Emotion-o1: Adaptive Long Reasoning for Emotion Understanding in LLMs

Changhao Song, Yazhou Zhang, Hui Gao, Kaiyun Huang, Peng ZhangSubmission to ICLR 2026 · Withdrawn

TL;DR

Current reasoning models either "overthink" simple tasks or reason too shallowly on complex ones. Emotion-o1 introduces an adaptive chain-of-thought (CoT) framework that tailors reasoning depth and structure to the difficulty of emotion understanding tasks. By combining structured distillation from DeepSeek-R1 with a multi-objective reinforcement learning reward, an 8B model matches or outperforms frontier LLMs like Grok-3 and OpenAI-o1 on affective benchmarks while slashing reasoning tokens by up to 83%.

What Problem Does This Paper Address?

Large reasoning models (LRMs) like OpenAI-o1 and DeepSeek-R1 rely on scaling test-time reasoning tokens to solve complex tasks. However, in affective computing, fixed-length CoT is a mismatch:

  • Simple tasks (e.g., binary sentiment analysis) trigger excessive verbosity, wasting compute and introducing noisy overthinking.
  • Pragmatic tasks (e.g., sarcasm and humor detection) require decomposing literal surface cues against contextual incongruities, but standard models often default to shallow, linear reasoning that misses nuanced subtext.

The paper aims to make test-time reasoning compute-adaptive: generating concise, direct rationales for straightforward emotion tasks, and reserved, multi-step, self-reflective reasoning paths for cognitively demanding pragmatic phenomena.

Methods

Emotion-o1 builds on an open-weight base model (LLaMA-3.1-8B) through a three-stage training pipeline:

  1. Structured Emotion Reasoning Distillation: The authors prompt DeepSeek-R1 to generate diverse reasoning paths varying across two dimensions: length (short vs. long) and structure (linear step-by-step vs. non-linear branching like Tree-of-Thought). Using rejection sampling, only rationales arriving at the ground-truth label are retained.

  2. Adaptive CoT-Augmented SFT: The model is fine-tuned to condition on a task identifier $\tau$, training it to associate simple tasks with shorter, direct CoTs and complex tasks with deeper reasoning.

  3. Verifiable Reward Reinforcement Learning (PPO): The policy is refined via PPO using a composite four-part reward function:

  4. Accuracy Reward ($r_{\text{acc}}$): Rewards ground-truth label matching and penalizes missing outputs.

  5. Task-Aware Variable-Length Reward ($r_{\text{length}}$): Evaluates response token counts against task-specific empirical and statistical length targets, applying a quadratic penalty if too short, an exponential penalty if too long, and a Gaussian reward near the optimal length.
  6. Reasoning Structure Reward ($r_{\text{struct}}$): Incentivizes explicit reasoning behaviors (decomposition, reflection, verification) and logical connectives for complex tasks.
  7. Repetition Penalty ($r_{\text{repeat}}$): Discourages redundant circular phrasing using combined Jaccard word-overlap and BERT semantic cosine similarity.

Datasets

The authors evaluate four emotion-related tasks mapped to standard benchmarks:

  • MELD: Used for both 3-class sentiment classification (positive/negative/neutral) and 7-class emotion recognition.
  • Sarcasm Corpus V2: Used for binary sarcasm detection.
  • Reddit Humor Detection: Used for binary humor classification.

From these, the authors construct a distilled reasoning corpus totaling ~87.7k examples (~21k–23k per task). The distribution maintains an ~7:3 linear-to-non-linear ratio across tasks, while setting long-to-short CoT ratios to ~1:1 for sentiment/emotion and ~8:2 for sarcasm/humor.

Key Insights and Ideas

  • Theoretical Justification of Task Difficulty: Sarcasm can be formalized as an XOR function between literal sentiment $S(x)$ and contextual stance $C(x, c)$. Missing context creates a strictly positive irreducible Bayes error. Furthermore, its VC-dimension bounds show it is at least as hard as the harder subtask, theoretically proving why sarcasm demands more representational capacity than sentiment classification.
  • More Reasoning Is Not Always Better: On sentiment and emotion tasks, long reasoning degraded performance. Short linear chains outperformed long CoT by ~5–7% in accuracy, proving that reducing computational redundancy directly prevents classification noise.
  • Pragmatics Demand Structural Depth: Sarcasm and humor required long CoTs with reflection and verification steps to detect incongruity, yielding up to an 11% F1 boost over short CoT.

Key Results

  • Massive Backbone Gains: Compared to the LLaMA-3.1-8B baseline, Emotion-o1 improved accuracy by 7% (Sentiment), 7% (Emotion), 14% (Humor), and 21% (Sarcasm), with F1 gains reaching up to 27%.
  • Competitiveness with Frontier LLMs:
  • Sentiment: Emotion-o1 achieved 0.6770 Acc / 0.6548 Macro-F1, outperforming Grok-3 (0.6559 Acc) and Claude-3.7 (0.6483 Acc).
  • Emotion: Matched OpenAI-o1 within 1% accuracy (0.6352 vs. 0.6398).
  • Sarcasm: Achieved 0.7469 Acc, outperforming all Short-CoT models and trailing GPT-4o by only ~1%.
  • Token Efficiency: Compared to OpenAI-o1, Emotion-o1 reduced average reasoning token length by 73% on sentiment, 52% on emotion, 83% on sarcasm, and 70% on humor.
  • Model Scaling: In scale analysis (LLaMA 1B, 3B, 8B), scaling from 1B to 3B provided the largest jump in F1 (up to 35%), demonstrating that small language models (SLMs) can master adaptive emotional reasoning at low cost.

Limitations and Future Work

  • Task Scope: Evaluation is limited to four curated text classification benchmarks, which may not capture the broader nuances of open-domain affective dialogue.
  • Text-Only Modality: Human emotion expression depends heavily on acoustic pitch, facial cues, and body language. The framework currently excludes multimodal signals.
  • Future Directions: Extending adaptive test-time reasoning to multimodal affective models and evaluating open-ended emotion generation tasks.

Summary updated

Abstract

Long chain-of-thought (CoT) reasoning has shown great promise in enhancing the emotion understanding performance of large language models (LLMs). However, current fixed-length CoT methods struggle to balance reasoning depth and efficiency. Simple tasks (e.g., sentiment classification) are over-reasoned, while complex tasks (e.g., sarcasm understanding) lack depth. To fill this gap, we present Emotion-o1, an adaptive CoT framework that dynamically adjusts reasoning length based on task complexity. Emotion-o1 is trained by distilling adaptive CoT patterns from a large reasoning model (LRM), followed by supervised fine-tuning and reinforcement learning with a four-part reward targeting accuracy, brevity, structure, and redundancy. Experimental results on four emotion tasks highlight: (1) Emotion-o1 demonstrates significant improvements over its backbone, with F1 score increases of 11\%↑(Sentiment), 14\%↑(Emotion), 18\%↑(Humor), and 27\%↑(Sarcasm). (2) In sentiment and emotion tasks, our 8B model demonstrates superior performance against SoTA LLMs, outperforming Grok‑3 by 2.1\% in sentiment and within 1\% of OpenAI‑o1 in emotion. (3) The framework maintains accuracy while reducing reasoning length by 83\% compared to OpenAI-o1, demonstrating effective precision-efficiency optimization. From a lower-cost perspective, the framework also empowers SLMs to achieve reasoning capabilities comparable to larger ones.
BibTeX
@misc{song2025emotion,
  author = {Changhao Song and Yazhou Zhang and Hui Gao and Kaiyun Huang and Peng Zhang},
  title = {{Emotion-o1: Adaptive Long Reasoning for Emotion Understanding in LLMs}},
  year = {2025},
  eprint = {2505.22548},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.22548},
}

Loading related content...

Figure