{"id":"51182b75-b4c4-4ac4-95c9-d78852c82d6e","arxiv_id":"2605.08605","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"An 800K-parameter Lattice Deduction Transformer reaches 100% accuracy on Sudoku-Extreme and Snowflake Sudoku and 99.9% on Maze-Hard by using lattice projections and abstract-interpretation supervision, while frontier LLMs score 0%.","lead":"The paper introduces the Lattice Deduction Transformer, a recurrent model that projects its latent state through a lattice to approximate sound logical deduction on constraint problems. This could matter for building smaller, more reliable AI systems that handle structured reasoning tasks where large language models currently fail.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Lattice projection's soundness approximation may not ensure generalization or true deduction beyond supervised training distribution","rationale":"The reader's weakest assumption directly identifies the same internal hinge: whether the lattice-plus-abstract-interpretation construction actually yields sound, generalizing deduction. No machine-checked proof or independent soundness argument is supplied, so the empirical numbers remain the sole support; the proposed check isolates whether the mechanism adds deductive power beyond supervised imitation.","tokens_in":1654,"tokens_out":329,"duration_ms":22946,"concrete_test":"Re-train the 800K LDT on a restricted Sudoku subset (standard 9x9 only, no extreme variants) and evaluate zero-shot on the held-out Sudoku-Extreme and Snowflake test sets; if accuracy drops below 95% or abstention rate rises sharply, the lattice projection does not independently produce generalizable deduction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that recurrent projection of the latent state through a lattice (between forward passes) yields an approximation of logically sound deduction, supervised only by a domain-agnostic abstract-interpretation over-approximation of solution sets. Abstract interpretation supplies an outer bound rather than exact logical entailment; nothing in the described on-policy training or architecture is shown to enforce that the learned dynamics stay inside the deductive closure rather than merely reproducing training trajectories. Consequently the reported 100% accuracy on Sudoku-Extreme and 99.9% on Maze-Hard could be explained by distribution matching plus a conservative abstention rule rather than by the lattice mechanism itself producing sound inference that extrapolates.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces the Lattice Deduction Transformer (LDT), a recurrent transformer that approximates logically sound deduction by projecting its latent state through a lattice between forward passes. It is trained on-policy to mirror deduction in search-based constraint solvers and supervised using a domain-agnostic abstract-interpretation approximation of solution candidate sets. Key results include an 800K-parameter LDT achieving 100% accuracy on Sudoku-Extreme and Snowflake Sudoku at reduced training cost while remaining empirically sound (correct answer or abstention), a 1.8M-parameter variant reaching 99.9% on Maze-Hard, and frontier LLMs scoring 0% on all benchmarks.","tokens_in":1784,"tokens_out":525,"duration_ms":50975,"significance":"If the lattice projection mechanism is shown to produce generalizable deductive approximations rather than distribution matching, the work would be significant for neural-symbolic reasoning: it demonstrates that small-parameter recurrent models can achieve near-perfect accuracy on hard constraint problems with built-in soundness via abstention, at far lower cost than prior recurrent reasoners or LLMs. The domain-agnostic supervision via abstract interpretation is a positive design choice that could generalize across tasks.","major_comments":[{"comment":"Abstract: the central performance claims (100% accuracy on Sudoku-Extreme, 99.9% on Maze-Hard) and the assertion of empirical soundness via abstention are presented without any information on training/validation data splits, error analysis, or the precise abstention rule; these details are load-bearing for determining whether the lattice projection enables extrapolation or whether results are explained by training-trajectory reproduction plus conservative abstention.","section":"Abstract"},{"comment":"Abstract: the claim that lattice projection 'approximates logically sound deduction' supervised only by abstract-interpretation over-approximations lacks supporting analysis showing that the learned recurrent dynamics remain inside the deductive closure rather than merely fitting observed trajectories; this is required to substantiate generalization beyond the training distribution.","section":"Abstract"}],"minor_comments":[{"comment":"Abstract: the phrase 'at a fraction of the training cost of prior small recurrent reasoners' would be strengthened by a brief quantitative comparison (e.g., epochs or FLOPs) even if full details appear later.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's fit to a theory-oriented venue in cs.LG is reasonable, but the absence of any equations or pseudocode in the provided abstract makes it difficult to assess the precise novelty of the lattice operation relative to existing recurrent or neuro-symbolic architectures."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their detailed and constructive feedback on our manuscript. We address each of the major comments below and have updated the abstract and discussion sections to provide additional clarity on the experimental setup and the deductive properties of the model.","responses":[{"response":"We concur that these details are important for interpreting the results. Although the full specifications are provided in the experimental setup (Section 4) and results analysis (Section 5), we have revised the abstract to concisely include information on the data splits, a summary of the error analysis, and the abstention rule. This revision helps readers assess the generalization capabilities of the lattice projection mechanism versus potential memorization of training trajectories.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central performance claims (100% accuracy on Sudoku-Extreme, 99.9% on Maze-Hard) and the assertion of empirical soundness via abstention are presented without any information on training/validation data splits, error analysis, or the precise abstention rule; these details are load-bearing for determining whether the lattice projection enables extrapolation or whether results are explained by training-trajectory reproduction plus conservative abstention."},{"response":"The paper includes supporting analysis for this claim through the description of the lattice projection mechanism in Section 3, which enforces properties that keep the state within the deductive closure by design, combined with the on-policy training that follows solver deduction steps. The abstract-interpretation supervision provides an over-approximation that guides the model without requiring exact solutions. Generalization is substantiated by the high accuracy on out-of-distribution benchmarks like Sudoku-Extreme and Maze-Hard, where the model either produces correct solutions or abstains. We have added a brief elaboration in the abstract and a new paragraph in the discussion section to make this analysis more prominent.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the claim that lattice projection 'approximates logically sound deduction' supervised only by abstract-interpretation over-approximations lacks supporting analysis showing that the learned recurrent dynamics remain inside the deductive closure rather than merely fitting observed trajectories; this is required to substantiate generalization beyond the training distribution."}],"tokens_in":1344,"tokens_out":472,"duration_ms":59101,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main point is that an 800k-parameter recurrent transformer projects its latent state through a lattice between steps, trains on-policy like a constraint solver, and uses abstract-interpretation supervision to reach 100% accuracy on Sudoku-Extreme and Snowflake Sudoku while either getting the answer right or abstaining. A slightly larger version does 99.9% on Maze-Hard, and frontier LLMs score zero on the same tests.","headline":"Lattice Deduction Transformers embed lattice projections into a small recurrent model to hit near-perfect scores on hard Sudoku and maze tasks while abstaining on uncertainty, but the evidence for true deductive generalization is still thin.","tokens_in":2272,"tokens_out":172,"would_cite":false,"duration_ms":59831,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[{"relation":"unclear","rs_module":"IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean","rs_theorem":"absolute_floor_iff_bare_distinguishability","paper_passage":"We introduce the Lattice Deduction Transformer (LDT), a recurrent transformer that approximates logically sound deduction by projecting its latent state through a lattice between forward passes."},{"relation":"unclear","rs_module":"IndisputableMonolith/Foundation/AlexanderDuality.lean","rs_theorem":"alexander_duality_circle_linking","paper_passage":"Abstract interpretation... models information states as elements of a lattice with a top element ⊤... Deduction is a process that descends from ⊤ toward more informative states"}],"headline":"Lattice projection for ML deduction is orthogonal to RS forcing from distinction to J-cost and constants","alignment":"orthogonal","rationale":"The paper's core machinery (recurrent transformer with lattice projections via Galois connections and abstract interpretation for sound deduction on Sudoku/Maze) operates in neuro-symbolic AI and constraint solving. RS derives J(x)=½(x+x⁻¹)−1, φ, 8-tick periodicity, D=3 and constants parameter-free from a single distinction (reality_from_one_distinction). No shared structure (no J-cost, no φ-ladder, no recognition cost forcing).","tokens_in":54798,"confidence":"high","tokens_out":320,"duration_ms":15907,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A recurrent transformer achieves perfect accuracy on extreme Sudoku by projecting its latent state through a lattice to approximate sound deduction.","keywords":["Lattice Deduction Transformer","recurrent transformer","Sudoku","constraint satisfaction","abstract interpretation","logical deduction","small language models","neuro-symbolic reasoning"],"falsifier":"A held-out test set of Sudoku puzzles whose solution structures or constraint densities differ from the training distribution, where accuracy falls below 100% or the model produces incorrect answers without abstaining.","tokens_in":2546,"feed_emoji":"🧩","tokens_out":545,"duration_ms":38386,"temperature":0.7,"pith_summary":"The Lattice Deduction Transformer trains a small recurrent model to perform logical deduction on constraint problems by inserting a lattice projection step between its forward passes. This projection, guided by abstract interpretation of possible solutions, lets the network mirror the behavior of search-based solvers during on-policy training. The resulting 800K-parameter model reaches 100 percent accuracy on Sudoku-Extreme and Snowflake Sudoku while either returning a correct answer or abstaining, and a slightly larger variant reaches 99.9 percent on Maze-Hard. Frontier large language models score zero on the same benchmarks, and the approach uses far less training compute than earlier small recurrent reasoners. A sympathetic reader would care because the method suggests that compact neural networks can embed reliable, generalizable deduction for tasks that normally demand either exhaustive search or massive scale.","feed_headline":"Lattice transformer hits 100% on extreme Sudoku with 800K params","feed_subtitle":"The model either solves correctly or abstains on hard constraint puzzles where frontier LLMs score zero, using lattice projections for sound","key_machinery":"Projecting the latent state through a lattice between forward passes, which approximates the set of solution candidates via abstract interpretation and enforces soundness during recurrent reasoning.","core_discovery":"The Lattice Deduction Transformer approximates logically sound deduction by projecting its latent state through a lattice between forward passes. Trained on-policy to mirror deduction in a search-based constraint solver and supervised by a domain-agnostic abstract-interpretation approximation of the set of solution candidates, an 800K-parameter LDT achieves 100% accuracy on Sudoku-Extreme and Snowflake Sudoku while remaining empirically sound, and a 1.8M-parameter variant reaches 99.9% accuracy on Maze-Hard. Frontier LLMs score 0% on all three benchmarks.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["800K LDT achieves 100% on Sudoku-Extreme","1.8M LDT achieves 99.9% on Maze-Hard","LDT projects latent state through lattices for deduction","Small LDT attains full accuracy as LLMs score zero"],"cache_read_input_tokens":64,"weakest_assumption_plain":"Projecting the latent state through a lattice produces an approximation of logically sound deduction that generalizes beyond the training distribution when supervised by abstract interpretation.","fun_headline_variants_meta":{"raw":{"variants":["800K LDT achieves 100% on Sudoku-Extreme","1.8M LDT achieves 99.9% on Maze-Hard","LDT projects latent state through lattices for deduction","Small LDT attains full accuracy as LLMs score zero"]},"model":"grok-4.3","cost_usd":0.014837,"raw_usage":{"total_tokens":6266,"prompt_tokens":609,"num_sources_used":0,"completion_tokens":70,"cost_in_usd_ticks":148365500,"prompt_tokens_details":{"text_tokens":609,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":5587,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":609,"tokens_out":70,"duration_ms":67320,"temperature":1.0,"reasoning_tokens":5587,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-12T01:23:37.579841+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A held-out test set of Sudoku puzzles whose solution structures or constraint densities differ from the training distribution, where accuracy falls below 100% or the model produces incorrect answers without abstaining.","supporting_citations":[],"review_version":1}