arxiv: 2604.00013 · v2 · submitted 2026-03-10 · 💻 cs.CL · cs.AI

Recognition: 2 theorem links

· Lean Theorem

C2F-Thinker: Coarse-to-Fine Reasoning with Hint-Guided Reinforcement Learning for Multimodal Sentiment Analysis

Miaosen Luo , Zhenhao Yang , Jieshen Long , Jinghu Sun , Yichu Liu , Sijie Mai

Authors on Pith no claims yet

Pith reviewed 2026-05-15 13:48 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords multimodal sentiment analysischain-of-thought reasoningreinforcement learningcoarse-to-fine reasoningcross-domain generalizationhint-guided optimizationmultimodal large language models

0 comments

The pith

C2F-Thinker combines distilled coarse-to-fine CoT reasoning with hint-guided RL to deliver competitive fine-grained multimodal sentiment performance and stronger cross-domain generalization.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that multimodal sentiment models can become both more interpretable and more robust by training in two explicit stages. First, supervised fine-tuning on chain-of-thought traces distilled from a larger teacher teaches the model a fixed sequence of polarity judgment, intermediate analysis, and fine-grained scoring. Second, a modified reinforcement-learning stage called hint-guided GRPO injects the correct polarity label as an early hint during sampling; this steers the model away from cascading errors on difficult examples while a multi-part reward keeps outputs formatted and accurate. A sympathetic reader would expect this structured path to reduce the usual brittleness of black-box MLLMs on unseen domains, because early polarity guidance shrinks the search space without requiring new manual labels.

Core claim

C2F-Thinker harmonizes coarse-to-fine structured reasoning with hint-guided reinforcement learning through a two-stage progressive training pipeline: cold-start supervised fine-tuning on high-quality CoT data distilled from a larger teacher model equips the base model with a three-phase emotional reasoning paradigm (polarity judgment, intermediate analysis, fine-grained scoring), after which hint-guided Group Relative Policy Optimization injects correct initial polarity predictions during sampling to mitigate cascading errors, improve hard-sample utilization, and refine predictions via a multi-faceted reward that balances classification, regression, and formatting constraints.

What carries the argument

Hint-guided Group Relative Policy Optimization (GRPO), which injects correct polarity predictions as sampling hints to steer the model toward valid reasoning paths while preserving the three-phase CoT structure.

If this is right

The model retains human-readable intermediate steps while matching or exceeding black-box baselines on fine-grained regression.
Cross-domain gains imply the framework can be deployed in new domains with less retraining data.
The same two-stage pipeline could be reused for other regression-style multimodal tasks that suffer from sparse rewards.
Multi-faceted rewards that include formatting constraints keep outputs structured without extra post-processing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The hint mechanism might transfer to any sequential reasoning task where early classification errors propagate, such as visual question answering or medical report generation.
If teacher-model quality varies across domains, the cold-start stage could become the new performance bottleneck rather than the RL stage.
The approach suggests a broader pattern: use cheap teacher distillation to install a reasoning skeleton, then use targeted hints inside RL to protect that skeleton on hard cases.

Load-bearing premise

High-quality chain-of-thought data distilled from a larger teacher supplies an effective structured emotional reasoning sequence, and injecting correct polarity predictions as hints during RL sampling reliably reduces cascading errors without adding new biases or limiting exploration.

What would settle it

An ablation that removes the polarity-hint injection from the GRPO stage and shows no gain (or a loss) in cross-domain regression accuracy relative to standard RL baselines on the same test sets.

Figures

Figures reproduced from arXiv: 2604.00013 by Jieshen Long, Jinghu Sun, Miaosen Luo, Sijie Mai, Yichu Liu, Zhenhao Yang.

**Figure 2.** Figure 2: The overall training pipeline for the proposed C2F-Thinker. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Comparison of reward curves during training for different configurations. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Illustration of C2T-Thinker reasoning. phenomenon suggests that relying solely on easy samples for optimization, while avoiding the risk of reward sparsity from challenging samples, prevents the model from learning the ability to distinguish complex sentiment boundaries from more difficult instances, thereby limiting its generalization performance. 3) Training Dynamics Analysis: To more intuitively demonst… view at source ↗

read the original abstract

Multimodal sentiment analysis aims to integrate textual, acoustic, and visual information for deep emotional understanding. Despite the progress of multimodal large language models (MLLMs) via supervised fine-tuning, their "black-box" nature hinders interpretability. While Chain-of-Thought (CoT) reasoning offers a potential remedy, it is constrained by high manual annotation costs and the inherent challenges of reinforcement learning (RL), such as reward sparsity and low exploration efficiency on hard samples. This paper presents C2F-Thinker, a framework that harmonizes coarse-to-fine structured reasoning with hint-guided RL through a two-stage progressive training pipeline. In the first stage, we conduct cold-start supervised fine-tuning using high-quality CoT data distilled from a larger teacher model, consisting of three distinct phases: polarity judgment, intermediate analysis, and fine-grained scoring. This equips the base model with a structured emotional reasoning paradigm. In the second stage, we introduce a hint-guided Group Relative Policy Optimization (GRPO) algorithm. By injecting correct initial polarity predictions as hints during the sampling process, the model is guided toward accurate reasoning paths, effectively mitigating cascading errors and enhancing the utilization of hard samples. Furthermore, a multi-faceted reward function incorporating classification, regression, and formatting constraints is designed to refine prediction accuracy while preserving interpretability. Experimental results demonstrate that C2F-Thinker achieves competitive performance on fine-grained sentiment regression tasks while significantly outperforming baselines in cross-domain generalization. This highlights its potential in building trustworthy and robust sentiment analysis systems for real-world applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

C2F-Thinker adds a three-phase CoT pipeline and hint-guided GRPO to multimodal sentiment work, but the hints during training likely drive the cross-domain claims more than independent reasoning.

read the letter

C2F-Thinker uses a two-stage setup: first cold-start supervised fine-tuning on distilled CoT data broken into polarity judgment, intermediate analysis, and fine-grained scoring, then hint-guided GRPO that feeds correct polarity predictions as hints while sampling to steer away from bad paths on hard examples. A multi-faceted reward mixes classification, regression, and formatting terms to keep outputs readable while pushing accuracy. This is a direct attempt to fix reward sparsity and low exploration in RL for sentiment without needing full manual CoT labels. The structure gives the model an explicit emotional reasoning sequence that standard SFT on MLLMs often lacks, and the hint step targets cascading errors in a practical way. That combination is the actual new piece within the subfield. The soft spot sits in the hint mechanism itself. Because training supplies ground-truth polarity labels as guidance, it is hard to tell whether the model learns the full coarse-to-fine path or simply follows the provided starting signal. Cross-domain gains could shrink or disappear once hints are removed at inference, and the abstract gives no numbers, ablations, or error bars to check this. If the full paper lacks those controls, the generalization story rests on weaker ground than claimed. The work is aimed at people building interpretable multimodal sentiment systems or applying RL to affective computing tasks. Readers already working on CoT distillation or policy optimization in this area will see concrete implementation choices worth examining. I would send it for peer review. The pipeline is specific enough that referees can test the hint dependence and ask for the missing quantitative checks without starting from scratch.

Referee Report

2 major / 2 minor

Summary. The paper introduces C2F-Thinker, a two-stage framework for multimodal sentiment analysis. The first stage performs cold-start supervised fine-tuning on high-quality coarse-to-fine Chain-of-Thought data distilled from a larger teacher model, structured into polarity judgment, intermediate analysis, and fine-grained scoring phases. The second stage applies hint-guided Group Relative Policy Optimization (GRPO), injecting correct initial polarity predictions as hints during sampling to mitigate cascading errors on hard samples, paired with a multi-faceted reward incorporating classification, regression, and formatting terms. The authors claim that this yields competitive performance on fine-grained sentiment regression tasks and significantly better cross-domain generalization than baselines.

Significance. If the empirical claims hold after addressing the training-inference mismatch, the work could contribute to more interpretable and robust multimodal sentiment systems by combining structured CoT reasoning with RL techniques that address reward sparsity. The progressive pipeline and hint-guided sampling represent a practical approach to improving exploration on difficult examples, with potential applicability beyond sentiment analysis to other multimodal reasoning tasks.

major comments (2)

[second stage / hint-guided GRPO description] The second-stage hint-guided GRPO explicitly injects ground-truth polarity labels as hints during the sampling process to guide reasoning paths. This mechanism is load-bearing for the cross-domain generalization claim, yet the manuscript does not clarify whether the model internalizes the coarse-to-fine structure or primarily learns to follow external oracle hints. At inference, no such hints are available, so the reported gains may reflect supervised guidance rather than learned reasoning; an ablation removing hints at test time or comparing against a no-hint GRPO baseline is required to substantiate the claim.
[reward function definition] The multi-faceted reward (classification + regression + formatting) is presented as balancing accuracy and interpretability, but no quantitative breakdown is given showing the contribution of each term or how the reward weights were chosen. Without this, it is unclear whether performance improvements stem from the hint mechanism, the reward design, or the stage-1 CoT initialization, weakening the attribution of gains to the proposed framework.

minor comments (2)

[Abstract] The abstract summarizes results only qualitatively (e.g., 'competitive performance' and 'significantly outperforming') without any numerical values, dataset names, or error bars; the full experimental section should include these to allow immediate assessment of effect sizes.
[Method / GRPO] The GRPO algorithm is introduced without an explicit equation or pseudocode for the policy update or the hint-injection step; adding a formal description would improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and insightful comments on our manuscript. We address each major comment point by point below and will revise the paper to incorporate the suggested clarifications and experiments.

read point-by-point responses

Referee: [second stage / hint-guided GRPO description] The second-stage hint-guided GRPO explicitly injects ground-truth polarity labels as hints during the sampling process to guide reasoning paths. This mechanism is load-bearing for the cross-domain generalization claim, yet the manuscript does not clarify whether the model internalizes the coarse-to-fine structure or primarily learns to follow external oracle hints. At inference, no such hints are available, so the reported gains may reflect supervised guidance rather than learned reasoning; an ablation removing hints at test time or comparing against a no-hint GRPO baseline is required to substantiate the claim.

Authors: We appreciate the referee's concern regarding the role of hints in the GRPO stage. The hints (correct initial polarity predictions) are provided exclusively during training to improve exploration efficiency and reduce cascading errors on hard samples, allowing the policy to learn more effective coarse-to-fine reasoning trajectories. At inference, the model generates the full reasoning chain without any external hints. To directly address whether the model internalizes the structure, we will add two ablations in the revised manuscript: (1) a no-hint GRPO baseline trained without polarity hints for comparison, and (2) an evaluation of the trained model under test-time hint removal or corruption to measure robustness. These experiments will clarify that performance gains arise from the learned reasoning policy rather than reliance on oracle guidance during training. revision: yes
Referee: [reward function definition] The multi-faceted reward (classification + regression + formatting) is presented as balancing accuracy and interpretability, but no quantitative breakdown is given showing the contribution of each term or how the reward weights were chosen. Without this, it is unclear whether performance improvements stem from the hint mechanism, the reward design, or the stage-1 CoT initialization, weakening the attribution of gains to the proposed framework.

Authors: We agree that a quantitative breakdown of the reward components would strengthen attribution of results. In the revised manuscript, we will include an ablation study quantifying the individual contributions of the classification, regression, and formatting reward terms to overall performance on both in-domain and cross-domain tasks. We will also detail the weight selection process, which was based on grid search over a held-out validation set to balance predictive accuracy against reasoning quality and format compliance. This will help isolate the effects of the reward design from the hint mechanism and stage-1 initialization. revision: yes

Circularity Check

0 steps flagged

No circularity in empirical two-stage training framework

full rationale

The paper describes C2F-Thinker as an empirical method consisting of cold-start SFT on teacher-distilled CoT data (polarity judgment, intermediate analysis, fine-grained scoring) followed by hint-guided GRPO with a multi-faceted reward. No equations, derivations, or predictions are presented that reduce by construction to the paper's own inputs or fitted parameters. All performance claims rest on external baseline comparisons rather than self-referential definitions, uniqueness theorems, or renamed known results. The framework is self-contained against external benchmarks with no load-bearing self-citation chains.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Abstract-only view provides no explicit fitted parameters; the framework rests on standard machine-learning assumptions about the value of structured CoT and guided RL sampling, with the hint mechanism presented as a novel but unverified addition.

invented entities (1)

hint-guided GRPO no independent evidence
purpose: To guide sampling toward accurate reasoning paths and mitigate cascading errors on hard samples during reinforcement learning
Presented as a new algorithmic variant in the second training stage; no independent evidence outside the paper is supplied.

pith-pipeline@v0.9.0 · 5606 in / 1326 out tokens · 51718 ms · 2026-05-15T13:48:21.750217+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

two-stage progressive training pipeline... hint-guided Group Relative Policy Optimization (GRPO)... injecting correct initial polarity predictions as hints
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

coarse-to-fine structured reasoning... polarity judgment, intermediate analysis, and fine-grained scoring

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 9 internal anchors

[1]

A comprehensive survey on multimodal sentiment analysis: Techniques, models, and applications,

H. Zhang, “A comprehensive survey on multimodal sentiment analysis: Techniques, models, and applications,”Advances in Engineering Inno- vation, vol. 12, pp. 47–52, 2024

work page 2024
[2]

Multimodal large language models for end-to-end affective computing: Benchmarking and boosting with generative knowledge prompting,

M. Luo, J. Long, Z. Li, Y . Yang, Y . Jiang, and S. Mai, “Multimodal large language models for end-to-end affective computing: Benchmarking and boosting with generative knowledge prompting,”arXiv preprint arXiv:2508.02429, 2025

work page arXiv 2025
[3]

Eemo-bench: a benchmark for multi-modal large language models on image evoked emotion assessment,

L. Gao, Z. Jia, Y . Zeng, W. Sun, Y . Zhang, W. Zhou, G. Zhai, and X. Min, “Eemo-bench: a benchmark for multi-modal large language models on image evoked emotion assessment,” inProceedings of the 33rd ACM International Conference on Multimedia, 2025, pp. 7064– 7073

work page 2025
[4]

Can large language models help multimodal language analy- sis? mmla: A comprehensive benchmark.arXiv preprint arXiv:2504.16427, 2025

H. Zhang, Z. Li, Y . Zhu, H. Xu, P. Wang, H. Zhu, J. Zhou, and J. Zhang, “Can large language models help multimodal language analysis? mmla: A comprehensive benchmark,”arXiv preprint arXiv:2504.16427, 2025

work page arXiv 2025
[5]

Emotionqueen: A bench- mark for evaluating empathy of large language models,

Y . Chen, S. Yan, S. Liu, Y . Li, and Y . Xiao, “Emotionqueen: A bench- mark for evaluating empathy of large language models,” inFindings of the Association for Computational Linguistics: ACL 2024, 2024, pp. 2149–2176

work page 2024
[6]

Exploring large language models for multimodal sentiment analysis: Challenges, benchmarks, and future directions,

S. Song, “Exploring large language models for multimodal sentiment analysis: Challenges, benchmarks, and future directions,”arXiv preprint arXiv:2411.15408, 2024

work page arXiv 2024
[7]

Evaluating signifi- cant features in context-aware multimodal emotion recognition with xai methods,

A. Khalane, R. Makwana, T. Shaikh, and A. Ullah, “Evaluating signifi- cant features in context-aware multimodal emotion recognition with xai methods,”Expert Systems, vol. 42, no. 1, p. e13403, 2025

work page 2025
[8]

Multimodal Chain-of-Thought Reasoning: A Comprehensive Survey

Y . Wang, S. Wu, Y . Zhang, S. Yan, Z. Liu, J. Luo, and H. Fei, “Mul- timodal chain-of-thought reasoning: A comprehensive survey,”arXiv preprint arXiv:2503.12605, 2025

work page internal anchor Pith review arXiv 2025
[9]

Multimodal Chain-of-Thought Reasoning in Language Models

Z. Zhang, A. Zhang, M. Li, H. Zhao, G. Karypis, and A. Smola, “Multimodal chain-of-thought reasoning in language models,”arXiv preprint arXiv:2302.00923, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[10]

Chain-of-thought prompting elicits reasoning in large language models,

J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhouet al., “Chain-of-thought prompting elicits reasoning in large language models,”Advances in neural information processing systems, vol. 35, pp. 24 824–24 837, 2022

work page 2022
[11]

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . Li, Y . Wuet al., “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,”arXiv preprint arXiv:2402.03300, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[12]

Boosting mllm reasoning with text-debiased hint-grpo,

Q. Huang, W. Dai, J. Liu, W. He, H. Jiang, M. Song, J. Chen, C. Yao, and J. Song, “Boosting mllm reasoning with text-debiased hint-grpo,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 4848–4857

work page 2025
[13]

Qwen3-Omni Technical Report

J. Xu, Z. Guo, H. Hu, Y . Chu, X. Wang, J. He, Y . Wang, X. Shi, T. He, X. Zhuet al., “Qwen3-omni technical report,”arXiv preprint arXiv:2509.17765, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[14]

Qwen2.5-Omni Technical Report

J. Xu, Z. Guo, J. He, H. Hu, T. He, S. Bai, K. Chen, J. Wang, Y . Fan, K. Dang, B. Zhang, X. Wang, Y . Chu, and J. Lin, “Qwen2.5-omni technical report,”arXiv preprint arXiv:2503.20215, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[15]

Multimodal large language models meet multimodal emotion recognition and reasoning: A survey,

Y . Shou, T. Meng, W. Ai, and K. Li, “Multimodal large language models meet multimodal emotion recognition and reasoning: A survey,”arXiv preprint arXiv:2509.24322, 2025

work page arXiv 2025
[16]

Sentiment analysis meets explainable artificial intelligence: A survey on explainable sentiment analysis,

A. Diwali, K. Saeedi, K. Dashtipour, M. Gogate, E. Cambria, and A. Hussain, “Sentiment analysis meets explainable artificial intelligence: A survey on explainable sentiment analysis,”IEEE Transactions on Affective Computing, vol. 15, no. 3, pp. 837–846, 2023

work page 2023
[17]

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Biet al., “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[18]

Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs

X. Wen, Z. Liu, S. Zheng, S. Ye, Z. Wu, Y . Wang, Z. Xu, X. Liang, J. Li, Z. Miaoet al., “Reinforcement learning with verifiable rewards implicitly incentivizes correct reasoning in base llms,”arXiv preprint arXiv:2506.14245, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[19]

Gencls++: Pushing the boundaries of generative classification in llms through comprehensive sft and rl studies across diverse datasets,

M. He, F. Zhao, C. Lu, Z. Liu, Y . Wang, and H. Qian, “Gencls++: Pushing the boundaries of generative classification in llms through comprehensive sft and rl studies across diverse datasets,”arXiv preprint arXiv:2504.19898, 2025

work page arXiv 2025
[20]

Think or not think: A study of explicit thinking in rule-based visual reinforcement fine-tuning.arXiv preprint arXiv:2503.16188, 2025

M. Li, S. Zhao, J. Zhong, Y . Lai, and K. Zhang, “Cls-rl: Image classification with rule-based reinforcement learning,”arXiv preprint arXiv:2503.16188, vol. 3, 2025

work page arXiv 2025
[21]

Affectgpt-r1: Leveraging reinforcement learning for open-vocabulary multimodal emotion recognition,

Z. Lian, F. Zhang, Y . Zhang, J. Tao, R. Liu, H. Chen, and X. Li, “Affectgpt-r1: Leveraging reinforcement learning for open-vocabulary multimodal emotion recognition,”arXiv preprint arXiv:2508.01318, 2025

work page arXiv 2025
[22]

Multimodal senti- ment intensity analysis in videos: Facial gestures and verbal messages,

A. Zadeh, R. Zellers, E. Pincus, and L.-P. Morency, “Multimodal senti- ment intensity analysis in videos: Facial gestures and verbal messages,” IEEE Intelligent Systems, vol. 31, no. 6, pp. 82–88, 2016

work page 2016
[23]

Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph,

A. B. Zadeh, P. P. Liang, S. Poria, E. Cambria, and L.-P. Morency, “Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph,” inProceedings of the 56th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), 2018, pp. 2236–2246

work page 2018
[24]

Ch- sims: A chinese multimodal sentiment analysis dataset with fine-grained annotation of modality,

W. Yu, H. Xu, F. Meng, Y . Zhu, Y . Ma, J. Wu, J. Zou, and K. Yang, “Ch- sims: A chinese multimodal sentiment analysis dataset with fine-grained annotation of modality,” inProceedings of the 58th annual meeting of the association for computational linguistics, 2020, pp. 3718–3727

work page 2020
[25]

Make acoustic and visual cues matter: Ch-sims v2. 0 dataset and av-mixup consistent module,

Y . Liu, Z. Yuan, H. Mao, Z. Liang, W. Yang, Y . Qiu, T. Cheng, X. Li, H. Xu, and K. Gao, “Make acoustic and visual cues matter: Ch-sims v2. 0 dataset and av-mixup consistent module,” inProceedings of the 2022 international conference on multimodal interaction, 2022, pp. 247–258

work page 2022
[26]

Pandagpt: One model to instruction-follow them all,

Y . Su, T. Lan, H. Li, J. Xu, Y . Wang, and D. Cai, “Pandagpt: One model to instruction-follow them all,”arXiv preprint arXiv:2305.16355, 2023

work page arXiv 2023
[27]

Emotion-llama: Multimodal emotion recognition and reasoning with instruction tuning,

Z. Cheng, Z.-Q. Cheng, J.-Y . He, K. Wang, Y . Lin, Z. Lian, X. Peng, and A. Hauptmann, “Emotion-llama: Multimodal emotion recognition and reasoning with instruction tuning,”Advances in Neural Information Processing Systems, vol. 37, pp. 110 805–110 853, 2024

work page 2024
[28]

MiniCPM-V: A GPT-4V Level MLLM on Your Phone

Y . Yao, T. Yu, A. Zhang, C. Wang, J. Cui, H. Zhu, T. Cai, H. Li, W. Zhao, Z. Heet al., “Minicpm-v: A gpt-4v level mllm on your phone,”arXiv preprint arXiv:2408.01800, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[29]

Ola: Pushing the frontiers of omni-modal language model with progressive modality alignment,

Z. Liu, Y . Dong, J. Wang, Z. Liu, W. Hu, J. Lu, and Y . Rao, “Ola: Pushing the frontiers of omni-modal language model with progressive modality alignment,”arXiv e-prints, pp. arXiv–2502, 2025

work page 2025
[30]

Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding

H. Zhang, X. Li, and L. Bing, “Video-llama: An instruction-tuned audio-visual language model for video understanding,”arXiv preprint arXiv:2306.02858, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[31]

Humanomni: A large vision-speech language model for human-centric video understanding,

J. Zhao, Q. Yang, Y . Peng, D. Bai, S. Yao, B. Sun, X. Chen, S. Fu, X. Wei, L. Boet al., “Humanomni: A large vision-speech language model for human-centric video understanding,”arXiv preprint arXiv:2501.15111, 2025

work page arXiv 2025