{"id":"43984e70-f92a-4c05-8b4e-cd579a5ab197","arxiv_id":"2607.26946","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A disentangled Belief head with uncertainty gating is claimed to replace MCTS correction and enable professional-level search-free Go on consumer GPUs, but the reported experiments do not demonstrate that claim.","lead":"The paper proposes a Go agent that splits move selection from a separate Belief head meant to catch overconfident errors, aiming to play well without heavy Monte Carlo tree search. If it worked as claimed, strong Go play would become practical on ordinary GPUs instead of large search clusters.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Central claim of search-free professional strength is unsupported by any game-outcome evidence","rationale":"The Reader correctly isolates the decisive gap: the architecture is motivated as an MCTS replacement, yet evaluation never leaves the supervised-proxy regime. I find the same soft spot and no stronger independent flaw that would alter the verdict. The correlation plots and visualizations are consistent with improved feature learning but do not constitute evidence of the claimed game-strength or hardware outcome. Because the missing experiment is both necessary and straightforward, the REJECT judgment stands without modification.","tokens_in":8384,"tokens_out":446,"duration_ms":10372,"concrete_test":"Run a fixed 400-game search-free match (temperature 0, no MCTS) of the final BGDM checkpoint versus an identically trained Policy-only Transformer baseline and versus a published search-free KataGo/MiniZero net of comparable parameter count; report win rate, mean score difference, and blunder rate (moves that drop estimated win-rate >15% per an independent strong engine). If BGDM win rate is not statistically above both baselines, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper’s strongest claim (Abstract, I, VI) is that a disentangled Belief head plus uncertainty gating “significantly improves search-free win rates and reduces hallucination, enabling professional-level play on limited hardware.” For this to hold, the Belief head must function as an effective internal substitute for MCTS verification at inference. Sections IV–V never test that premise: they report only supervised pre-training loss curves, Acc@1/Acc@5 correlation with Belief error, qualitative heatmaps, and an unsupported “30% blunder reduction.” No win rates, Elo, head-to-head matches against search-free AlphaZero/KataGo baselines, gating ablation, or RTX-2060 latency/throughput numbers appear. Gradient detachment (Eq. 1, III.C) and the gating mechanism itself are described but never evaluated under actual play. Consequently the load-bearing causal link—Belief calibration \to professional search-free strength—remains an untested assertion.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a Belief-Guided Decision Model (BGDM) for Go that disentangles a Policy head from a separate Belief head intended to model epistemic uncertainty and strategic stability. A Transformer backbone with memory, gradient detachment between Actor and Belief, and an uncertainty gating mechanism are introduced so that Belief can act as an internal critic/filter in place of deep MCTS. The stated goal is higher-quality search-free play on consumer GPUs (e.g., RTX 2060), with reduced hallucination and professional-level strength. Experiments in §IV report supervised pretraining dynamics, Belief-error vs Acc@1/Acc@5 correlation, and qualitative policy/belief heatmaps; the Discussion asserts end-game robustness and a 30% blunder reduction.","tokens_in":8604,"tokens_out":1527,"duration_ms":43588,"significance":"If the central claim held—that a calibrated, detached Belief head plus gating can substitute for MCTS verification well enough to raise search-free win rates to professional level on limited hardware—the result would matter for deployable game AI and, more broadly, for uncertainty-aware policy filtering without runtime search. The architectural separation of Policy from an independent critic and the explicit gradient-detachment protocol (III.C, Eq. 1) are reasonable design ideas. As written, however, the manuscript does not supply the game-outcome, Elo, baseline, ablation, or latency evidence needed to establish that significance; the contribution remains a proposal supported only by supervised move-matching metrics.","major_comments":[{"comment":"Abstract, §I, and §VI claim that BGDM “significantly improves search-free win rates,” “reduces hallucination,” and enables “professional-level play on limited hardware.” §IV–V report only supervised pretraining loss/accuracy curves (Fig. 5), Belief-error vs Acc@k correlation on expert/KataGo-style labels (Fig. 6), and qualitative heatmaps (Fig. 7). There are no win rates, Elo ratings, head-to-head matches against search-free Policy-Value or KataGo baselines, self-play strength curves, or any search-free playing evaluation. The load-bearing causal link—Belief calibration and gating → search-free strength—is therefore untested.","section":"Abstract; §I; §IV–VI"},{"comment":"§V states the agent reduces blunders by 30% due to more accurate perception of passing conditions. No definition of “blunder,” no counting protocol, no baseline comparison, and no table or figure support this number. An unsupported quantitative claim in the Discussion cannot stand in for missing outcome experiments.","section":"§V Discussion"},{"comment":"The gating mechanism is central to the method (Abstract; §III.A “Control Relation”: Belief filters high-risk Policy moves). The manuscript never specifies the gate (threshold, risk rule, or how belief logits modify action selection at inference), nor does it report an ablation with vs. without gating under play or even under supervised top-k. Without that, the claim that gating filters overconfident policy errors is not evidenced.","section":"§III.A; Abstract"},{"comment":"Algorithm 1 and §III.B describe only supervised pretraining of a belief_policy distribution via cross-entropy on expert actions (5 epochs, embed 128, 4 layers/heads). Gradient detachment (Eq. 1, §III.C) and the multi-head Policy/Value/Belief setup are not evaluated in RL/self-play, and no search-free inference protocol on RTX-class hardware (latency, nodes/sec, or strength at fixed budget) appears. The experimental section therefore does not instantiate the architecture’s intended use case as argued in §I against MCTS cost on consumer GPUs.","section":"§III.B–C; Algorithm 1; §IV"},{"comment":"Fig. 1 and the text disagree on heads and memory: the figure caption lists Policy, Value, and Territory Belief; the abstract promises Transformer/GRU for Ko and long-term dependency; the body emphasizes a Vision Transformer and an MLP Belief head after mean pooling, with Value largely undeveloped in the methods. This inconsistency makes the proposed model hard to reproduce and weakens the claim that Belief is a well-specified “internal simulator” distinct from standard value/ownership heads (cf. KataGo auxiliary targets discussed in §II.C).","section":"§III.A; Fig. 1; Abstract"}],"minor_comments":[{"comment":"Algorithm 1 is poorly typeset (line breaks mid-sentence, “belief_policy” vs. separate Belief head, incomplete evaluation block). It should be cleaned and aligned with the architecture description.","section":"Algorithm 1"},{"comment":"Eq. (1) writes ∇θ LTotal with LActor, LCritic, LBelief but does not define the individual losses or the stop-gradient operator formally; a clear expression for the detached update would help.","section":"§III.C Eq. (1)"},{"comment":"References [37] and [38] (tennis video annotation; Ma’qeli script cellular automata) are unrelated to Computer Go or belief/uncertainty modeling and should be removed or justified.","section":"§II; References"},{"comment":"Typos and wording: “Fixed2DPositionalEncoding,” “nosep,” “gati ng,” “Efficiency,” “halluci- nations” line breaks; “Vision Transformer backbone” vs. earlier CNN-centric Go literature could use a short related-work note on ViT-for-Go ([33] is cited but lightly used).","section":"§III.A; passim"},{"comment":"Fig. 6 reports correlation between Belief error and Acc@k but no sample size, confidence intervals, or dataset split; axes and “Mean Belief Error” units are unspecified.","section":"Fig. 6; §IV.B"},{"comment":"Input is described as 4 channels including “History,” but Ko handling and history length are not specified despite the abstract’s emphasis on Ko and memory.","section":"§III.A.1; Abstract"}],"recommendation":"reject","confidential_remarks":"I agree with the reader’s high-confidence reject: the abstract and conclusion advertise demonstrated search-free professional strength and hardware gains that the body never measures. That is not a minor revision item; it is a missing evaluation of the paper’s main claim. A resubmission would need actual playing strength (Elo/win rate), search-free baselines, gating ablations, and hardware timing—not only more supervised curves. Novelty relative to KataGo-style auxiliaries and uncertainty-aware RL should also be sharpened. Scope is fine for cs.AI/games, but the evidence gap is decisive."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple: they propose a clean architectural split—Policy head detached from a Belief head that is meant to act as an internal critic and gate—and then never test the claim that this yields better search-free win rates or professional play on limited hardware.\n\nWhat is actually new is modest but real. They isolate a Belief pathway (global pool + MLP for stability/uncertainty), block actor gradients into it (Eq. 1), and talk about using it as an inference-time filter rather than just another multi-task auxiliary. The Transformer board backbone, history channels for Ko, and supervised grounding on expert/KataGo data are sensible engineering choices, and the training curves plus Belief-error vs Acc@k plots are coherent for a pretraining stage. The qualitative heatmaps are readable. Citation coverage of AlphaZero/MuZero/KataGo and uncertainty work is adequate; they are not inventing the literature.\n\nThe soft spot is load-bearing and not minor. Abstract, Intro, and Conclusion promise significantly improved search-free win rates, reduced hallucination, and professional-level play on RTX-class hardware. Sections IV–V only show supervised loss/accuracy dynamics, a correlation between Belief error and move accuracy, heatmaps, and an unsupported “30% blunder” remark. No Elo, no head-to-head win rates, no search-free baselines, no gating ablation, no latency or throughput numbers on consumer GPUs. Gradient detachment and the gate itself are described, not evaluated under play. So the causal story—Belief calibration substitutes for MCTS verification—remains an untested assertion. Reproducibility is also thin (no code/data, free parameters around width, loss weights, and the gate rule).\n\nThis is for people already working on search-light board agents or uncertainty gating who want an architecture sketch and a pretraining recipe. It is not yet evidence that the sketch works at the strength claimed. I would still send it to referees rather than desk-reject: the problem is real, the proposal is concrete enough to evaluate, and a revision with proper matches could be useful. I would not cite it in its current form, and I would not bring it to reading group until the game-outcome numbers exist.","headline":"Clear architecture idea for search-light Go, but the professional-strength and win-rate claims are asserted without any game-outcome evidence.","tokens_in":9294,"tokens_out":554,"would_cite":false,"duration_ms":10745,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A separate Belief head can replace deep search in Go by filtering overconfident policy errors at inference time.","keywords":["Computer Go","Belief head","uncertainty gating","search-free play","Monte Carlo Tree Search","Vision Transformer","hallucination","epistemic uncertainty"],"falsifier":"Play matched search-free games on the same backbone with and without Belief gating against a fixed opponent, and check whether win rate, blunder rate, and Elo actually rise as claimed; if gating does not improve game outcomes over the raw policy, the central claim fails.","tokens_in":9170,"feed_emoji":"♟️","tokens_out":733,"duration_ms":15329,"temperature":0.7,"pith_summary":"Standard Go engines lean on Monte Carlo Tree Search to fix the neural network’s mistakes. On ordinary GPUs that search is too slow, and without it the network often plays high-confidence blunders. This paper argues that the fix is architectural, not more search: keep a Policy head for choosing moves, but add a distinct Belief head that scores how stable and trustworthy the current position is. Memory layers track long-term threats and Ko, and a gate suppresses policy moves the Belief head flags as risky. Intelligence moves from runtime tree expansion into learned parametric intuition, so search-free play stays strong on limited hardware.","feed_headline":"Belief head cuts Go hallucinations without deep search","feed_subtitle":"A detached critic gates overconfident moves so search-free play stays strong on ordinary GPUs","key_machinery":"Belief-Guided Decision Model (BGDM): a Vision-Transformer backbone with a detached Belief head (global pool plus MLP critic) whose uncertainty signal gates Policy logits, trained first on expert/KataGo data then used as an independent filter at inference.","core_discovery":"Disentangling Policy from a supervised Belief head that models epistemic uncertainty and strategic stability, then gating the policy with that belief signal, internalizes the verification role usually left to MCTS and yields stronger search-free Go play with fewer hallucinations on consumer hardware.","pith_inferences":["If belief calibration truly tracks fatal traps, the same detached critic could regularize any overconfident discrete policy, not only Go.","The missing head-to-head Elo tables against search-starved AlphaZero-style baselines are the natural next measurement the architecture invites.","Gating on belief may trade some tactical sharpness for blunder avoidance, a quality-vs-creativity tension worth measuring in endgame suites."],"forward_implications":["Search-free Go bots can keep higher move quality on single consumer GPUs by rejecting overconfident policy errors before they are played.","Gradient detachment lets the Belief head act as a fixed critic, so policy training cannot rewrite the network’s sense of positional risk.","Memory over history (Transformer/GRU) becomes part of the belief state, helping detect Ko and slow multi-move attacks without tree expansion.","The same belief-plus-gate pattern is proposed as transferable to other domains with structural uncertainty, such as real-time strategy games."],"fun_headline_variants":["Belief head gates policy errors for stronger search-free Go","Detached Belief critic cuts Go hallucinations on consumer GPUs","Uncertainty gating replaces deep MCTS in belief-guided Go","Policy-Belief split internalizes MCTS checks without tree search","Belief-guided Go plays pro-level with less hallucination, no deep search"],"cache_read_input_tokens":128,"weakest_assumption_plain":"A supervised Belief head plus gating can stand in for deep MCTS verification well enough to reach professional-level strength without search.","fun_headline_variants_meta":{"raw":{"variants":["Belief head gates policy errors for stronger search-free Go","Detached Belief critic cuts Go hallucinations on consumer GPUs","Uncertainty gating replaces deep MCTS in belief-guided Go","Policy-Belief split internalizes MCTS checks without tree search","Belief-guided Go plays pro-level with less hallucination, no deep search"]},"model":"grok-4.5","effort":"low","cost_usd":0.004458,"raw_usage":{"total_tokens":1252,"prompt_tokens":709,"num_sources_used":0,"completion_tokens":88,"cost_in_usd_ticks":44584000,"prompt_tokens_details":{"text_tokens":709,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":455,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":709,"tokens_out":88,"duration_ms":7609,"temperature":1.0,"reasoning_tokens":455,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T16:11:17.718614+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Play matched search-free games on the same backbone with and without Belief gating against a fixed opponent, and check whether win rate, blunder rate, and Elo actually rise as claimed; if gating does not improve game outcomes over the raw policy, the central claim fails.","supporting_citations":[],"review_version":1}