{"id":"e524de38-2227-489e-9817-73360782cfd7","arxiv_id":"2505.22548","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Emotion-o1 uses distillation, supervised fine-tuning, and reinforcement learning to make an 8B LLM adjust its chain-of-thought length per emotion task, improving F1 and reducing reasoning cost.","lead":"This paper trains an 8B language model to adapt how much it reasons before answering emotion classification questions, switching between short and long chains of thought depending on task difficulty. It reports accuracy gains on sentiment, emotion, humor, and sarcasm benchmarks while cutting reasoning length versus large reasoning models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never states whether distillation and RL use train or test splits; if test labels enter via Eq. (2) or the PPO reward, the reported F1 gains are leakage, not adaptive reasoning.","rationale":"Reading the paper in good faith, the core idea—distilling variable-length CoT from a strong reasoner, SFT, then RL with a multi-part reward—is coherent and the ablation study (Fig. 4) suggests length and structure choices matter. However, the paper's most consequential empirical claim rests on a foundational detail that is never stated: whether the distillation set D and the RL reward labels come from the same splits as the evaluation. The methodology says 'for each sample (x_i, y_i)' without restricting to train; the only explicit split in the paper is the evaluation set. Since Eq. (2) uses ground-truth labels for rejection sampling and Eq. (6) uses ground-truth labels as the accuracy reward, any test instance in D would leak the answer into SFT/RL. This would trivially produce large F1 gains over the untrained backbone and explain the apparent superiority over zero-shot larger models. The reader's concern about the quality of intermediate reasoning traces is valid but not the primary risk; even perfect traces cannot rescue a pipeline that trains on the test set. Therefore, the central claim is unverdictable without disclosure of the split. I propose UNVERDICTED until the authors clarify this; if the split is proper, the condition is easily resolved and the paper could return to conditional acceptance. My agreement with the reader is partial: we both focus on the supervision signal, but the split leakage is a more elementary and more damaging failure mode.","tokens_in":14320,"tokens_out":8336,"duration_ms":92003,"concrete_test":"Request the authors to disclose the exact data split used in Eq. (1)-(3) and in the PPO reward, and to provide the train/test split filenames in the released code. Alternatively, rerun the entire pipeline using only the official training splits of MELD, Sarcasm Corpus V2, and Reddit Humor, keeping evaluation on the official test splits; if the reported F1 gains over LLaMA-3.1-8B (10/5/18/27%) shrink materially or disappear, the published numbers are contaminated.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"Section 'Structured Emotion Reasoning Distillation' builds dataset D (Eq. 3) by sampling from DeepSeek-R1 'for each sample (x_i, y_i)' from MELD, Sarcasm Corpus V2, and Reddit Humor, then filtering by ground-truth label (Eq. 2). The paper never states that these samples come from the training split; the only split mention is 'Evaluation Setup: We conduct tasks on the test set of the benchmark dataset.' If D includes test instances, then SFT (Eq. 5) and the PPO accuracy reward (Eq. 6) directly expose the model to test labels, and the length targets (Eqs. 7-9) are computed from test-response distributions. This would fully explain the 10/5/18/27% F1 gains over the backbone and the apparent parity with much larger zero-shot models. The reader's concern about flawed reasoning traces is secondary: even if every trace were human-validated, test-set contamination would invalidate all reported numbers. The split is not a reproducibility detail; it is a necessary condition for the central claim to be meaningful.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Emotion-o1, an 8B-parameter LLaMA-3.1-based model for emotion understanding that generates chain-of-thought reasoning whose length adapts to task complexity. Training proceeds in three stages: distillation of variable-length, linear/non-linear reasoning traces from DeepSeek-R1; supervised fine-tuning on those traces; and PPO reinforcement learning with a reward combining accuracy, task-specific length targets, structural keywords, and a repetition penalty. The authors report large F1/accuracy gains over the LLaMA-3.1 backbone on MELD sentiment and emotion recognition, Sarcasm Corpus V2, and Reddit Humor, competitive or superior results against much larger closed-source models, and substantially shorter CoTs than OpenAI-o1 and DeepSeek-R1. The paper claims that Emotion-o1 learns to balance reasoning depth and efficiency automatically.","tokens_in":14614,"tokens_out":7716,"duration_ms":86076,"significance":"If the empirical results hold, the work is significant for practical emotion understanding: an 8B model could approach frontier-level performance with much lower inference cost by allocating reasoning depth per task. The paper also provides a reasonably complete pipeline (distillation, SFT, RL) and a broad comparison against nine baselines, with ablations and length-distribution plots. However, the paper does not release code, data, or hyperparameters, and the central 'adaptive' claim is partly built into the reward design. The significance therefore depends on resolving the data-split and reproducibility concerns below.","major_comments":[{"comment":"The paper never specifies that the distilled dataset D is constructed from the training splits of the four benchmarks. Since the distillation prompt p(x_i, y_i, c) includes the gold label and Eq. (2) retains only responses whose final label matches y_i, any test instances in D would leak test labels into SFT (Eq. 5) and into the RL accuracy reward (Eq. 6). The only explicit split statement concerns evaluation on the test set. Please state the exact split construction and verify that no test example appears in D or in any reward computation; if this is not already the case, the experiments must be rerun with strict train/test separation.","section":"Structured Emotion Reasoning Distillation / Evaluation Setup"},{"comment":"The length reward is task-specific: L_min and L_max are P5/P95 quantiles of the distilled response lengths and L_base is a weighted combination of an author-set L_exp and the median of those same lengths. Consequently, the finding in Table 2 that Sentiment gets shorter CoTs and Sarcasm/Humor get longer ones is substantially determined by the reward, not by the model autonomously discovering task complexity. The authors should provide evidence of adaptivity beyond the encoded targets, such as an ablation without the length reward or an evaluation on unseen emotion tasks, or should soften the 'automatically adjusts' framing.","section":"Verifiable Reward RL, Eqs. (7)-(9) / CoT Length Analysis"},{"comment":"The abstract's 'F1 score increases of 10%, 5%, 18%, and 27%' do not correspond to a single metric in Table 1: the sentiment number is Macro-F1 (+10.1 percentage points), the emotion number is accuracy (+4.9 points) rather than Macro-F1 (+14.5 points) or Weighted-F1 (+10.2 points), and the humor/sarcasm numbers are Macro/Weighted-F1. The paper should report all four tasks with the same metric, and because Table 1 is an average of only 3 runs, standard deviations or confidence intervals are needed to support claims such as '+1.1% over Grok-3' and '+2% over Claude-3.7'.","section":"Abstract and Main Result"},{"comment":"The rejection sampling criterion checks only whether the final label matches the gold label, and the prompt itself provides the gold label, so the distilled traces can be fluent post-hoc rationalizations of a known answer rather than valid reasoning. No human evaluation or consistency check is reported for the intermediate reasoning steps. Without such validation, the improved benchmark scores could reflect better label-conditioned generation rather than improved emotion reasoning, and the interpretability claim is weakened.","section":"Structured Emotion Reasoning Distillation, Eq. (2)"},{"comment":"The method depends on many hyperparameters—alpha, L_exp, L_min/L_max, s_min/s_max/s_base, lambda, N_A/N_C, w_acc/w_length/w_struct/w_repeat, beta, tau, epsilon_acc, and the sampling count N—yet none are reported, and no code or data is provided. This makes the empirical results impossible to reproduce or independently check; please provide the full hyperparameter configuration and release the distillation/filtering scripts and the constructed dataset (or a representative sample) with guaranteed train/test separation.","section":"Training Setup / hyperparameters"}],"minor_comments":[{"comment":"The Introduction reports length reductions as '73% (54%), 52% (27%), 83% (70%), 70% (58%)' while the Visualization section gives 83% for Sarcasm and 52% for Emotion; clarify whether these are medians or means, and which baseline (OpenAI-o1 or DeepSeek-R1) each number refers to.","section":"Introduction and Visualization of CoT Length"},{"comment":"The reference list contains several entries that appear unrelated to the manuscript (e.g., [7]-[10], [13], [17], [22], [28]-[30]) and duplicates ([16] appears to duplicate [11]; [18] duplicates [17]; [29] and [30] duplicate each other); these should be cleaned up.","section":"References"},{"comment":"The probability density plots are difficult to read; provide the median, percentiles, and standard deviations in tabular form and specify that lengths are measured in tokens.","section":"Figure 5"},{"comment":"The statement 'We adopt the LLaMA-3.1-8B as our base model' contains a grammar error; it should read 'We adopt LLaMA-3.1-8B as our base model.'","section":"Experiments, Training Setup"},{"comment":"The model name is inconsistently capitalized; please unify to 'Emotion-o1' everywhere, including the Main Result paragraph where 'emotion-o1' appears.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The train/test split ambiguity is the most serious issue; if the authors cannot confirm that D is built exclusively from training splits and that no test labels enter the RL reward, the paper should not be published in its current form. I would also ask the editor to verify the reference list integrity, as several entries seem unrelated to the paper's topic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a plausible adaptive-CoT recipe for emotion tasks, and the gains over the backbone are consistent across four benchmarks. But there's one issue that has to be resolved before I'd trust any number in it: the paper never states whether distillation and the PPO reward use training or test splits. Eq. (2) filters distilled responses by ground-truth label, and Eq. (6) rewards accuracy against ground truth; if test labels enter anywhere in that pipeline, the 10/5/18/27% gains are leakage, not adaptive reasoning. The evaluation section only says \"we conduct tasks on the test set,\" which doesn't tell us what the distillation/RL saw. This is a necessary clarification, not a reproducibility nicety.\n\nWhat's genuinely new: combining length/structure-distilled CoT with SFT plus a four-part reward for emotion classification is a legit integration. The ablation in Fig. 4 is the most informative part—short CoT matches long on sentiment and emotion, long CoT clearly helps humor and sarcasm. That's a real empirical finding. The efficiency numbers (83% length reduction vs OpenAI-o1) are striking, though they follow partly from the length reward the authors designed; calling this \"adaptive\" is a bit generous, but the task-dependent gains hold up in the ablation.\n\nSoft spots beyond the split ambiguity: the abstract mixes metrics—the 10% and 5% are Macro-F1/accuracy while the 18% and 27% are Weighted-F1—and that's misleading. Table 1 reports \"average results from 3 experiments\" but no standard deviations or significance tests. Hyperparameters like alpha, L_exp, lambda, N_A, N_C, and all reward weights are listed as \"empirically tuned\" but never given. The references include duplicate and extraneous entries, which suggests a hasty compile.\n\nBottom line: if the split is clean and the metrics are reported consistently, the central claim is likely supportable. As submitted, the paper is a conditional accept at best. It deserves a serious referee round—the idea is useful and the ablation is honest—but the authors need to state splits, report variance, and fix the abstract before I'd believe the headline numbers.","headline":"Useful adaptive-CoT recipe for emotion tasks, but a train/test split ambiguity and metric sloppiness keep the headline gains from being verifiable.","tokens_in":15105,"tokens_out":2716,"would_cite":false,"duration_ms":30332,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Emotion-o1 claims that letting a model choose how long to reason—short for simple emotions, long for sarcasm and humor—raises F1 by up to 27% and cuts reasoning length by 83%.","keywords":["emotion understanding","chain-of-thought reasoning","adaptive reasoning length","reinforcement learning","knowledge distillation","sarcasm detection","humor detection","sentiment classification"],"falsifier":"Take the distilled training set and corrupt the intermediate steps of every trace that ends in the correct label—shuffle the steps or replace them with generic filler—then retrain the same pipeline. If final-label accuracy stays roughly unchanged, the model's gains come from the label-conditioned distribution of the traces rather than from the reasoning content, and the adaptive-reasoning interpretation of the result fails.","tokens_in":14137,"feed_emoji":"🧠","tokens_out":10455,"duration_ms":103420,"temperature":0.7,"pith_summary":"This paper argues that for emotion understanding, a language model's reasoning budget should be per input, not fixed. Emotion-o1 learns to emit short, direct chains of thought for simple tasks like sentiment and longer, reflective chains for pragmatically complex cases like sarcasm and humor. The training pipeline distills variable-length reasoning traces from a strong reasoner, supervises an 8-billion-parameter student on them, then applies reinforcement learning with a reward that balances accuracy, brevity, structure, and redundancy. On four emotion benchmarks the paper reports F1 gains of 10%, 5%, 18%, and 27% over its backbone, while cutting median reasoning length by up to 83% compared with a fixed long-reasoning system. If these results hold, adaptive allocation of reasoning depth is a practical way to get competitive emotion understanding from a small model.","feed_headline":"Adaptive reasoning length lifts emotion understanding up to 27% in F1","feed_subtitle":"An 8-billion-parameter model matches or beats much larger systems on sentiment and sarcasm while thinking up to 83% less.","key_machinery":"The load-bearing mechanism is adaptive chain-of-thought length control, where chain-of-thought is the sequence of intermediate reasoning steps a model writes before its final answer. The policy that maps each input to a reasoning trace of appropriate length and structure is induced by distilling a dataset of short/long and linear/non-linear traces from a strong reasoner, keeping only traces whose final label matches the ground truth, fine-tuning on those traces, and then running PPO with a reward that has four terms: accuracy, closeness to quantile-derived length targets, presence of structured reasoning actions (decomposition, reflection, verification), and a penalty for repeated content. The length targets are computed from the distilled trace-length distribution blended with hand-set expectations, so the reward literally encodes the premise that simple emotion tasks should be answered briefly and hard ones should be allowed to think at length.","core_discovery":"The central claim is that reasoning depth is a task-specific resource and can be optimized. Emotion-o1 learns to compress reasoning for straightforward cases (sentiment, basic emotion recognition) and to expand it for pragmatically demanding cases (sarcasm, humor). It is built in three stages: rejection-filtered distillation of short/long and linear/non-linear reasoning traces from DeepSeek-R1, supervised fine-tuning on those traces, and PPO reinforcement learning guided by a four-part reward for accuracy, closeness to task-specific length targets, structured reasoning actions, and redundancy. On four benchmarks the paper reports F1 gains of 10%, 5%, 18%, and 27% over its LLaMA-3.1-8B backbone, and reported margins of 1.1% over Grok-3 and 2% over Claude-3.7 on sentiment and sarcasm respectively, while reducing median reasoning length by up to 83% compared with OpenAI-o1. The discovery is that the precision-efficiency trade-off in emotion reasoning can be actively shaped rather than accepted as fixed.","pith_inferences":["Beyond the paper, the same adaptive-length recipe should transfer to other classification domains with mixed difficulty, such as hate-speech detection or clinical note triage, because the reward needs only a verifiable label and a length distribution.","A testable extension not run here is a sweep of the expected-length parameter $L_{\\text{base}}$ with all other components fixed; if F1 is insensitive to that target, the gains come from the distilled traces, not from length adaptation.","Since the paper itself notes that it considers only text, a natural next test is multimodal emotion input (speech, facial expression): the same reward structure could learn to spend more reasoning on cues that are easy to miss in text."],"forward_implications":["Fixed-length chain-of-thought is suboptimal for emotion: simple tasks get over-reasoned and complex tasks get under-reasoned, so adaptive length should be the default setting.","An 8-billion-parameter model can reach or beat much larger closed models on sentiment and sarcasm, meaning scale is not the only path to strong emotion understanding.","Reasoning length becomes a trainable efficiency knob, with median reductions of 52-83% relative to a fixed long-reasoning baseline while maintaining accuracy.","The model's learned length distribution reflects task difficulty: sentiment stays short, sarcasm and humor grow long, matching the paper's simple-versus-complex split."],"supporting_citations":[{"why":"Supplies the teacher model whose reasoning traces are distilled with rejection sampling in Stage 1.","marker":"[11]"},{"why":"Serves as the fixed long-reasoning baseline used for the 52-83% length reduction comparisons.","marker":"[24]"},{"why":"Establishes chain-of-thought prompting, the base paradigm the paper extends to adaptive length.","marker":"[36]"},{"why":"Defines Tree-of-Thoughts, one of the non-linear reasoning structures used as a prompt strategy in distillation.","marker":"[40]"},{"why":"Defines Graph-of-Thoughts, the other non-linear reasoning structure used to diversify distilled traces.","marker":"[2]"},{"why":"Provides the MELD dataset used for sentiment classification and emotion recognition evaluation.","marker":"[27]"},{"why":"Provides the Sarcasm Corpus V2 used for the sarcasm detection benchmark.","marker":"[25]"},{"why":"Provides the Reddit humor detection benchmark used for humor understanding.","marker":"[37]"},{"why":"Motivates treating sarcasm detection as a step-by-step reasoning task, grounding the complex-task distinction.","marker":"[39]"}],"fun_headline_variants":["Adaptive reasoning lengths boost emotion F1 by 27%","Emotion-o1: Thinking less on easy tasks, more on hard ones","Task-aware CoT lifts sarcasm F1 27%, cuts reasoning 83%","Smarter reasoning length improves emotion understanding","Adaptive chain-of-thought beats fixed CoT on emotion tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper treats a distilled reasoning trace as valid training data whenever it ends at the correct label, so a fluent but internally flawed chain of thought can be baked into the student during fine-tuning and reinforcement learning.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive reasoning lengths boost emotion F1 by 27%","Emotion-o1: Thinking less on easy tasks, more on hard ones","Task-aware CoT lifts sarcasm F1 27%, cuts reasoning 83%","Smarter reasoning length improves emotion understanding","Adaptive chain-of-thought beats fixed CoT on emotion tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1312,"prompt_tokens":1029,"completion_tokens":283,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":193}},"tokens_in":645,"tokens_out":283,"duration_ms":3580,"temperature":1.0,"reasoning_tokens":193,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:04:04.281442+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the distilled training set and corrupt the intermediate steps of every trace that ends in the correct label—shuffle the steps or replace them with generic filler—then retrain the same pipeline. If final-label accuracy stays roughly unchanged, the model's gains come from the label-conditioned distribution of the traces rather than from the reasoning content, and the adaptive-reasoning interpretation of the result fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the teacher model whose reasoning traces are distilled with rejection sampling in Stage 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the fixed long-reasoning baseline used for the 52-83% length reduction comparisons."},{"cited_title":"Chain-of-thought prompting elicits reasoning in large language models.Advances in neural informa- tion processing systems, 35:24824–24837, 2022","cited_arxiv_id":null,"evidence_quote":"Establishes chain-of-thought prompting, the base paradigm the paper extends to adaptive length."},{"cited_title":"Graph of thoughts: Solving elaborate problems with large language models","cited_arxiv_id":null,"evidence_quote":"Defines Graph-of-Thoughts, the other non-linear reasoning structure used to diversify distilled traces."},{"cited_title":"Humor detection: A transformer gets the last laugh.”Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing”, November 2019","cited_arxiv_id":null,"evidence_quote":"Provides the Reddit humor detection benchmark used for humor understanding."},{"cited_title":"Is sarcasm detection a step-by-step reasoning process in large language models? InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 25651–25659, 2025","cited_arxiv_id":null,"evidence_quote":"Motivates treating sarcasm detection as a step-by-step reasoning task, grounding the complex-task distinction."}],"review_version":1}