{"id":"1c5019a8-8439-4b73-9b7f-af743c4fcde5","arxiv_id":"2608.12781","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Fast non-thinking inference in hybrid-thinking MLLMs produces far more response-pattern failures (CoT leakage, repetition, contradiction, performative reasoning) than thinking inference, and PatternRL reduces this gap by about 13-14 points with a small accuracy cost.","lead":"This paper introduces PatternEval, a diagnostic benchmark of 2,415 multimodal prompts that measures four types of bad response behavior in hybrid-thinking AI models, and reports that the fast 'non-thinking' mode triggers these failures far more often than the slow 'thinking' mode across 25 model configurations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Shared LLM judge for PatternEval and PatternRM training may inflate both the measured mode gap and the PatternRL improvement; human labels on PatternEval itself are needed.","rationale":"The reader's weakest assumption correctly identifies the load-bearing dependence on Seed-2.0-Pro as the pattern judge. I agree that this is the most consequential risk: the benchmark's headline gap and the PatternRL improvement are both measured by this judge, and the same judge family contributes to PatternRM's training labels, so the evaluation and optimization signals are not cleanly separated. The moderate F1 on contradiction and performative reasoning is less decisive by itself because those labels are rare, but the risk of systematic bias—especially length- or style-based bias—is amplified by the strong length-Trigger correlations and by the absence of any human-annotated PatternEval sample. I do not see an internal inconsistency or a reason to reject: the paper is transparent about the failure-enriched design and reports calibration honestly. However, the conditional verdict is appropriate because the central quantitative claims would be substantially weakened if human annotation on PatternEval showed a much smaller PatternRL effect than the judge reports. A focused human-annotation study on the trained models would settle whether the observed reduction reflects genuine response-pattern improvement or alignment to judge-specific cues.","tokens_in":30440,"tokens_out":6469,"duration_ms":71089,"concrete_test":"Human-annotate a stratified random sample of 400 non-thinking responses from Table 5—100 per backbone and condition for BaseRL and PatternRL on Qwen3-VL-4B/8B—using the paper's four-label rubric and CoT-priority rule, with two annotators and disagreement adjudication. Compare human Trigger rates against Seed-2.0-Pro's labels on the same responses. If the human-evaluated BaseRL-to-PatternRL Trigger reduction is less than 5 percentage points while Seed-2.0-Pro reports 13.08/14.35, the shared-judge family is inflating the headline improvement. As a cheaper auxiliary check, re-score the same 400 responses with GPT-5.5 image+text, which was not used to build PatternRM; an independent judge that does not reproduce the gap would further support the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"All headline Trigger numbers—the universal positive gap in Table 3 and the 13.08/14.35 percentage-point PatternRL reductions in Table 5—are produced by Seed-2.0-Pro with image access. The only human validation is a separate 2,500-response calibration set, on which this judge reaches macro-F1 75.3% but only 62.5% on logical contradiction and 60.0% on performative reasoning (Table 2). The aggregate Trigger is dominated by CoT leakage and repetition, where the judge is more accurate but may rely on surface cues such as 'let me' phrasing or long process narration, which are strongly correlated with inference mode; Section 4.3 reports response-length correlations of r=0.64 (non-thinking) and r=0.84 (thinking) with Trigger. More importantly, PatternRM's supervision labels come from unanimous agreement among Kimi-K2.6, Seed-2.0-Pro, and Qwen3.5-397B (Section 5.1), so the training reward and the evaluation metric are not independent. PatternRL could reduce judged Trigger by conforming to Seed-2.0-Pro's stylistic preferences without improving user-facing response quality, and the paper reports no human labels on PatternEval itself and only a single training run (Appendix D.2). The central claim therefore requires that the judge's labels are valid on the benchmark and not exploitable by the training objective.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces PatternEval, a failure-enriched multimodal benchmark of 2,415 prompts that evaluates four response-pattern failures (chain-of-thought leakage, response repetition, logical contradiction, and performative reasoning) under matched thinking and non-thinking inference. Using Seed-2.0-Pro as the pattern judge, the authors report that all 25 evaluated model configurations exhibit higher non-thinking Trigger rates (Section 4.2, Table 3). They then train PatternRM, a text-only reward model distilled from unanimous labels of three judges including Seed-2.0-Pro (Section 5.1), and PatternRL, which adds a pattern penalty to the GRPO verifier reward during RL (Eq. (7)). On Qwen3-VL-4B and Qwen3-VL-8B, PatternRL reduces non-thinking Trigger relative to correctness-only BaseRL by 13.08 and 14.35 percentage points while changing PatternEval accuracy by less than one percentage point (Section 5.3, Table 5), though it produces small accuracy declines on ten external reasoning and document-understanding benchmarks (Table 6).","tokens_in":30746,"tokens_out":9220,"duration_ms":90751,"significance":"The paper's strengths are its matched-mode paired experimental design, the explicit and detailed four-label taxonomy with a full judge prompt in Appendix E, the use of a separate calibration set for judge selection, and the transparent acknowledgment of limitations such as single-run training and the failure-enriched nature of the benchmark. If the judge-validity concern is resolved, PatternEval would be a useful diagnostic resource for hybrid-thinking MLLMs, and PatternRL would be a practical demonstration that response-pattern penalties can be injected into GRPO training without collapsing accuracy. The reported 13-14 percentage-point Trigger reductions are substantial, and the observation that correctness-only RL aggravates pattern failures is interesting. However, the central numbers are all produced by the same judge family that shapes the training reward, so the current evidence does not yet establish the headline claims to the standard expected for a journal paper.","major_comments":[{"comment":"The evaluation and training signals are not independent. PatternEval's Trigger is computed with Seed-2.0-Pro as Jpat (Eq. (3), Section 3.2), and PatternRM is trained on 57,578 instances whose labels are kept only when Kimi-K2.6, Seed-2.0-Pro, and Qwen3.5-397B agree unanimously (Section 5.1). PatternRL then optimizes PatternRM's penalties (Eq. (7)) and is evaluated on Trigger measured by the same Seed-2.0-Pro (Table 5). Because Seed-2.0-Pro is one of the three annotators of PatternRM's supervision, the 13.08 and 14.35 percentage-point reductions may in part reflect the policy conforming to Seed-2.0-Pro's stylistic preferences rather than a genuine improvement in user-visible response quality. This concern is compounded by the calibration results in Table 2, where this judge has F1 of only 62.5% for logical contradiction and 60.0% for performative reasoning. The paper needs human labels on a sample of PatternEval responses themselves (not only the separate 2,500-response calibration set) or an independent judge to validate both the universal gap in Table 3 and the PatternRL gain in Table 5.","section":"§3.2, §5.1, Eq. (7), Table 2"},{"comment":"The length confound is not resolved. Section 4.3 reports strong correlations between average response length and Trigger (r=0.64 non-thinking, r=0.84 thinking), and Figure 6 shows a monotone increase in Trigger across length sextiles. The operational judge prompt (Appendix E) instructs the judge not to use length, but the cited literature on LLM judges (refs. [4,21,30,39,52]) shows that such instructions are often insufficient. Because non-thinking responses are typically longer than thinking responses, a length-sensitive judge would inflate both the universal mode gap in Table 3 and the PatternRL improvement in Table 5. The paper should provide a length-controlled version of the headline results (for example, Trigger within matched-length bins or a human-annotated subset stratified by length) before the gap is attributed to inference mode rather than response form.","section":"§4.3, §5.3"},{"comment":"The reward design makes the pattern penalty conditional on correctness, leaving the worst failure regime unoptimized. In Eq. (8), incorrect responses always receive reward 0 regardless of the pattern signal, because s_ver=0 and s_pat<=0 are clipped to zero. Consequently, PatternRL cannot reduce pattern failures in wrong answers, even though Section 4.3 and Figure 6 identify long, incorrect non-thinking responses as the highest-Trigger regime. Since the paper explicitly frames quality as orthogonal to correctness, the authors should report Trigger split by correctness and ideally run an ablation that also penalizes pattern failures on incorrect responses, or at least state this limitation prominently.","section":"§5.2, Eq. (8)"},{"comment":"The central training result rests on a single run. Appendix D.2 states that the reported configuration corresponds to a single run and that multi-seed training manifests are unavailable. Table 5 therefore reports no variance, confidence intervals, or significance tests for the 13.08/14.35 percentage-point Trigger reductions or for the less-than-one-percentage-point PatternEval accuracy change. Given the known variance of GRPO-style training, the 4B versus 8B comparisons and the claim that PatternRL preserves task accuracy need at least a small number of seeds or bootstrap intervals to be load-bearing.","section":"Appendix D.2, Table 5"}],"minor_comments":[{"comment":"The sentence 'aggregate accuracy changes by less than one percentage point' in the introduction should be explicitly scoped to PatternEval accuracy in Table 5; on the external benchmarks in Table 6 the 4B model loses 1.87 points on average, which is not a negligible change.","section":"§1, §5.3"},{"comment":"Table 3 is dense and reports point estimates without any measure of uncertainty; adding standard errors or at least noting the paired nature of the comparisons would help readers assess the robustness of the 'every evaluated pair has a positive gap' claim.","section":"Table 3"},{"comment":"Figure 1(a) shows gap values but does not clearly label which bars correspond to thinking and non-thinking Trigger rates; a legend or explicit axis labels would improve readability.","section":"Figure 1"},{"comment":"The table header states 'thinking-mode results after reinforcement learning are not reported'; please state explicitly whether PatternRL updates only non-thinking rollouts, and if so, report the post-RL thinking-mode Trigger to show the full cross-mode gap.","section":"Table 5"},{"comment":"The paper notes that the complete 8B training configuration is unavailable; this should be provided if possible, since reproducibility of the central RL result depends on it.","section":"Appendix D.3"},{"comment":"The paper is transparent that PatternEval is failure-enriched, but Takeaway 1 in Section 4.2 presents the universal positive gap as a general cross-family finding; consider softening this to 'on this failure-enriched diagnostic benchmark' to avoid overgeneralization.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The shared-judge issue is the main risk: the evaluation judge is also one of the three annotators for the training reward's supervision, and no human labels exist on PatternEval itself. I would ask the authors to provide human-annotated PatternEval samples or an independent judge, plus length-controlled analyses, before publication. The paper is within scope for an evaluation-and-alignment venue, but the current evidence does not yet separate genuine response-quality improvement from reward hacking toward the judge's stylistic preferences."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth taking seriously. It defines a plausible and under-studied problem—user-visible response-pattern failures in hybrid-thinking MLLMs—and builds a benchmark plus a training method around it. The matched-mode evaluation across 25 model configurations is a lot of work, and the failure taxonomy is well specified; the appendix prompt is detailed enough that the label definitions are auditable, including the CoT-priority rule and the distinction between performative reasoning and genuine but wrong reasoning. Giving the judge a human-labeled calibration set and reporting per-label F1 is also more transparent than what most benchmark papers do.\n\nThe central finding—that non-thinking responses carry more pattern failures than thinking responses—is probably real, at least for the two dominant labels, CoT leakage and repetition. Those labels are measured with high F1, and the same positive gap appears across every evaluated model. The weaker labels, contradiction and performative reasoning, are where judge F1 drops to 62.5 and 60.0, and the paper itself notes these labels are harder. So the aggregate gap direction is likely robust, but the precise magnitudes are not.\n\nThe bigger soft spot is the shared judge between evaluation and training. PatternEval labels come from Seed-2.0-Pro, and Seed-2.0-Pro is one of three annotators that generate PatternRM's supervision corpus. That means PatternRL is optimizing toward a distilled version of the same judge that defines the headline Trigger metric. If the judge has stylistic preferences—and the strong length correlations (r=0.64 non-thinking, r=0.84 thinking) suggest it might—then the reported 13-point and 14-point Trigger reductions could partly be reward hacking, not genuine improvement in user-facing quality. The calibration set being separate from PatternEval is good practice, but it does not rule this out, because the judge's accuracy on the benchmark prompts themselves is never directly measured with human labels.\n\nOther issues are more minor: no multi-seed runs or error bars on the RL results, and the benchmark and training data are not released. The paper is honest about these limitations in the appendix, which I appreciate.\n\nThis is a paper for MLLM alignment and RL practitioners. It deserves serious peer review, not desk rejection, but the reviewers should ask for human labels on a sample of PatternEval itself, evaluation with at least one independent judge, and public artifacts. I would not cite the headline numbers as established until those pieces exist, but I would cite it as the first to frame response-pattern alignment across hybrid-thinking modes.","headline":"Useful benchmark and training recipe for response-pattern alignment, but the headline numbers rest on a judge that also labels the training data; referees should push for independent evaluation before trusting effect sizes.","tokens_in":31331,"tokens_out":1990,"would_cite":true,"duration_ms":24698,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Hybrid-thinking multimodal models behave worse in fast non-thinking mode than in thinking mode on all 25 configurations tested — and a pattern-aware reinforcement-learning penalty removes most of that gap without hurting accuracy.","keywords":["hybrid-thinking multimodal LLMs","response-pattern alignment","PatternEval","chain-of-thought leakage","response repetition","performative reasoning","reward model","reinforcement learning"],"falsifier":"Have human raters apply the paper's own four-label rubric to a random sample of perhaps 300 of the actual 2,415 PatternEval prompts and recompute each model's thinking and non-thinking Trigger rates from those labels. If the human-measured gap is near zero, or if most of the gap disappears when responses are matched for length before judging, the central diagnosis collapses. A second check targets the remedy: if PatternRL's Trigger reduction is fully explained by shorter outputs rather than by fewer repetitions, contradictions, and leaked traces at matched length, the pattern-penalty mechanism claim fails.","tokens_in":30237,"feed_emoji":"🤖","tokens_out":16902,"duration_ms":135827,"temperature":0.7,"pith_summary":"The paper's aim is to establish response-pattern alignment as a requirement that sits alongside answer correctness for hybrid-thinking multimodal models — models that can switch between a slow, deliberative thinking mode and a fast, low-latency non-thinking mode. Using PatternEval, a deliberately failure-enriched benchmark of 2,415 multimodal prompts, it claims that every one of the 25 model configurations it tests shows the same systematic gap: non-thinking answers carry far more user-visible failures, above all leaked reasoning traces, but also needless repetition, logical contradictions, and performative reasoning that stages analysis without real evidence. The paper then claims this gap is fixable at the training stage: PatternRM, a reward model trained on 57,578 judge-consensus examples to spot the four failure types, feeds small pattern-specific penalties into reinforcement learning, and the resulting PatternRL cuts non-thinking failure triggers by 13.08 and 14.35 percentage points on two Qwen3-VL backbones while PatternEval accuracy moves less than one point. The broader point is that correctness metrics alone cannot characterize delivered response quality, and that user-visible behavior is a separate, measurable, optimizable axis of model quality.","feed_headline":"Fast AI mode leaks reasoning in all 25 hybrid models","feed_subtitle":"A new 2,415-prompt benchmark exposes the gap; pattern-aware RL cuts fast-mode failures by 13+ points at flat accuracy.","key_machinery":"The load-bearing object is the four-category response-pattern taxonomy — chain-of-thought leakage, response repetition, logical contradiction, and performative reasoning — with an operational judge (Seed-2.0-Pro with image access) that assigns all four binary labels to each response. The second mechanism is the PatternRL reward design, which turns those labels into a bounded penalty on top of the correctness signal: the final reward is $r = \\mathrm{clip}(s_{\\mathrm{ver}} + s_{\\mathrm{pat}}, 0, 1)$, where the pattern penalty is clamped at 0.1 in magnitude, leakage and repetition each carry weight 0.05, contradiction and performative reasoning each carry 0.02, and the judge is invoked per rollout with probability 0.6. The design keeps a hard correctness gate — wrong answers always receive zero — while slightly down-scoring correct answers that carry bad patterns, so the pattern term nudges behavior without overriding the primary task reward.","core_discovery":"The central claim is stated directly: response-pattern alignment is a requirement complementary to correctness in hybrid-thinking MLLMs, because switching inference modes should not degrade the user-visible quality of the final answer. On 2,415 PatternEval prompts, evaluated under matched thinking and non-thinking conditions, every one of 25 model configurations shows a positive response-pattern gap — non-thinking Trigger is consistently higher, exceeding 20 percentage points for 17 of 25 pairs, with chain-of-thought leakage the dominant failure and failure rates uneven across task families and categories. Making the diagnosis actionable, the paper trains PatternRM on 57,578 examples that all three judge models labeled identically, and folds its four category weights into the GRPO reward as $s_{\\mathrm{pat}} = -z \\min(0.1, \\sum_p w_p \\hat{b}_p)$, with leakage and repetition weighted 0.05 and contradiction and performative reasoning weighted 0.02. On Qwen3-VL-4B and Qwen3-VL-8B, the resulting PatternRL lowers non-thinking Trigger on PatternEval from 51.55% to 38.47% and from 44.33% to 29.98% relative to correctness-only BaseRL — reductions of 13.08 and 14.35 points at under one point of PatternEval accuracy change — while a ten-benchmark accuracy suite still shows a modest trade-off that is larger for the 4B model.","pith_inferences":["A length-controlled re-scoring of PatternEval — matching thinking and non-thinking responses for token count before the judge runs — would show how much of the claimed gap survives once the paper's own strong length correlations (r=0.64 and r=0.84) are neutralized; the paper reports the correlations but does not perform this control.","The same judge-and-penalty loop transfers beyond hybrid modes to any deployment with two output routes (interactive versus batch, cheap versus expensive decoding), where pattern drift between routes could be monitored and penalized the same way.","Because of the CoT-priority attribution rule, the headline per-label rates understate contradiction and performative reasoning; re-annotating leaked deliberation independently would shift some fraction of the CoT rate into the other two categories.","PatternRM is text-only while the operational judge sees images; feeding image-grounded supervision into the reward model could recover some of the contradiction and performative-reasoning judgments that the current training signal underweights."],"forward_implications":["Accuracy-only evaluation understates a systematic quality gap: a hybrid model can score well on task correctness while its fast mode returns verbose, self-contradicting, or reasoning-leaking answers.","Correctness-only reinforcement learning can actively worsen response patterns, since BaseRL raises non-thinking Trigger from 44.22% to 51.55% on the 4B model and from 36.77% to 44.33% on the 8B model.","Pattern-aware penalties can be added to broader task training, increasing the proportion of usable responses at marginal accuracy cost rather than being limited to a dedicated alignment stage.","The accuracy cost of pattern constraints scales with model capacity, so the penalty weights or reward mixture need to be tuned per model size and task difficulty.","Auxiliary RL penalties alone do not eliminate the failures, which points to earlier training stages as the place where pattern-biased behavior is baked in."],"supporting_citations":[{"why":"Documents the reward-hacking failure in which a policy satisfies a non-thinking reward while visibly deliberating, the exact problem PatternRL is built to counter.","marker":"[7]"},{"why":"Defines the hybrid thinking/non-thinking interface under study and supplies the backbone model family used for the training experiments.","marker":"[10]"},{"why":"Provides the GRPO reinforcement-learning algorithm used for both the correctness-only baseline and PatternRL.","marker":"[34]"},{"why":"Supplies most of the multimodal reinforcement-learning training prompts used for both the baseline and PatternRL runs.","marker":"[49]"},{"why":"Establishes the deliberative-reasoning paradigm that the thinking mode instantiates and that the benchmark contrasts against fast mode.","marker":"[3]"},{"why":"Contributes reasoning-heavy mathematics training data to the RL mixture that PatternRL optimizes.","marker":"[31]"},{"why":"Contributes chart-understanding training data to the RL mixture, covering the structure-sensitive document family.","marker":"[27]"},{"why":"Grounds the decision to treat chain-of-thought leakage as an observable response-style label rather than evidence of exposed internal states.","marker":"[36]"}],"fun_headline_variants":["PatternRL trims fast-mode reasoning leaks by 13+ points","New benchmark exposes response failures in hybrid MLLMs","Beyond accuracy: aligning response patterns in hybrid thinking","Fast-mode AI leaks chain-of-thought; pattern RL fixes it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every headline failure rate is produced by a single AI judge scoring the 2,415 benchmark prompts, yet humans validated that judge only on a separate 2,500-response set where it reached 62.5% F1 on logical contradiction and 60.0% on performative reasoning — and since longer responses trigger more failures, a judge that leans on length would inflate the thinking–non-thinking gap.","fun_headline_variants_meta":{"raw":{"variants":["PatternRL trims fast-mode reasoning leaks by 13+ points","New benchmark exposes response failures in hybrid MLLMs","Beyond accuracy: aligning response patterns in hybrid thinking","Fast-mode AI leaks chain-of-thought; pattern RL fixes it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000488,"raw_usage":{"total_tokens":2496,"prompt_tokens":1132,"completion_tokens":1364,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":748,"completion_tokens_details":{"reasoning_tokens":1308}},"tokens_in":748,"tokens_out":1364,"duration_ms":10441,"temperature":1.0,"reasoning_tokens":1308,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:24:05.856546+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have human raters apply the paper's own four-label rubric to a random sample of perhaps 300 of the actual 2,415 PatternEval prompts and recompute each model's thinking and non-thinking Trigger rates from those labels. If the human-measured gap is near zero, or if most of the gap disappears when responses are matched for length before judging, the central diagnosis collapses. A second check targets the remedy: if PatternRL's Trigger reduction is fully explained by shorter outputs rather than by fewer repetitions, contradictions, and leaked traces at matched length, the pattern-penalty mechanism claim fails.","supporting_citations":[{"cited_title":"Thinking-Based Non-Thinking: Solving the Reward Hacking Problem in Training Hybrid Reasoning Models via Reinforcement Learning","cited_arxiv_id":"2601.04805","evidence_quote":"Documents the reward-hacking failure in which a policy satisfies a non-thinking reward while visibly deliberating, the exact problem PatternRL is built to counter."},{"cited_title":"Openmmreasoner: Pushing the frontiers in multimodal reasoning with an open and general recipe","cited_arxiv_id":null,"evidence_quote":"Supplies most of the multimodal reinforcement-learning training prompts used for both the baseline and PatternRL runs."}],"review_version":1}