of dynamic domains. In the only fully end-to-end experiment (MNIST Addition, Sec. 4.1) the hypothesis space is tiny (10^{2} models). In CLEVR (Sec. 4.2) the attribute EBMs are trained with a fully supervised hinge loss, not the free-energy objective; the ASP program is used only at inference. In MOT (Sec. 4.3) the EBM is trained offline with pairwise contrastive loss and then frozen; ASP merely performs assignment. Consequently the paper never demonstrates that gradients obtained by solving the ASP program once with and once without supervision constraints actually train the neural substrate on a non-trivial combinatorial space. Tables 1–2 report only inference-time statistics, not training-time gradient quality or scaling of the free-energy computation. If those gradients are uninformative or the free-energy solve becomes intractable, the “joint optimisation \\ldots fully incorporating \\ldots non-monotonic inference” claim collapses to ordinary hybrid post-processing.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes ASPEn, a neurosymbolic framework that couples Answer Set Programming with Energy-Based Models so that stable models become preferred configurations in a continuous energy landscape. Energised atoms (energise(...)) inject neural energy factors into an ASP program; inference is exact MAP via Clingo’s #minimize over the sum of active energies Eθ(M|x); learning uses a contrastive free-energy objective L = F+ − F− obtained by solving once with and once without supervision constraints. Three case studies are presented: weakly supervised MNIST digit addition (end-to-end), CLEVR visual question answering (attribute EBMs trained with supervised hinge loss, ASP used at inference), and MOT multi-object tracking (pairwise EBM trained offline, ASP used for constrained assignment). Runtime tables for MOT show that grounding and solving remain practical for hundreds of energy atoms per frame.","tokens_in":14726,"tokens_out":1083,"duration_ms":9649,"significance":"If the claimed joint optimisation truly works, ASPEn would give the ASP community a clean, modular route to end-to-end training that preserves non-monotonic semantics, defaults and abductive model selection—capabilities that probabilistic logic and neural-predicate approaches typically sacrifice. The practical Clingo/PyTorch implementation and the explicit energy-based semantics over stable models are genuine engineering contributions. The MNIST addition experiment already shows that weak supervision via ASP constraints can train a digit classifier to high accuracy. However, the stronger claim of “robust, end-to-end training for dynamic domains” is only partially substantiated by the current experiments, so the significance remains conditional on a clearer demonstration that ASP-derived free energies produce usable gradients on non-trivial combinatorial spaces.","major_comments":[{"comment":"Sec. 3.2 defines the contrastive free-energy objective L = F+ − F− as the mechanism that lets non-monotonic ASP semantics participate in gradient-based training. Only the MNIST Addition experiment (Sec. 4.1) actually uses this objective; the hypothesis space is tiny (100 models). In CLEVR (Sec. 4.2) attribute EBMs are trained with a fully supervised margin hinge loss; ASP is used only for inference. In MOT (Sec. 4.3) the EBM is trained offline with pairwise contrastive loss and then frozen. Consequently the central claim of joint optimisation that “fully incorporates \to non-monotonic inference” is not demonstrated on any non-trivial combinatorial space. Either an end-to-end free-energy training experiment on CLEVR or MOT, or an explicit ablation showing that the free-energy gradients remain informative when the number of stable models grows, is required.","section":null},{"comment":"CLEVR VQA accuracy is reported as 62.85 % (Fig. 3b) with no comparison to the neuro-symbolic ASP baselines already cited (Eiter et al. 2022, Yi et al. 2018) and no ablations that isolate the contribution of energy-based ranking versus pure ASP or pure neural attribute prediction. Without these numbers the claim that the energy formulation “successfully bridges visual perception and logical reasoning” cannot be evaluated.","section":null},{"comment":"MOT results (HOTA 43.26, MOTA 36.37, IDF1 49.89) are described as “comparable with trackers of similar character,” yet no concrete baseline numbers or public leaderboard references are supplied. Tables 1–2 report only inference-time statistics; they do not address training-time cost or gradient quality of the free-energy objective. A short comparison table against at least one standard tracking-by-detection method is needed to support the dynamic-domain claim.","section":null}],"minor_comments":[{"comment":"Fig. 1 caption and surrounding text use both “energised” and “energized”; pick one spelling consistently.","section":null},{"comment":"Sec. 4.2: the functional-program-to-ASP translation is only illustrated by one example; a short formal sketch or reference to the translation algorithm would improve reproducibility.","section":null},{"comment":"Table 2 reports “Number of Energy Atoms per Frame 339 \to 1058”; it would be useful to state whether these counts include only match atoms or also the auxiliary energy atoms generated by grounding.","section":null},{"comment":"The IoU threshold used as a hard constraint in MOT is listed as a free parameter but never given a concrete value or sensitivity analysis.","section":null}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid engineering contribution that will interest the ASP and neurosymbolic communities, but the gap between the abstract’s “end-to-end \to dynamic domains” rhetoric and the actual experimental design (only MNIST is truly end-to-end) is large enough that a major revision is the appropriate gate. If the authors can add even one non-trivial free-energy training run or a clear ablation, the paper becomes a strong accept for a KR/AI journal."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real move here is treating ASP stable models as configurations in an energy landscape: energise atoms, sum their EBM scores into E_theta(M|x), then do exact MAP via Clingo #minimize and train with the free-energy contrast L = F+ - F- (supervision constraints on vs off). That is a legitimate generalisation of the NeurASP / ASPMT / probabilistic-ASP line, and they ship a working Clingo+PyTorch loop plus three demos.\n\nWhat they do well: the formal definitions are crisp, the MNIST addition experiment really is end-to-end under weak sum supervision and hits high digit/sum accuracy, and the MOT tables show that grounding and solving with hundreds of energy atoms stays cheap enough for real video. CLEVR and MOT at least prove the modular pipeline can be wired into existing detectors and trackers.\n\nThe soft spot is exactly the one the stress-test flags, and it is not minor. Only MNIST uses the ASP-derived free-energy gradients to train the network; CLEVR trains attribute heads with ordinary supervised hinge loss and uses ASP only at inference, while MOT freezes a pairwise contrastive EBM and lets ASP do assignment. So the headline claim of “joint optimisation \to continuous latent space through explicit ASP-based declarative semantics” is demonstrated only on a 100-model space. Accuracies are modest (CLEVR 62.85 %, MOT HOTA 43), baselines and ablations are missing, and no code is released. That does not kill the framework, but it does mean the “robust \to dynamic domains” rhetoric runs ahead of the evidence.\n\nThis is for people already building ASP-centric neurosymbolic systems who want a reusable energy substrate. It is not a breakthrough result, but it is a usable platform paper with honest formal scaffolding. I would send it to referees; the core construction is sound enough to improve under review. Worth a look if you are in the area; not urgent otherwise.","headline":"Clean ASP-EBM integration that works as a platform, but the full end-to-end free-energy claim is only shown on the toy case.","tokens_in":15336,"tokens_out":505,"would_cite":false,"duration_ms":15080,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"ASP stable models become preferred energy minima so non-monotonic logic can train neural perception end-to-end.","keywords":["answer set programming","energy-based models","neurosymbolic AI","stable models","end-to-end learning","visual question answering","multi-object tracking","non-monotonic reasoning"],"falsifier":"Train the MNIST-addition or MOT pipeline with the published architecture; if digit-sum accuracy or HOTA/IDF1 collapses relative to the reported figures once the energy terms are replaced by uniform costs or the positive/negative free-energy contrast is ablated, the claim that ASP semantics meaningfully shape learning fails.","tokens_in":15240,"feed_emoji":"⚡","tokens_out":863,"duration_ms":10210,"temperature":0.7,"pith_summary":"The paper presents ASPEn, a method that fuses answer set programming with energy-based models so that high-level declarative knowledge, constraints, and non-monotonic inference fully participate in continuous optimisation and gradient-based learning. Instead of treating logic as a post-hoc filter on neural outputs, the method interprets each stable model as a configuration whose energy is the sum of neural energy factors attached to its atoms; inference finds the lowest-energy stable model, and learning lowers the energy of models that satisfy supervision relative to those that do not. The authors argue this removes the conceptual separation that has limited earlier neurosymbolic ASP systems in dynamic, partially observed settings. They supply a working implementation that injects neural energies into Clingo’s optimisation engine and demonstrate the pipeline on digit addition under sum-only supervision, compositional visual question answering on CLEVR, and multi-object tracking on MOT.","feed_headline":"Logic programs train neural nets by scoring stable models as energies","feed_subtitle":"ASP constraints and defaults become the energy landscape for end-to-end perception learning","key_machinery":"Energised atoms of the form energise(rule_name([Args])) couple ASP atoms to neural energy functions; the energy of a stable model is the sum of its active energy factors, MAP inference is exact Clingo optimisation of that sum, and contrastive free-energy (F+ − F−) supplies the learning objective.","core_discovery":"ASPEn realises end-to-end neurosymbolic reasoning and learning by treating ASP stable models as preferred configurations inside a continuous energy landscape: logical rules restrict the feasible set while learnable energy factors attached to atoms rank those configurations, so that non-monotonic declarative semantics generate structured training signals for the neural substrate.","pith_inferences":["If the free-energy contrast remains informative under heavier occlusion or longer temporal horizons, the method could support online belief revision in continuous video without re-engineering the neural backbone.","The energy-atom interface suggests a natural route for attaching other continuous modules (physics simulators, language embeddings) to the same ASP optimisation engine.","Benchmark suites that explicitly score how well learned energies respect non-monotonic defaults would become necessary once the method is adopted more widely."],"forward_implications":["Declarative background knowledge and non-monotonic defaults can be written once and then directly shape both inference and neural training without a separate loss-design step.","Perception pipelines for dynamic domains can maintain identity and spatial consistency by minimising an energy that already encodes hard logical constraints.","ASP becomes a practical substrate for end-to-end training rather than only a discrete post-processor of neural scores.","The same modular energy-atom interface can be reused across static VQA and multi-object tracking without redesigning the learning loop."],"fun_headline_variants":["ASP stable models score as energies to train nets end-to-end","ASP programs turn stable models into energy configs for neural training","Neurosymbolic learning: ASP rules shape continuous energy landscapes","Stable models rank as preferred energies for ASP-neural joint training","ASP constraints generate energy signals for end-to-end perception"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The contrastive free-energy difference obtained by solving the same program once with and once without supervision constraints must yield usable gradients, and exact MAP optimisation over the grounded program must stay tractable for the combinatorial spaces that appear in real dynamic scenes.","fun_headline_variants_meta":{"raw":{"variants":["ASP stable models score as energies to train nets end-to-end","ASP programs turn stable models into energy configs for neural training","Neurosymbolic learning: ASP rules shape continuous energy landscapes","Stable models rank as preferred energies for ASP-neural joint training","ASP constraints generate energy signals for end-to-end perception"]},"model":"grok-4.5","effort":"low","cost_usd":0.006946,"raw_usage":{"total_tokens":1639,"prompt_tokens":674,"num_sources_used":0,"completion_tokens":89,"cost_in_usd_ticks":69460000,"prompt_tokens_details":{"text_tokens":674,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":876,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":674,"tokens_out":89,"duration_ms":8023,"temperature":1.0,"reasoning_tokens":876,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T12:30:43.203213+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the MNIST-addition or MOT pipeline with the published architecture; if digit-sum accuracy or HOTA/IDF1 collapses relative to the reported figures once the energy terms are replaced by uniform costs or the positive/negative free-energy contrast is ablated, the claim that ASP semantics meaningfully shape learning fails.","supporting_citations":[],"review_version":1}