{"id":"40c8faf1-4fec-416d-a8b8-4812455f02c1","arxiv_id":"2509.01236","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper finds that as exemplar count increases, LLMs shift from pretrained priors to in-context signals, and long chain-of-thought prompts induce longer reasoning chains that often improve accuracy.","lead":"This paper studies why chain-of-thought reasoning works in large language models, using lexical analysis and noisy-exemplar experiments to trace how models balance prior knowledge against examples in the prompt. It finds that more exemplars shift the model toward in-context signals, and that long chain-of-thought prompts induce longer, often better reasoning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3.2's label-flipping claim is inferred from aggregate accuracy only; no per-label outputs or exact Coin Flip accuracies are reported, so ICL-override could be copying/confusion rather than systematic mapping.","rationale":"The reader's weakest assumption identifies the same gap: the paper infers 'systematic label flipping' from aggregate accuracy without analyzing actual outputs. I agree this is the most load-bearing uncertainty because the paper's novel challenge to prior work (e.g., Wei et al. 2023) depends on showing that an 8B model learns an inverted input-label mapping, not merely that its accuracy degrades under noisy demonstrations. However, I would add a nuance: in a balanced binary task like Coin Flip, an accuracy drop to near 0% would itself demonstrate systematic flipping, so the issue is best framed as insufficient reporting rather than a definitively false inference. The paper does not give exact Coin Flip accuracies or per-label error rates, so the reader cannot tell whether the curve falls to ~0% (flipping) or ~50% (copying/confusion). This underdetermination is a correctness risk, not a stylistic concern. The abstract's blanket downstream-improvement claim and the absence of released prompts/data are secondary issues already noted by the reader. My concrete test would resolve the mechanism, and the reader's CONDITIONAL verdict remains appropriate: the paper is promising but currently under-evidenced on its central interpretive claim.","tokens_in":9333,"tokens_out":6393,"duration_ms":77290,"concrete_test":"Rerun the 40-shot false-answer CoT condition on Coin Flip with LLaMA3.1-8B and Gemma2-9B, recording the predicted label per test item. Compute P(pred='no' | true='yes') and P(pred='yes' | true='no') separately, plus overall accuracy, over 5 seeds with greedy decoding. Also compute a 'copy-random-exemplar' baseline (sample one false-answer exemplar per test item and output its answer). If both per-label flip rates are ≥0.8 and significantly exceed the copy baseline, the label-flipping interpretation is confirmed. If flip rates are near chance or if errors are concentrated on one label, the correct interpretation is confusion/label prior, and Section 3.2's central claim is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that increasing false-answer exemplars shifts decisions from pretrained priors to ICL and produces systematic label flipping in closed-domain tasks—rests on Fig. 4, but the paper reports only aggregate accuracy curves. In Coin Flip, 'systematic label flipping' is invoked to explain the decline, and this is used to challenge Wei et al.'s claim that small models cannot learn flipped labels (Section 3.2). However, a decline in accuracy is also consistent with (a) copying answers from the false-answer exemplars (which, if balanced, would yield ~50%), (b) a global response bias (e.g., always 'no'), or (c) general confusion/instability from contradictory evidence. The paper provides no output-level analysis (per-true-label flip rates), no exact accuracy values, and no error bars, so the mechanism—which is load-bearing for 'ICL overrides pretrained priors'—is underdetermined. If the Coin Flip accuracy at 40-shot is near 0%, systematic flipping is supported; if it is near 50%, it is not. The confidence trajectories in Fig. 5 are a single case study and do not resolve this.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies Chain-of-Thought (CoT) prompting through the dual lens of in-context learning (ICL) and pretrained priors. It reports three experiments: (i) a lexical-level analysis of model-generated rationales comparing zero-shot CoT, manual CoT, and task-agnostic CoT; (ii) a noisy-exemplar scaling study that increments false-answer and false-rationale exemplars, claiming that sufficient exemplars shift the model from pretrained priors to ICL and that 8B models can flip labels in closed-domain tasks; and (iii) a prompt-engineering study that uses long CoT prompts distilled from R1/QwQ models to induce slow thinking. The paper claims three findings: models learn lexical and deeper reasoning structures while relying on pretrained priors; exemplar scaling shifts decision-making toward ICL while misleading prompts introduce instability; and long CoT prompting improves downstream performance.","tokens_in":9636,"tokens_out":6345,"duration_ms":71455,"significance":"If the label-flipping and prior-shift claims held, the paper would contribute to the debate on the nature of ICL and would challenge prior results (Wei et al., 2023) about model-size thresholds for learning flipped labels. The RQ3 results also have practical value for prompt design. The paper has notable strengths: it spans multiple model families and sizes (Gemma2-9B/27B, LLaMA3.1-8B, Qwen2.5-32B), the false-exemplar perturbation design is simple and clearly described, and the authors include a limitation section acknowledging dataset, noise-level, and decoding issues. However, the central mechanistic interpretation is not tightly supported by the reported measures, and several quantitative details are missing.","major_comments":[{"comment":"The central inference of RQ2 is that 'the model tends to learn an input-label mapping that leads to systematic label flipping' and that this refutes Wei et al.'s claim that smaller models cannot learn flipped labels. The evidence is an aggregate accuracy decline in Coin Flip. A decline from high to near-chance accuracy is equally compatible with copying the (possibly balanced) false answers, with a global response bias, or with general confusion from contradictory evidence; only a decline to ~0% (or per-label flip rates close to 100%) would establish systematic flipping. The paper reports no exact Coin Flip accuracies, confusion matrices, or error bars. Please add output-level analysis (accuracy by true label, flip rate, and answer distribution at 4-shot and 40-shot) before the 'ICL overrides pretrained priors' claim is accepted.","section":"Section 3.2 (Fig. 4)"},{"comment":"The confidence analysis is one case study: a single trace of per-step probabilities for three prompts. The claim that CoT prompts maintain stable probabilities while false prompts fluctuate is not substantiated without aggregation over a sample of questions. Please provide mean and variance/autocorrelation statistics across the test set, or at least across multiple examples, with prompt type as the independent variable. As presented, Fig. 5 is anecdotal and cannot carry the 'misleading prompts introduce instability' takeaway.","section":"Section 3.2 (Fig. 5)"},{"comment":"The paper states that 'a positive correlation' exists between average verb count and accuracy, and that an 'optimal count' is evident, but no correlation coefficients, fitted curves, or confidence intervals are reported. The scatterplots combine multiple prompt types and model sizes in one panel, making within-task relationships difficult to assess. Please quantify these lexical-level claims (e.g., Spearman correlation per model/task, or a regression controlling for model size and task difficulty) and state the criteria for an 'optimal' verb count.","section":"Section 3.1 (Figs. 2 and 3)"},{"comment":"The 'improving performance' claim is based on point estimates without variance or significance tests. Some entries contradict the broad claim, e.g., LLaMA3.1-8B-Instruct on GSM8K with QwQ-32B-Long-CoT (0.7695 to 0.6202) and LLaMA3.1-8B-Instruct on DATE with DS-Qwen32B-Long-CoT (0.6504 to 0.6449). The 'Avg. tokens total' is reported but not analyzed as a continuous variable; the optimal-length conclusion would benefit from a systematic length-versus-performance analysis over a range of token budgets rather than selected model/task pairings.","section":"Table 1 (Section 3.3)"}],"minor_comments":[{"comment":"Typo: 'Our findings challenge previousous claims' should read 'previous claims'.","section":"Section 3.2"},{"comment":"'A vg. tokens total' should be 'Avg. tokens total'.","section":"Table 1 caption"},{"comment":"The legend uses 'LLaMA3-8B' while the text uses 'LLaMA3.1-8B'; make consistent.","section":"Fig. 3"},{"comment":"The formalization of CoT prompting is incomplete: 'Input: X = {(q_i, r_i, a_i)}N_{i=1}, q' lacks clear set notation and the target question q is not incorporated into the conditioning set. Please rewrite the equation cleanly.","section":"Introduction"},{"comment":"The 'token generation probability at each time step' is used as confidence, but it is not specified whether this is the softmax probability over the full vocabulary, top-k logits, or a normalized logit transformation. Clarify how the 'Probability Value' in Fig. 5 is computed.","section":"Section 3.2"},{"comment":"The authors acknowledge that greedy decoding 'often resulted in endless repetitions' for long CoT prompts, but the frequency of such repetitions is not reported. This could bias the RQ3 accuracy numbers; please state how many generations (per condition) were truncated or excluded.","section":"Section 6 (Limitations)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for the journal, but the hinge between the experimental observation (aggregate accuracy declines) and the mechanistic interpretation ('systematic label flipping' in 8B models) needs to be strengthened. The authors can address this within the manuscript's scope by adding output-level statistics for Coin Flip and aggregating the confidence analysis. I recommend major revision rather than rejection because the core observations are reproducible and the missing analyses are standard. The RQ3 section would also benefit from error bars or significance tests, given that several point estimates go against the paper's summary claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look for two reasons. First, the Coin Flip scaling results in RQ2 give a real datapoint against the 'small models can't override pretrained priors through ICL' line. Second, the verb-count analysis in RQ1 is a new, if rough, way to see what models absorb from exemplars. The paper documents behavior carefully across models and tasks, uses greedy decoding consistently, and is honest about its own limits (repetition, task scope). The RQ3 experiments with distilled long-CoT prompts are practically useful, and Table 1 is detailed enough to show the non-monotonic length effect.\n\nThe soft spot is Section 3.2. The claim that the accuracy decline in Coin Flip reflects systematic label flipping—the model learning an input-label mapping—is inferred from aggregate accuracy curves only. No per-label flip rates, no exact accuracy values, no variance. The same curve could appear from copying false answers or a response bias. That matters because the ICL-overrides-priors interpretation depends on ruling those out. The single confidence-trace case study doesn't settle it. The stress-test framing is right: at 40-shot, accuracy near 0% would support flipping; near 50% would support copying. The paper doesn't report which.\n\nThe abstract also overstates RQ3: long CoT prompting improves performance, full stop. Table 1 has several clear declines (e.g., LLaMA3.1-8B-Instruct on GSM8K with QwQ-32B-Long-CoT; Qwen2.5-32B-Instruct on GSM8K with DS-LLaMA8B-Long-CoT). The body is more careful, but the abstract needs rewording. And no prompts or data are released, no error bars reported, so the main empirical claims are not directly replicable as-is. The citation pattern is fine—the paper engages the right prior work, and the challenges to Wei et al. and Schaeffer et al. are clearly framed.\n\nWho this is for: people working on CoT mechanisms, ICL, or prompt robustness. It's not a breakthrough, but it's a useful empirical addition that deserves discussion. I'd send it to review. The right call is a revise-and-resubmit with a request for output-level analysis in RQ2 and abstract/table alignment.","headline":"A useful empirical paper whose central RQ2 label-flipping claim is under-evidenced; deserves a referee but needs per-label output analysis.","tokens_in":10109,"tokens_out":4636,"would_cite":true,"duration_ms":49269,"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":"As exemplars accumulate, chain-of-thought reasoning shifts from pretrained priors to in-context signals, so enough wrong examples can flip even an 8B model's answers.","keywords":["Chain-of-Thought prompting","In-context learning","Pretrained priors","Large language models","Noisy exemplars","Slow thinking","Reasoning mechanisms","Prompt engineering"],"falsifier":"Record the actual final-answer tokens the model emits under 40 false-answer CoT exemplars on Coin Flip. If the model's output distribution still concentrates on correct labels, or collapses onto copied exemplar answers regardless of the input, then the claimed systematic input-label mapping and label flipping are not real, and the ICL-overrides-priors conclusion would need revision.","tokens_in":9248,"feed_emoji":"🧠","tokens_out":7348,"duration_ms":79908,"temperature":0.7,"pith_summary":"Chain-of-thought prompting is standard practice, but why it works is unsettled. This paper argues that the answer lies in a shifting balance between two sources: pretrained knowledge the model already has and in-context signals from the exemplars in the prompt. With few examples, pretrained priors dominate and noise barely matters; as the number of exemplars grows, in-context signals strengthen and can override prior knowledge, so systematically wrong answers begin to drive the model's choices—even in an 8B model on a binary task. The same logic implies exemplar quality matters more as prompt size grows. Finally, the paper shows that prompts distilled from long-reasoning models induce longer chains and better downstream accuracy, with an optimal length.","feed_headline":"Noisy examples override a model's pretrained knowledge at scale","feed_subtitle":"With enough misleading demonstrations, even an 8B model flips labels, and long CoT prompts elicit slower, better reasoning.","key_machinery":"The exemplar-count dial plus a lexical decomposition of rationales. Starting with a small number of CoT exemplars and incrementally adding corrupted ones turns the strength of in-context learning into a controllable scalar. A complementary lexical analysis splits model-generated rationales into structure words, feature words, verbs, and location/person entities to separate imitation of form from reliance on semantic priors; confidence is tracked through per-token generation probabilities.","core_discovery":"The paper's central discovery is a mechanism with a dial: the number of exemplars controls whether CoT reasoning follows pretrained priors or in-context signals. In controlled experiments, task-agnostic CoT prompts leave semantic reasoning intact while changing lexical form, and accuracy tracks the number of reasoning verbs with an optimum, indicating the model copies structure while relying on prior knowledge for content. Then, replacing answers with wrong answers in increasing numbers of exemplars shows little effect at 4–5 shots but progressive accuracy collapse at 40 shots; in the closed-domain binary Coin Flip task, an 8B model systematically flips labels. False rationales likewise degr","pith_inferences":["The abrupt shift from priors-dominated to ICL-dominated behavior as exemplars accumulate suggests a phase-transition-like dynamic; a natural test is to fit accuracy as a function of exemplar count and look for a critical count that changes with model size.","If the mechanism is general, closed-domain tasks with small label spaces (multiple choice, yes/no, classification) are the most vulnerable to prompt poisoning, while open-ended generation is more robust.","The slow-thinking transfer result could support self-improvement loops: a model's own long outputs, once filtered for quality, could be reused as prompts or training data, provided the decoding repetition problem is solved.","The optimal-length finding implies adaptive control of reasoning length per question—stop early on easy items, extend on hard ones—could capture most gains at lower inference cost."],"forward_implications":["Few noisy exemplars are harmless, but at scale they can outweigh what the model knows, so prompt hygiene matters increasingly with prompt length.","Contrary to earlier claims that only very large models can learn flipped labels in context, CoT prompting lets an 8B model flip answers on a closed-domain binary task.","Misleading rationales make a model less confident as well as less accurate, so unreliable prompts are detectable in generation probabilities.","Long-CoT prompts distilled from stronger reasoners can transfer slow-thinking behavior to smaller models and improve accuracy, up to an optimal length.","Reasoning length is not 'more is better': beyond an optimum, longer chains hurt, and the optimum grows with model size and task difficulty."],"supporting_citations":[{"why":"Supplies the standard CoT prompts, exemplars, and tasks (GSM8K, Coin Flip) that the paper's experiments build on and measure against.","marker":"[30]"},{"why":"Counterfactual study claiming noisy rationales and answers have little effect; the paper's accuracy-collapse results challenge this baseline.","marker":"[16]"},{"why":"Earlier result that invalid logic yields equivalent gains; the paper's false-rationale experiments show degradation at scale.","marker":"[22]"},{"why":"Prior scaling claim that smaller models cannot learn flipped labels; the paper's 8B closed-domain flipping result directly contradicts it.","marker":"[31]"},{"why":"Framework distinguishing task recognition from task learning; motivates the paper's method of replacing labels and answers to separate ICL signals from priors.","marker":"[20]"},{"why":"Claim that models rely on pretrained priors rather than task-specific knowledge from demonstrations; the paper's early-shot results align with and extend this.","marker":"[18]"},{"why":"Used to distill additional exemplars from training data, enabling the 40-shot noisy-exemplar study.","marker":"[19]"},{"why":"Supplies the zero-shot CoT instruction and the Last Letters task, used as a baseline and evaluation setting.","marker":"[9]"},{"why":"Shows that more CoT length is not always better; the paper's optimal-length finding is consistent with this result.","marker":"[32]"},{"why":"Provides thinking-optimal scaling of test-time compute; supports the interpretation that an optimal CoT length exists.","marker":"[33]"}],"fun_headline_variants":["Exemplar count flips CoT from prior knowledge to in-context copying","More examples shift CoT reasoning from pretrained priors to context","Misleading prompts destabilize CoT, long prompts slow it down","The reasoning dial: how exemplars control CoT reliance on priors","8B model flips labels when given enough noisy CoT examples"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the accuracy collapse under many false-answer exemplars means the model actually learns to output the flipped labels, rather than merely copying final answers or getting confused; the paper's own limitation section adds that greedy decoding often produced endless repetitions, a separate risk for the slow-thinking result.","fun_headline_variants_meta":{"raw":{"variants":["Exemplar count flips CoT from prior knowledge to in-context copying","More examples shift CoT reasoning from pretrained priors to context","Misleading prompts destabilize CoT, long prompts slow it down","The reasoning dial: how exemplars control CoT reliance on priors","8B model flips labels when given enough noisy CoT examples"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1457,"prompt_tokens":729,"completion_tokens":728,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":643}},"tokens_in":473,"tokens_out":728,"duration_ms":7252,"temperature":1.0,"reasoning_tokens":643,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:43:53.747142+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the actual final-answer tokens the model emits under 40 false-answer CoT exemplars on Coin Flip. If the model's output distribution still concentrates on correct labels, or collapses onto copied exemplar answers regardless of the input, then the claimed systematic input-label mapping and label flipping are not real, and the ICL-overrides-priors conclusion would need revision.","supporting_citations":[{"cited_title":"In: Advances in Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"Supplies the standard CoT prompts, exemplars, and tasks (GSM8K, Coin Flip) that the paper's experiments build on and measure against."},{"cited_title":"In: Findings of the Association for Computational Linguistics: EMNLP 2023","cited_arxiv_id":null,"evidence_quote":"Counterfactual study claiming noisy rationales and answers have little effect; the paper's accuracy-collapse results challenge this baseline."},{"cited_title":"In: Findings of the Association for Computational Linguistics: ACL 2023","cited_arxiv_id":null,"evidence_quote":"Framework distinguishing task recognition from task learning; motivates the paper's method of replacing labels and answers to separate ICL signals from priors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Claim that models rely on pretrained priors rather than task-specific knowledge from demonstrations; the paper's early-shot results align with and extend this."},{"cited_title":"In: Advances in Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"Supplies the zero-shot CoT instruction and the Last Letters task, used as a baseline and evaluation setting."}],"review_version":1}