CIMR: Contextualized Iterative Multimodal Reasoning for Robust Instruction Following in LVLMs

Christian Ng; Heng Chen; Kexin Qiu; Xinyi Jiang; Yangshu Yuan

REVIEW 4 major objections 4 minor 27 references

CIMR: Contextualized Iterative Multimodal Reasoning for Robust Instruction Following in LVLMs

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CIMR reports that a self-correcting iterative loop lifts vision-language instruction following to 91.5% on its new MAP benchmark, ahead of GPT-4V and the base LLaVA-1.5.

desk verdict CIMR is a readable description of an iterative self-correction wrapper for LVLMs, but its headline 91.5% accuracy is built on an unreleased synthetic benchmark that mirrors the training template, and the paper is not ready for peer review without a major empirical rebuild. read the letter →

arxiv 2507.22074 v1 pith:QDO56K6W submitted 2025-07-22 cs.LG cs.CL

Yangshu Yuan , Heng Chen , Xinyi Jiang , Christian Ng , Kexin Qiu This is my paper

classification cs.LGcs.CL

keywords largevision-languagemodelsiterativereasoningself-correctioninstructionfollowingmulti-modalactionplanningdynamicfusioncontext-aware

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a large vision-language model can follow complex, multi-step instructions much more reliably when it is allowed to check its own work and revise: generate an initial response, parse visual and textual feedback, update its running context, and generate a corrected response. Concretely, it reports that its two-stage system CIMR reaches 91.5% task completion on a new synthetic benchmark called MAP, beating GPT-4V (89.2%) and the base model LLaVA-1.5-7B (78.5%) it is built on. If correct, this is a way to make instruction-following models more robust on tasks like arranging objects or correcting placement without needing a larger base model.

What carries the argument

The carrier of the argument is a closed-loop refinement cycle wrapped around a base vision-language model. At each iteration $t$, the model updates its task context as $C_t = U(C_{t-1}, R_{t-1}, S_t)$, parses multi-modal feedback as $S_t = P(R_{t-1}, F_{V'}, F_{C_t}, F_T)$, and produces a refined response $R_t = R(R_{t-1}, S_t, F_{C_t})$. A multi-head cross-attention module fuses textual, visual, and contextual features at every step, so the feedback signal is grounded in the original instruction as well as the latest observation. The paper's argument is that this cycle converts approximate first attempts into corrected final answers and that its convergence after two or three iterations is what the MAP accuracy table shows.

What would settle it

Take an independently constructed set of multi-step visual instruction tasks that are not generated from 'initial erroneous response - corrective feedback - final correct response' triples, for example logged episodes of a real robot fixing misplacements or human-written planning prompts, run CIMR and a single-pass LLaVA-1.5 baseline on it, and compare task completion. If the gap drops to near zero, the claim that iterative self-correction with dynamic context is the source of the MAP improvement is refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the loop, not the base model, is what pays off. CIMR's first pass is just LLaVA-1.5-7B and scores 78.5%, identical to the baseline, while the full loop reaches 91.5%. The ablation numbers make the same point in reverse: removing the self-correction phase drops accuracy to 79.1%, and removing only the dynamic context update drops it to 84.7%. The paper also reports human evaluation on 100 MAP scenarios, with CIMR scoring 4.6 out of 5 for overall task quality and 92.1% error recovery, versus 4.3 out of 5 and 88.5% for GPT-4V.

Load-bearing premise

The claim rests on the MAP dataset, which is synthetic and constructed by the same team that designed CIMR and its training triples, so the gap over baselines may partly measure how well the model learns the dataset's own 'wrong answer, feedback, right answer' template rather than general multi-step instruction following.

Editorial extensions

If this is right

On the MAP benchmark, wrapping LLaVA-1.5-7B in the CIMR loop raises task completion from 78.5% to 91.5%, and removing the self-correction phase returns accuracy to 79.1%, so the loop itself is the source of the gain.
Most of the benefit appears early: accuracy goes from 78.5% on iteration 1 to 88.0% after one correction and 91.0% after two, which bounds the extra compute if early stopping is used.
A feedback loop without a running memory of task state is not enough: CIMR without dynamic context update reaches only 84.7%, 6.8 points below the full system.
Human raters agree with the automatic metric, scoring CIMR 4.6 out of 5 for overall task quality and 92.1% error recovery versus 4.3 out of 5 and 88.5% for GPT-4V.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the MAP benchmark was generated by the same pipeline used to construct training triples, the 91.5% figure should be read as a controlled-template result; an independent, human-annotated or real-robot test set would show how much of the gain generalizes.
Even though the paper only tests LLaVA-1.5-7B, the modules are described as generic wrappers around any base vision-language model, so the same loop is a plausible add-on to other single-pass models.
The internal consistency check can run on the instruction and context alone, so a purely textual version of CIMR could be tested on multi-step planning; if it retains the accuracy gain, the mechanism is about reasoning rather than vision.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Mike —

Quick take on arXiv:2507.22074. The framework itself is sensible: take LLaVA-1.5, add an iterative loop that parses multi-modal feedback and updates a context state, and ablate the components. The related work is competently surveyed, the equations (1)–(8) are readable at a high level, and the ablation design (drop self-correction, drop dynamic context) is textbook. There's no new mechanism here — it's an application of the iterative self-correction loop that already exists in QueryAgent, Agent-R, and Embodied Reasoner — but as a systems proposal it's coherent.

The problem is the empirical claim. The MAP dataset is self-constructed, and the scenarios are explicitly built to require iterative refinement. Meanwhile, the model is fine-tuned on corrective-feedback triples of the same format. That is a direct train/test template overlap. The 13-point jump over LLaVA-1.5 (91.5 vs. 78.5) is exactly what you'd predict if the baseline simply never saw that template. There are no error bars, no test-set size, no split details, and no code or data release. The 2.3-point margin over GPT-4V is not statistically grounded. Table IV is also too neat: iteration 1 matches the LLaVA baseline exactly, which is possible but also consistent with the loop just re-prompting the same model on the same distribution.

What's genuinely useful is the intent: the human evaluation, small as it is, is a good instinct, and the paper is honest about the computational overhead. The citation pattern is fine, including some self-citations that are to related work rather than to this method. But none of that supports the state-of-the-art claim.

Bottom line: this is a decent proposal write-up, not a verifiable research result. If the authors release the MAP dataset and code and add an out-of-distribution test, it could become a useful empirical study. As it stands, I wouldn't send it to referees; the central evidence is uncheckable and likely circular. I'd desk-reject in current form.

Referee Report

4 major / 4 minor

Summary. The paper proposes CIMR, a two-stage iterative reasoning and self-correction framework for large vision-language models (LVLMs). CIMR first generates an initial response from fused textual, visual, and contextual features, then iteratively parses multi-modal feedback, updates a dynamical context, and refines the response. The authors fine-tune LLaVA-1.5-7B on a Visual Instruction Tuning dataset and on paired corrective-feedback triples, then evaluate on a newly constructed synthetic MAP dataset, reporting 91.5% task-completion accuracy versus 89.2% for GPT-4V, 78.5% for LLaVA-1.5, 75.3% for MiniGPT-4, and 72.8% for InstructBLIP. They also report an ablation study, a human evaluation, iteration-dynamics results, and qualitative examples. The central claim is that iterative reasoning and self-correction, rather than the base model, produce the reported gains.

Significance. If the central claim were established, the paper would make a useful contribution to multi-modal instruction following by demonstrating that iterative context-aware refinement improves LVLM performance on multi-step tasks. The framework is clearly motivated and the two-stage formulation is easy to follow. However, the empirical evidence as presented is not sufficient to support the claim. The MAP benchmark is author-constructed, the training data and test scenarios are built around the same corrective-feedback template, and no dataset, code, test-set size, error bars, or statistical tests are provided. The equations in Section III describe modules only symbolically, without implementation details or hyperparameters. The reported 13-point gain over the base LLaVA-1.5 model and the 2.3-point margin over GPT-4V could plausibly reflect template overlap and evaluation artifacts rather than a general capability. The paper does provide a straightforward ablation and a human evaluation, which are useful in intent, but both inherit the same benchmark-related limitations. Overall, the contribution is potentially interesting but empirically unverified as presented.

major comments (4)

[IV-A1 and IV-A2] The central 91.5% claim is undermined by train/test circularity. Section IV-A1 says CIMR is fine-tuned on paired data of the form 'initial erroneous response - corrective feedback - final correct response', while Section IV-A2 states that the MAP dataset was constructed by the authors and is 'specifically designed to evaluate the model's ability to engage in iterative reasoning and error recovery.' The MAP scenarios therefore appear to be generated from the same corrective-feedback template used for training. Since no MAP data, train/test split, or template-overlap analysis is released, the reported gain over LLaVA-1.5 (91.5 vs 78.5) and over GPT-4V (91.5 vs 89.2) could largely reflect overlap between the training and test distributions rather than a general instruction-following improvement. The paper needs an out-of-distribution evaluation or at minimum a formal analysis of template overlap to support its attribution of the gains to iterative reasoning.
[III-C2 and IV-G] The feedback loop is not reproducible because the environment that generates multi-modal feedback is never described. Section III-C2 assumes new visual observations I'_visual after executing a response, and Section IV-G's qualitative examples refer to a 'simulated robot's camera' and 'a slightly different viewpoint', but Section IV-A2 provides no simulator, rendering pipeline, number of scenes, or procedure for generating feedback images. Equations (6)-(8) define the context update, feedback parsing, and refinement functions only symbolically, and no implementation of P, R, or U is given. Without this information, neither the MAP benchmark nor the CIMR feedback mechanism can be reconstructed or independently evaluated.
[IV-C and Table I] The quantitative comparison is statistically unsupported. Table I reports single accuracy numbers with no number of test tasks, no confidence intervals, and no error bars, and the text says CIMR 'significantly outperforms' baselines without any significance test. For the 2.3-point margin over GPT-4V (91.5 vs 89.2), the absence of sample-size or variance information makes the comparison unverifiable. The same issue affects Table IV's iteration dynamics, where accuracy values are averaged over test scenarios but the number of scenarios is not stated, and Table III's human evaluation, where no inter-annotator agreement or confidence intervals are reported. The paper should report the MAP dataset size, the number of runs, and appropriate statistical measures before claiming state-of-the-art performance.
[III] The method description is not sufficiently concrete to support the empirical claims. Equations (1)-(8) introduce encoders E_T, E_V, E_C, a CrossAttention fusion mechanism, a context update U, a feedback parser P, and a refinement function R, but the manuscript does not specify how these are implemented, how they interact with the frozen or trainable parameters of LLaVA-1.5, or what training configurations are used. The maximum iteration count and the termination confidence metric are mentioned but never defined, and the number of fine-tuning epochs, learning rates, and optimization details are absent. Since the contribution is a framework whose effectiveness is measured empirically, the lack of implementation detail prevents replication and makes the reported accuracy numbers impossible to audit.

minor comments (4)

[References] Reference [5] is cited as the source of LLaVA-1.5-7B, but the listed paper by Cocchi et al. is about LLaVA-MoRe, not the original LLaVA-1.5; the paper should cite the actual LLaVA-1.5 publication. Reference [6] is cited for the Visual Instruction Tuning (VIT) dataset, but the listed COLING 2025 paper is about synthesizing visual instructions and does not clearly correspond to the standard VIT dataset; please clarify the exact dataset and version used.
[IV-A2] The phrase 'novel, fabricated dataset' should be 'novel, synthetic dataset' to avoid the unintended connotation of fabrication; the text elsewhere uses 'synthetic' and should be consistent.
[IV-F, Table IV] The '4+' iteration row should state whether it means exactly four iterations or all iterations beyond four, and the maximum iteration count used in the experiments should be reported explicitly.
[IV-E] The 'Error Recovery Rate' metric is not defined; the paper should state its denominator, whether it is computed only on tasks where the initial response contained an error, and whether human evaluators had access to the full iterative feedback traces.

Circularity Check

2 steps flagged · score 8.0 of 10

CIMR's 91.5% accuracy is an in-distribution score: the MAP test is built from the same 'erroneous response - corrective feedback - corrected response' template used to train the model, so the claimed improvement is a fit to the authors' synthetic benchmark rather than a validated prediction.

fitted input called prediction [Section IV-A1 (Base Model and Training Strategy); Section IV-C (Quantitative Results, Table I)]
"This involved generating paired data consisting of “initial erroneous response - corrective feedback - final correct response” sequences, specifically designed to teach the model how to identify errors and perform effective self-correction."

The only task-specific training signal for CIMR is the corrective-feedback triple (erroneous response, feedback, corrected response). The headline number is then reported as Task Completion Accuracy on MAP, defined as “successfully executes all required steps and achieves the final specified goal, including successful self-correction when necessary.” Thus the test metric measures exactly the behavior the model was trained to produce, and the MAP scenarios are authored by the same group to require the identical “self-correct if wrong” template. The 91.5% vs. 78.5% gain over LLaVA-1.5 is therefore an in-distribution measurement, not a prediction of a distinct capability.
self definitional [Section I (Introduction); Section IV-A2 (Datasets)]
"To rigorously test the model’s proficiency in complex, multi-step tasks requiring iterative correction, we constructed a novel, fabricated dataset: the Multi-modal Action Planning (MAP) Dataset. This dataset comprises a series of scenarios involving visual observations and multi-turn instructions, explicitly designed to simulate situations where dynamic adjustments and self-correction are imperative (e.g., “adjust item position until specified layout, self-correct if wrong”)."

CIMR’s claimed contribution is iterative reasoning and self-correction; the evaluation dataset is explicitly defined as requiring iterative reasoning and self-correction. The inference from Table II (removing the self-correction phase drops accuracy from 91.5% to 79.1%) is therefore entailed by the benchmark’s construction rather than by an independent test: a single-pass model is scored on scenarios whose definition is “you will need to correct yourself.” The empirical claim reduces to “the model trained and configured to self-correct scores higher on a test built to demand self-correction,” which is true by definition and does not independently validate the mechanism.

full rationale

Neither circular step is a citation-based loop: the reference list contains no prior work by this paper’s authors, and no self-citation is load-bearing. The circularity is in the benchmark/training identity. CIMR’s specialized training uses synthetic triples of the form initial erroneous response, corrective feedback, final correct response (Section IV-A1). The paper then claims state-of-the-art performance on MAP, a dataset the authors constructed to require exactly the same dynamic adjustment and self-correction behavior (Section IV-A2), with the official metric defined as successful execution “including successful self-correction when necessary” (Section IV-C). Consequently, the reported 91.5% versus 78.5% gap over LLaVA-1.5 and the 2.3-point margin over GPT-4V are measurements of how well the model matches the authors’ own synthetic template; they do not provide an independent test of generalized iterative reasoning. The absence of a released dataset/code and of any external benchmark means the central claim cannot be checked outside this template. A well-designed independent benchmark would resolve this, but as written the result is forced by the definition of the evaluation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 3 invented entities

The central claim rests on a self-contained, unreleased synthetic benchmark, an unspecified confidence stopping rule, and unstated training choices. The paper does not ship the dataset or code, so the measured accuracy is inseparable from the authors' private generation templates.

free parameters (3)

Maximum iteration count = not reported (implicitly 4 or more)
Table IV shows accuracy rising to 91.5 only at 4+ iterations; the stopping condition is described only as predefined, so the number of refinement rounds is a free choice that directly affects the reported accuracy.
Termination confidence threshold = not reported
Section IV-C3 says the loop stops when an internal confidence metric indicates completion, but the metric and threshold are never defined, making convergence non-objective and unreproducible.
Number of fine-tuning epochs and training steps = not reported
The multi-stage fine-tuning on VIT and the custom correction triples is described without epochs, batch sizes, learning rates, or data volumes, leaving the central behavior dependent on unreported choices.

assumptions (4)

ad hoc to paper The synthetic MAP distribution is representative of real-world complex instruction-following.
Section IV-A2 introduces MAP as a custom, synthetic dataset; no external dataset validates transfer, so the central accuracy claim inherits this untested assumption.
domain assumption The base LVLM can produce useful corrective feedback from its own outputs.
Section IV-C2 says feedback parsing uses an internal consistency check that relies on the base LVLM's reasoning capabilities; if self-critique is weak, the iterative loop cannot improve results.
domain assumption New visual observations are available after each action at test time.
Section IV-C2 integrates new visual observations after executing the previous response, but many real-world tasks do not provide simulated camera or sensor feedback after every action.
domain assumption Training on the custom correction triples does not leak the MAP test content.
The MAP dataset and the training triples are generated by the same authors with overlapping templates, and no separation procedure or contamination check is described.

invented entities (3)

MAP benchmark
purpose: Test CIMR on synthetic multi-step action planning requiring self-correction.
The dataset is created for this paper, is not released, and its size and split are not reported, so it provides no external validation for the central claim.
Internal confidence metric
purpose: Determine when to stop iterative refinement.
The metric is mentioned in Section IV-C3 but never defined, making the stopping condition unverifiable and unreproducible.
Dynamic fusion module
purpose: Fuse textual, visual, and contextual features at each iteration.
The module is described only through a CrossAttention equation; no architectural details or ablations isolate its contribution.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CIMR: Contextualized Iterative Multimodal Reasoning for Robust Instruction Following in LVLMs." pith.science (2026). https://pith.science/paper/QDO56K6W

@misc{pith2026250722074,
  author       = {Pith},
  title        = {Pith review of: CIMR: Contextualized Iterative Multimodal Reasoning for Robust Instruction Following in LVLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QDO56K6W}},
  note         = {Machine review of arXiv:2507.22074}
}

read the original abstract

The rapid advancement of Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) has enhanced our ability to process and generate human language and visual information. However, these models often struggle with complex, multi-step multi-modal instructions that require logical reasoning, dynamic feedback integration, and iterative self-correction. To address this, we propose CIMR: Contextualized Iterative Multimodal Reasoning, a novel framework that introduces a context-aware iterative reasoning and self-correction module. CIMR operates in two stages: initial reasoning and response generation, followed by iterative refinement using parsed multi-modal feedback. A dynamic fusion module deeply integrates textual, visual, and contextual features at each step. We fine-tune LLaVA-1.5-7B on the Visual Instruction Tuning (VIT) dataset and evaluate CIMR on the newly introduced Multi-modal Action Planning (MAP) dataset. CIMR achieves 91.5% accuracy, outperforming state-of-the-art models such as GPT-4V (89.2%), LLaVA-1.5 (78.5%), MiniGPT-4 (75.3%), and InstructBLIP (72.8%), demonstrating the efficacy of its iterative reasoning and self-correction capabilities in complex tasks.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 15 canonical work pages

[1]

A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly,

Y . Yao, J. Duan, K. Xu, Y . Cai, E. Sun, and Y . Zhang, “A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly,” CoRR, 2023

2023
[2]

Visual in-context learning for large vision-language models,

Y . Zhou, X. Li, Q. Wang, and J. Shen, “Visual in-context learning for large vision-language models,” in Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024 . Association for Computational Linguis- tics, 2024, pp. 15 890–15 902

2024
[3]

Draw all your imagine: A holistic benchmark and agent framework for complex instruction-based image generation,

Y . Zhou, J. Yuan, and Q. Wang, “Draw all your imagine: A holistic benchmark and agent framework for complex instruction-based image generation,” arXiv preprint arXiv:2505.24787 , 2025

arXiv 2025
[4]

Complexbench- edit: Benchmarking complex instruction-driven image editing via com- positional dependencies,

C. Wang, Y . Zhou, Q. Wang, Z. Wang, and K. Zhang, “Complexbench- edit: Benchmarking complex instruction-driven image editing via com- positional dependencies,” arXiv preprint arXiv:2506.12830 , 2025

arXiv 2025
[5]

Llava-more: A comparative study of llms and visual backbones for enhanced visual instruction tuning,

F. Cocchi, N. Moratelli, D. Caffagni, S. Sarto, L. Baraldi, M. Cornia, and R. Cucchiara, “Llava-more: A comparative study of llms and visual backbones for enhanced visual instruction tuning,” CoRR, 2025

2025
[6]

What makes for good visual instructions? synthesizing complex visual reasoning instructions for visual instruc- tion tuning,

Y . Du, H. Guo, K. Zhou, W. X. Zhao, J. Wang, C. Wang, M. Cai, R. Song, and J. Wen, “What makes for good visual instructions? synthesizing complex visual reasoning instructions for visual instruc- tion tuning,” in Proceedings of the 31st International Conference on Computational Linguistics, COLING 2025, Abu Dhabi, UAE, January 19-24, 2025 . Association f...

work page 2025
[7]

A review of multi-modal large language and vision models,

K. Carolan, L. Fennelly, and A. F. Smeaton, “A review of multi-modal large language and vision models,” CoRR, 2024

work page 2024
[8]

Visual instruction tuning towards general-purpose multimodal model: A survey,

J. Huang, J. Zhang, K. Jiang, H. Qiu, and S. Lu, “Visual instruction tuning towards general-purpose multimodal model: A survey,” CoRR, 2023

work page 2023

Show all 27 references

[9]

Visual question answering instruction: Unlocking multimodal large language model to domain- specific visual multitasks,

J. Lee, S. Cha, Y . Lee, and C. Yang, “Visual question answering instruction: Unlocking multimodal large language model to domain- specific visual multitasks,” CoRR, 2024

2024
[10]

Sharegpt4v: Improving large multi-modal models with better captions,

L. Chen, J. Li, X. Dong, P. Zhang, C. He, J. Wang, F. Zhao, and D. Lin, “Sharegpt4v: Improving large multi-modal models with better captions,” in Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XVII. Springe...

2024
[11]

Multi-modal large language models are effective vision learners,

L. Sun, C. Ahuja, P. Chen, M. D’Zmura, K. Batmanghelich, and P. Bontrager, “Multi-modal large language models are effective vision learners,” in IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2025, Tucson, AZ, USA, February 26 - March 6, 2025 . IEEE, 2025,...

2025
[12]

Modeling event-pair relations in external knowledge graphs for script reasoning,

Y . Zhou, X. Geng, T. Shen, J. Pei, W. Zhang, and D. Jiang, “Modeling event-pair relations in external knowledge graphs for script reasoning,” Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, 2021

2021
[13]

Claret: Pre-training a correlation-aware context-to-event transformer for event-centric gener- ation and classification,

Y . Zhou, T. Shen, X. Geng, G. Long, and D. Jiang, “Claret: Pre-training a correlation-aware context-to-event transformer for event-centric gener- ation and classification,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: L...

2022
[14]

Eventbert: A pre- trained model for event correlation reasoning,

Y . Zhou, X. Geng, T. Shen, G. Long, and D. Jiang, “Eventbert: A pre- trained model for event correlation reasoning,” in Proceedings of the ACM Web Conference 2022 , 2022, pp. 850–859

2022
[15]

UNIFIED- IO: A unified model for vision, language, and multi-modal tasks,

J. Lu, C. Clark, R. Zellers, R. Mottaghi, and A. Kembhavi, “UNIFIED- IO: A unified model for vision, language, and multi-modal tasks,” in The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023

2023
[16]

Beyond empathy: Integrating diagnostic and therapeutic reasoning with large language models for mental health counseling,

H. Hu, Y . Zhou, J. Si, Q. Wang, H. Zhang, F. Ren, F. Ma, and L. Cui, “Beyond empathy: Integrating diagnostic and therapeutic reasoning with large language models for mental health counseling,” arXiv preprint arXiv:2505.15715, 2025

2025
[17]

Queryagent: A reliable and efficient reasoning framework with environmental feedback based self-correction,

X. Huang, S. Cheng, S. Huang, J. Shen, Y . Xu, C. Zhang, and Y . Qu, “Queryagent: A reliable and efficient reasoning framework with environmental feedback based self-correction,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume...

2024
[18]

Au- tomatically correcting large language models: Surveying the landscape of diverse self-correction strategies,

L. Pan, M. Saxon, W. Xu, D. Nathani, X. Wang, and W. Y . Wang, “Au- tomatically correcting large language models: Surveying the landscape of diverse self-correction strategies,” CoRR, 2023

2023
[19]

Adareasoner: Adaptive reasoning enables more flexible thinking,

X. Wang, Y . Huang, Y . Wang, X. Luo, K. Guo, Y . Zhou, and X. Zhang, “Adareasoner: Adaptive reasoning enables more flexible thinking,” CoRR, 2025

2025
[20]

Plan-on-graph: Self-correcting adaptive planning of large language model on knowledge graphs,

L. Chen, P. Tong, Z. Jin, Y . Sun, J. Ye, and H. Xiong, “Plan-on-graph: Self-correcting adaptive planning of large language model on knowledge graphs,” in Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, Neu...

2024
[21]

Embodied-reasoner: Synergizing visual search, reasoning, and action for embodied interactive tasks,

W. Zhang, M. Wang, G. Liu, H. Xu, Y . Jiang, Y . Shen, G. Hou, Z. Zheng, H. Zhang, X. Li, W. Lu, P. Li, and Y . Zhuang, “Embodied-reasoner: Synergizing visual search, reasoning, and action for embodied interactive tasks,” CoRR, 2025

2025
[22]

Agent-r: Training language model agents to reflect via iterative self-training,

S. Yuan, Z. Chen, Z. Xi, J. Ye, Z. Du, and J. Chen, “Agent-r: Training language model agents to reflect via iterative self-training,” CoRR, 2025

2025
[23]

Mam: Modular multi-agent framework for multi-modal medical diagnosis via role-specialized collaboration,

Y . Zhou, L. Song, and J. Shen, “Mam: Modular multi-agent framework for multi-modal medical diagnosis via role-specialized collaboration,” arXiv preprint arXiv:2506.19835 , 2025

2025 arXiv
[24]

A theoretical under- standing of self-correction through in-context alignment,

Y . Wang, Y . Wu, Z. Wei, S. Jegelka, and Y . Wang, “A theoretical under- standing of self-correction through in-context alignment,” in Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouve...

2024
[25]

Improving medical large vision- language models with abnormal-aware feedback,

Y . Zhou, L. Song, and J. Shen, “Improving medical large vision- language models with abnormal-aware feedback,” arXiv preprint arXiv:2501.01377, 2025

2025 arXiv
[26]

Igniting language intelligence: The hitchhiker’s guide from chain-of-thought reasoning to language agents,

Z. Zhang, Y . Yao, A. Zhang, X. Tang, X. Ma, Z. He, Y . Wang, M. Ger- stein, R. Wang, G. Liu, and H. Zhao, “Igniting language intelligence: The hitchhiker’s guide from chain-of-thought reasoning to language agents,” ACM Comput. Surv., pp. 206:1–206:39, 2025. 9

2025
[2024]

5014–5035

Association for Computational Linguistics, 2024, pp. 5014–5035

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly,

Y . Yao, J. Duan, K. Xu, Y . Cai, E. Sun, and Y . Zhang, “A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly,” CoRR, 2023

2023

[2] [2]

Visual in-context learning for large vision-language models,

Y . Zhou, X. Li, Q. Wang, and J. Shen, “Visual in-context learning for large vision-language models,” in Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024 . Association for Computational Linguis- tics, 2024, pp. 15 890–15 902

2024

[3] [3]

Draw all your imagine: A holistic benchmark and agent framework for complex instruction-based image generation,

Y . Zhou, J. Yuan, and Q. Wang, “Draw all your imagine: A holistic benchmark and agent framework for complex instruction-based image generation,” arXiv preprint arXiv:2505.24787 , 2025

arXiv 2025

[4] [4]

Complexbench- edit: Benchmarking complex instruction-driven image editing via com- positional dependencies,

C. Wang, Y . Zhou, Q. Wang, Z. Wang, and K. Zhang, “Complexbench- edit: Benchmarking complex instruction-driven image editing via com- positional dependencies,” arXiv preprint arXiv:2506.12830 , 2025

arXiv 2025

[5] [5]

Llava-more: A comparative study of llms and visual backbones for enhanced visual instruction tuning,

F. Cocchi, N. Moratelli, D. Caffagni, S. Sarto, L. Baraldi, M. Cornia, and R. Cucchiara, “Llava-more: A comparative study of llms and visual backbones for enhanced visual instruction tuning,” CoRR, 2025

2025

[6] [6]

What makes for good visual instructions? synthesizing complex visual reasoning instructions for visual instruc- tion tuning,

Y . Du, H. Guo, K. Zhou, W. X. Zhao, J. Wang, C. Wang, M. Cai, R. Song, and J. Wen, “What makes for good visual instructions? synthesizing complex visual reasoning instructions for visual instruc- tion tuning,” in Proceedings of the 31st International Conference on Computational Linguistics, COLING 2025, Abu Dhabi, UAE, January 19-24, 2025 . Association f...

work page 2025

[7] [7]

A review of multi-modal large language and vision models,

K. Carolan, L. Fennelly, and A. F. Smeaton, “A review of multi-modal large language and vision models,” CoRR, 2024

work page 2024

[8] [8]

Visual instruction tuning towards general-purpose multimodal model: A survey,

J. Huang, J. Zhang, K. Jiang, H. Qiu, and S. Lu, “Visual instruction tuning towards general-purpose multimodal model: A survey,” CoRR, 2023

work page 2023