{"id":"4e28df76-3b99-4548-b710-9c706310f8f1","arxiv_id":"2601.07036","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A prompt combining the 'Okay' reasoning cue with the '</think>\\n\\n' no-think cue gives LLMs an intermediate reasoning budget without training, and also speeds up RL fine-tuning.","lead":"This paper shows that a hybrid LLM's 'think' vs 'no-think' behavior is switched by a few trigger tokens, chiefly a leading 'Okay' and a newline after '</think>'. It packages this into a prompt format, Mid-Think, that produces intermediate-length reasoning without retraining and also cuts RL fine-tuning time by ~15%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Trigger-token attribution lacks causal control; observed 'Okay' effect may be an attention-sink or position artifact.","rationale":"The paper's most novel and load-bearing claim is not merely that Mid-Think works, but that a small number of token-level triggers causally govern reasoning mode. The reader's weakest assumption correctly identifies the lack of causal control around 'Okay'. I agree: the evidence is correlational and confounded by position and attention-sink phenomena. However, the empirical Mid-Think prompt may still deliver its practical gains even if the mechanism is misattributed, so the appropriate verdict remains CONDITIONAL rather than REJECT. The paper would need a swap-control or intervention to support the mechanistic framing. The budget-baseline and missing-error-bars issues are secondary; they affect the Pareto and RL claims but not the core mechanistic assertion. Therefore, I recommend keeping the reader's CONDITIONAL verdict unchanged, pending the concrete substitution experiment.","tokens_in":14769,"tokens_out":4710,"duration_ms":50614,"concrete_test":"On Qwen3-8B, run MATH500 with the Mid-Think template from Table 3 but replace 'Okay' with semantically neutral tokens ('Hello', 'The', 'So') keeping all other context identical. Compare accuracy, avg length, and wait count to the original Mid-Think run. If the metrics do not change materially across substitutions, the 'Okay' trigger attribution is unsupported; the token-level mechanism should be replaced by a positional/structure explanation. Additionally, one could ablate the 'Okay' token embedding or patch attention, but the substitution experiment suffices as a first causal check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that a leading 'Okay' token is an active trigger for reasoning mode and the newline after '</think>' suppresses it. The evidence in §2.2–2.3 is correlational: average attention is higher on 'Okay' and prompting variants containing 'Okay' produce thinking-like behavior. No swap-control replaces 'Okay' with a different token in the same position, and no intervention (patching/ablating) is performed. The attention-sink literature (Xiao et al. 2023; Gu et al. 2024) shows that early tokens routinely attract high attention regardless of content, so attention concentration alone does not establish causation. Moreover, the Mid-Think format inserts '<reason>\\n' before 'Okay', so 'Okay' is not the first token; without a control that substitutes a non-reasoning token in that exact slot, the lexical trigger story is not identified. If any token in that position induces the same reasoning behavior, the mechanistic claim in §2 and the design rationale for Mid-Think (§3.1) collapse to a much weaker positional effect. The paper's own controlled experiments (Table 2) vary multiple factors simultaneously (e.g., presence of '<think>', newline patterns, and 'Okay'), so they cannot attribute the observed differences to 'Okay' specifically.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims that the Think/No-think behavior of hybrid reasoning models is governed not by high-level instructions but by a small set of token-level triggers: a leading “Okay” token induces reasoning, while the newline pattern after “</think>” suppresses it. The authors support this with attention visualizations and prompting experiments on Qwen3-8B, then propose Mid-Think, a training-free prompt format (<think>\\n\\n</think>\\n\\n<reason>\\nOkay...) that combines both triggers to obtain an intermediate reasoning budget. They evaluate Mid-Think across Qwen3-8B/14B/32B and DeepSeek-Qwen-7B on MATH500, AIME22–24, and GPQA, comparing against fixed-token and prompt-based budget-control baselines. They also apply Mid-Think to GRPO post-training after SFT and report improved accuracy and ~15% training-time reduction for Qwen3-8B on AIME and GPQA.","tokens_in":15050,"tokens_out":6698,"duration_ms":74286,"significance":"If the token-trigger claim were established, it would be a noteworthy mechanistic finding about hybrid reasoning models and could inform both inference-time control and RL post-training. The practical Mid-Think prompt is simple, training-free, and the reported accuracy–length trade-offs in Table 4 are encouraging. The paper also ships code and evaluates on multiple benchmarks and model families, which is a strength. However, the central causal attribution is not yet supported by the evidence as presented; the budget-control baseline is a proxy; and the RL results lack variance. These issues are addressable, and the empirical prompt format may still be a useful contribution even if the mechanistic story needs to be softened.","major_comments":[{"comment":"The central claim that a leading “Okay” token actively induces reasoning and the newline after “</think>” suppresses it is not identified by the presented evidence. Fig. 2 reports average attention, and Table 2 compares prompt formats, but no condition replaces “Okay” with a control token in the same slot, and no intervention (e.g., attention patching or ablation) is performed. Because early tokens are known to attract high attention regardless of content (attention sinks; Xiao et al. 2023, Gu et al. 2024), the attention concentration alone cannot establish causation. Moreover, the comparisons in Table 2 vary several factors at once: “No Think + Okay” differs from “Think” in both the position of “Okay” (after “</think>\\n\\n” rather than after “<think>”) and the presence of the “</think>” boundary. A minimal control would compare <think>\\nOkay, <think>\\nThe, <think>\\n###, and <think>\\n\\n</","section":"§2.2–2.3, Fig. 2, Table 2"},{"comment":"The budget-controlled baseline does not measure a model’s reasoning under a token budget; it measures continuation from a truncated prefilled chain. The protocol first generates a full Think-mode trace, keeps the first fraction of the reasoning tokens, concatenates them into the prompt, and lets the model finish. Under this protocol the model is effectively given the beginning of the answer, so the monotonic accuracy–budget curve in Fig. 3 is expected even without any notion of intermediate reasoning. Consequently, the statements that Mid-Think “corresponds to a budget of 0.5” and “surpasses the Pareto frontier” (Figs. 5–6) are comparisons against an axis that is not a true budget-control mechanism. Please either replace this baseline with an actual constrained-generation protocol (e.g., max-token early stopping or a forced-exit mechanism) or explicitly reframe the comparison as “prefill","section":"§3.2.1, Fig. 4"},{"comment":"The RL training results are reported as single runs with no variance or seed information. For Qwen3-8B, RL Mid-Think improves Think-test AIME accuracy from 69.8 to 72.4, but the Mid-Think run also produces longer outputs (15,318 vs. 13,330 tokens) and more “wait” tokens (44,142 vs. 34,701), so the gain may partly reflect a test-time compute effect. The claimed ~15% training-time reduction is not backed by a precise definition (wall-clock? GPU-hours?) in the experimental section; Figure 8 shows relative time but no quantitative table. Please report at least three seeds with standard deviations for both accuracy and training time, and, if the test-time length differs, include a length-controlled comparison or discuss the confound.","section":"§4, Table 5"}],"minor_comments":[{"comment":"The notation “Okay...” is ambiguous: does the literal prompt contain “Okay” followed by an ellipsis, or is the ellipsis a placeholder for model-generated content? Please clarify in the caption or table footnote.","section":"Table 3 and Fig. 1"},{"comment":"The attention analysis averages over all layers and heads. This may hide head-specific or layer-specific roles. Please report whether the concentration on “Okay” and on the newline is consistent across layers and heads, or at least across a few representative layers.","section":"§2.2"},{"comment":"The checkmark table is vague. For each model/dataset, report the frequency of “Okay” and related tokens, or at least a concrete count rather than a binary checkmark.","section":"Table 1"},{"comment":"The sentence “The specific token choice is not essential; we evaluate three variants: <reason>, <begin>, and <less think>” refers to the wrapper tags, not to the lexical trigger. This is fine, but it should be stated more precisely to avoid confusing tag choice with trigger-token choice.","section":"§3.1"},{"comment":"Wording such as “Pareto-optimal” and “consistently outperforms” is stronger than what single-seed, no-error-bar experiments can support. Please either add error bars or soften the wording.","section":"§3.2 and §4"}],"recommendation":"major_revision","confidential_remarks":"The practical Mid-Think prompt appears to be a useful empirical contribution, and the paper’s controlled comparisons in Table 4 are a step in the right direction. However, the paper’s central mechanistic claim is not yet supported because the key experiments lack causal controls, and the budget baseline is a proxy. These issues are fixable within the scope of a revision: add token-swap controls and/or attention interventions, rework or reinterpret the budget baseline, and report RL variance. If the authors cannot provide the causal controls, the manuscript should be reframed as an empirical prompt-engineering study without the title-level trigger-token causality claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this paper shows you can get Qwen3-class hybrid models to reason at intermediate length with a simple prompt tweak — put '<think>\\n\\n</think>\\n\\n<reason>\\nOkay...' at the start of the response. The empirical finding that a leading 'Okay' token flips the model into thinking mode, and a newline after </think> suppresses it, looks real, and the Mid-Think prompt reliably lands between the Think and No-Think regimes across benchmarks and model scales. The RL training result — roughly 15% time savings with slightly better AIME/GPQA after GRPO — is a nice bonus.\n\nWhat's genuinely new: the specific trigger-token observation and the Mid-Think combination. It builds on No-Wait, SpecExit, and Speculative Thinking, but nobody has isolated this particular pair of cues. The paper does solid multi-model, multi-benchmark evaluation, with a reasonable fixed-budget baseline, and it ships code. That counts.\n\nThe soft spots are the causal story and the lack of error bars. The attention analysis in §2.2 shows attention concentrating on 'Okay' and on the newline, but attention concentration is not causation — early tokens get attention-sink mass regardless. The behavioral controls in Table 2 are better than the stress test implies: comparing No-think with and without 'Okay' keeps everything else fixed, so the token does change behavior. But there is no control where 'Okay' is swapped for a neutral token in the same slot, so 'lexical trigger' is not actually identified. It might be that any content token in that position flips the mode. The Mid-Think prompt works pragmatically, but the claim that reasoning is 'governed by' these specific tokens goes beyond the evidence. Single-seed point estimates without variance also make the Pareto-frontier claims less crisp than they look.\n\nMinor: the budget baseline is a truncation proxy, not an oracle, and a direct comparison to No-Wait/SpecExit would have made the positioning cleaner. The paper itself notes it cannot do fine-grained budget control, so the title's 'intermediate-budget' claim is appropriately scoped.\n\nThis is worth a serious referee. The method is simple, reproducible, and likely useful to anyone doing inference-time reasoning control or RL post-training. Send it to review, but ask the authors to add a token-swap control and seed variability, and to soften the causal language to match what the experiments show.","headline":"A practical training-free budget dial for hybrid reasoning models, with a mechanistic story that overreaches but a trick that works.","tokens_in":15541,"tokens_out":2513,"would_cite":true,"duration_ms":26823,"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":"Reasoning in hybrid LLMs is switched by a few trigger tokens, not by instructions, and the paper exploits this with a training-free prompt format that delivers intermediate-budget thinking.","keywords":["token-level triggers","hybrid reasoning models","attention sink","intermediate budget","reasoning control","prompt engineering","reinforcement learning post-training","Pareto trade-off"],"falsifier":"A concrete test: in the standard Think prompt, replace the leading 'Okay' with a neutral token such as 'The' or 'Mm' in the same position, and measure wait count and accuracy. If the model still produces full reasoning traces, the 'Okay' token is not a necessary trigger and the mechanistic claim collapses. Alternatively, patch or ablate the attention from generated tokens to the 'Okay' token during generation; if reasoning persists, attention to that token is not causal.","tokens_in":14692,"feed_emoji":"🧠","tokens_out":3643,"duration_ms":33823,"temperature":0.7,"texified_at":"2026-08-05T20:46:37.409122+00:00","texify_note":"no math","pith_summary":"The paper argues that hybrid reasoning models enter 'think' mode because of a handful of token cues, not because of the high-level instructions they are given. In particular, a leading 'Okay' token flips the model into reasoning, while the newline that follows the closing </think> tag pushes it into no-think mode. The authors propose Mid-Think, a simple prompt format that combines both cues, so the model produces a middle amount of reasoning with no extra training. That matters because it offers a cheap, interpretable dial for trading accuracy against output length, and it also makes reinforcement-learning post-training faster and more accurate.","texify_model":"deepseek-v4-flash","feed_headline":"Trigger tokens, not instructions, switch LLM reasoning mode","feed_subtitle":"Leading 'Okay' flips thinking on; a newline flips it off. Mid-Think combines both for free middle-budget reasoning.","key_machinery":"The central objects are the token-level triggers: the leading 'Okay' token that acts as a reasoning-on switch, and the newline pair after </think> that acts as a reasoning-off switch. The Mid-Think prompt format stacks both in one context. It works by keeping the no-think structural cue while adding the think trigger, so the model settles into a middle reasoning budget automatically, with no training and no per-instance token limit. Attention analysis is the evidence, and the prompt itself is the mechanism.","core_discovery":"On the paper's own terms: for Qwen3-series hybrid models, the mode switch between thinking and not thinking is not controlled by the user's instruction but by a small set of token-level triggers. Attention analysis shows that later generated tokens concentrate their attention on an opening 'Okay' token when the model reasons, and on the newline pattern after </think> when it does not. Controlled prompt experiments confirm that adding 'Okay' reproduces Think-mode accuracy and wait counts, while the </think>+newline pattern suppresses them. Mid-Think (<think>\\n\\n</think>\\n\\n<reason>\\nOkay…) combines both triggers to induce intermediate-budget reasoning that sits on or beyond the Pareto frontie","pith_inferences":["The trigger-token story is likely not limited to these specific models: any model fine-tuned on reasoning traces that uniformly start with 'Okay' could develop the same overfit, so the format may transfer to other open and closed models as long as their SFT data shares that lexical pattern.","A natural extension is to find other trigger tokens (e.g., 'Let's', 'Hmm', 'wait') and see whether each maps to a different reasoning depth; if so, one could build a discrete family of budget settings from token choice alone.","A stronger causal test would be to patch or ablate the 'Okay' token's attention during generation; if reasoning persists without it, the mechanism is likely an attention-sink artifact rather than true causality, though the Mid-Think prompt might still work for other reasons.","If token-level triggers are so dominant, safety or evaluation work on hybrid models should control for them; otherwise a small prompt change could silently switch a model in or out of reasoning."],"forward_implications":["If the trigger-token claim holds, reasoning budget can be tuned without retraining or per-instance token limits; a single prompt change yields a middle accuracy-length operating point.","The same format carries over to RL after SFT, implying that other post-training phases can inherit the efficiency and accuracy gains without architectural changes.","Because the trigger works across model scales and training regimes listed in the paper, the finding suggests a general mechanism in hybrid reasoning models: the policy has overfit to low-level lexical cues, which can be exploited for control.","Mid-Think can be used as a baseline for future budget-control research, since it needs no training and no search over token limits.","In RL, Mid-Think preserves no-think behavior better than Think training while achieving higher think-mode accuracy, suggesting it is a better objective for hybrid models that need to do both."],"fun_headline_variants":["Mid-Think: token triggers pre-set LLM reasoning budget","LLM reasoning toggled by a token, not your instructions","One 'Okay' token flips LLM thinking on; newline off","Trigger tokens control thinking mode without any training","Mid-Think gets intermediate reasoning with just token tweaks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the 'Okay' token is the active cause of reasoning behavior, not a side effect of attention-sink or position effects; the paper's evidence is observational (attention patterns and prompt variants) and does not include a control that replaces 'Okay' with another token in the same position.","fun_headline_variants_meta":{"raw":{"variants":["Mid-Think: token triggers pre-set LLM reasoning budget","LLM reasoning toggled by a token, not your instructions","One 'Okay' token flips LLM thinking on; newline off","Trigger tokens control thinking mode without any training","Mid-Think gets intermediate reasoning with just token tweaks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00115,"raw_usage":{"total_tokens":4590,"prompt_tokens":715,"completion_tokens":3875,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":3790}},"tokens_in":459,"tokens_out":3875,"duration_ms":29921,"temperature":1.0,"reasoning_tokens":3790,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T11:11:21.803700+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: in the standard Think prompt, replace the leading 'Okay' with a neutral token such as 'The' or 'Mm' in the same position, and measure wait count and accuracy. If the model still produces full reasoning traces, the 'Okay' token is not a necessary trigger and the mechanistic claim collapses. Alternatively, patch or ablate the attention from generated tokens to the 'Okay' token during generation; if reasoning persists, attention to that token is not causal.","supporting_citations":[],"review_version":1}