Pith. sign in

REVIEW 5 major objections 5 minor 12 references

CASHEW claims that iterative aggregation of candidate reasoning trajectories, with explicit visual verification, stabilizes multimodal reasoning, and its reinforced variant internalizes this behavior.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Iteratively aggregating multiple reasoning trajectories with Grounding-DINO-verified visual evidence improved VLM accuracy on 13 multimodal benchmarks, and an RL-trained version internalizes the aggregation.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A useful test-time aggregation method with a promising RL variant, but the headline gains rest on an undocumented train/eval split. the 5 major comments →

arxiv 2601.08010 v3 pith:AC6Y7XZT submitted 2026-01-12 cs.CV

CASHEW: Stabilizing Multimodal Reasoning via Iterative Trajectory Aggregation

Chaoyu Li , Fei Tao , Pooyan Fazli This is my paper
classification cs.CV
keywords multimodal reasoningvision-language modelstest-time scalingtrajectory aggregationvisual groundinghallucination filteringreinforcement learningvideo understanding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 multi-step vision-language reasoning is unstable because single trajectories are sampled independently, and that the right fix is to treat reasoning as an evolving population: sample multiple candidate trajectories, verify the objects they mention against the visual input with a frozen detector, and iteratively synthesize the survivors into a more grounded trace. If this is right, test-time scaling for multimodal models does not need larger backbones or longer chains; it needs aggregation with external visual grounding. The authors further claim that the same aggregation behavior can be trained into a single model with reinforcement learning, so a one-pass model can approximate the iterative method's gains. Across 13 image and video benchmarks they report consistent improvements, with the largest claimed at +26.2 percentage points on ScienceQA and +9.1 on EgoSchema.

Core claim

Multimodal reasoning is treated as aggregation: rather than selecting one trajectory, merge several while grounding their object claims against the visual input with a frozen detector. The loop samples N=8 candidates, aggregates subsets of size M=4 conditioned on objects verified by Grounding DINO, and repeats for T=3 iterations before a final merge. CASHEW-RL internalizes this via Group Sequence Policy Optimization—comparing several rollouts per prompt—to produce a structured <think>/<visual_keys>/<answer> trace, with rewards for correctness, visual-key precision-recall, and difficulty-adjusted length. The paper reports gains at T=1, larger gains at T=3, and wins over majority voting, selec

What carries the argument

The engine is the iterative aggregation loop: population initialization, random subset sampling, grounded object verification by a frozen Grounding DINO detector (threshold 0.35), and conditioned synthesis of a new trajectory. The learned variant's engine is GSPO—group-relative policy optimization that compares multiple rollouts for the same prompt and candidate set—driven by a composite reward: answer accuracy (exact match), visual-key quality (balanced precision and recall against a ground-truth key set), and a length penalty scaled by an EMA of the empirical solve rate so easy prompts are penalized for verbosity while hard prompts keep longer reasoning.

Load-bearing premise

The central claim depends on the 10,000 ScienceQA and 2,500 NExT-QA instances used for RL training (Table 4) being disjoint from the evaluation sets in Tables 1-2, which the paper never states; if they overlap, the headline gains on those benchmarks measure memorization rather than aggregation.

What would settle it

Inspect the RL corpus for question-id overlap with the ScienceQA and NExT-QA evaluation splits; if overlap exists, retest on a held-out subset. Also reconcile the abstract's +26.2/+9.1 percentage point claims with the body's +23.6/+8.1, and rerun the ablation with the detector's outputs randomized to confirm visual verification is what drives the gains.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Frozen vision-language models, from 4B to 8B parameters, can be improved at inference time without retraining, so test-time scaling can be decoupled from backbone scale.
  • Explicit visual verification is necessary: removing Grounding DINO from the loop reduces gains on ScienceQA, POPE, EgoSchema, and VSI-Bench, so object-grounding is doing real work beyond aggregation alone.
  • Iterative aggregation outperforms majority voting, model-based trajectory selection, and single-round synthesis, suggesting that merging partial insights beats choosing one trajectory.
  • CASHEW-RL at a single aggregation pass already exceeds the base model on most benchmarks, meaning the reinforcement-learned policy captures much of the benefit without repeated test-time sampling.
  • The method applies beyond images: consistent gains on video understanding and video reasoning benchmarks imply the same grounded-aggregation loop transfers to temporal and egocentric reasoning.

Where Pith is reading between the lines

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

  • If the training and evaluation splits are confirmed disjoint, the large ScienceQA and NExT-QA gains suggest the RL objective teaches a transferable grounding policy rather than memorizing answers; this is testable by publishing question ids in the RL corpus.
  • Because verification relies on an open-set detector, the method inherits the detector's blind spots: true claims about objects outside its recognition vocabulary may be filtered out, so gains should shrink on tasks with rare, abstract, or heavily occluded entities.
  • The difficulty-aware length penalty makes a per-item prediction — easy prompts should produce shorter reasoning and hard prompts longer reasoning — which can be measured directly from the T=1 rollouts without retraining.
  • The qualitative examples show the framework correcting temporal claims (e.g., which hand is used), but the formal verification is object-level only; extending grounding to actions and events would be a natural next step and could be tested by using a spatio-temporal verifier in the same loop.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes CASHEW, a test-time framework that stabilizes multimodal reasoning by sampling multiple candidate reasoning trajectories, verifying object-level claims with Grounding DINO, and iteratively aggregating the trajectories into a grounded consensus. It also introduces CASHEW-RL, a post-trained variant that internalizes this aggregation behavior via supervised fine-tuning followed by Group Sequence Policy Optimization (GSPO) with a composite reward combining answer correctness, visual-key overlap, and a difficulty-aware length penalty. Experiments on 13 image/video benchmarks report consistent improvements over base VLMs and over prior test-time scaling baselines, with headline gains of up to +26.2 points on ScienceQA and +9.1 points on EgoSchema.

Significance. If the empirical claims hold, this is a meaningful contribution to multimodal test-time scaling: it shows that aggregating multiple trajectories with explicit visual verification can reduce hallucinations and improve reasoning stability across a wide range of benchmarks and backbone models. The framework is modular (frozen VLM + detector) and the RL formulation with a composite reward for evidence-grounded, effort-adaptive reasoning is a useful design. The ablation RQ1 demonstrates that grounding verification contributes beyond plain aggregation, and the paper ships substantial experiments (13 benchmarks, multiple backbones). However, the evaluation has integrity issues—most importantly the potential train/eval overlap on two benchmarks—that must be resolved before the main claims can be accepted.

major comments (5)
  1. [Table 4 vs. Tables 1–2, 8] The RL training corpus explicitly includes ScienceQA (10,000 instances) and NExT-QA (2,500 instances), and CASHEW-RL results are reported on ScienceQA (Tables 1 and 8) and NExT-QA (Table 2). The paper never states that the training and evaluation instances are disjoint. Since Racc (Eq. 8) is exact-match against the ground-truth answer, any overlap would allow the model to memorize answers and would invalidate the headline gains (+26.2 on ScienceQA in Table 8; +4.9 on NExT-QA in Table 2). Please report the exact split (e.g., question IDs, version of the benchmark) or retrain without these two benchmarks in the RL/SFT corpora.
  2. [Eq. (9) and Section 5.1] Rkey is defined against a 'ground-truth visual-key set G derived from dataset annotations,' but Section 5.1 describes G as being extracted from teacher-generated reasoning chains (Qwen3-VL-30B-Thinking) and verified by Grounding DINO. This is not independent ground truth: it is the output of the same teacher model that generates the candidate trajectories. Consequently, Rkey rewards agreement with the teacher's stated objects, not with objectively annotated visual entities. Please rename/clarify the reward term and, ideally, validate G against human annotations or use dataset-supplied object annotations.
  3. [Tables 1, 2, 8] The table captions state that '95% confidence intervals (CIs, ±) were computed via bootstrap resampling (10,000 iterations, percentile method)' and that significance is judged by whether the CI excludes zero, yet no intervals are shown anywhere in the tables—only § marks for some rows. The reader cannot verify the significance claims, and the summary statement that 'most gains are statistically significant' is unsubstantiated. Please include the CIs (or p-values) for all reported improvements.
  4. [Abstract/Introduction vs. Tables] The Abstract reports gains of up to +26.2 on ScienceQA and +9.1 on EgoSchema; the Introduction reports +23.6 and +8.1. Table 8 indeed shows +26.2 for Qwen3-VL-4B CASHEW-RL on ScienceQA, but the largest EgoSchema gain in Table 2 is +8.1 (Qwen2.5-VL-7B, CASHEW), not +9.1, and CASHEW-RL's EgoSchema gain is +4.3. Please reconcile these numbers and explicitly state which configuration produces the headline result.
  5. [Section 6.2 / Table 8] The paper claims CASHEW-RL 'generalizes across backbone models,' but this is only tested on Qwen3-VL-4B and Qwen3-VL-8B, which share the same architecture family and training data. To support a cross-backbone generalization claim, at least one architecturally distinct backbone (e.g., InternVL3.5-8B or a LLaVA variant) should be evaluated with the same RL recipe. Otherwise, the claim should be limited to 'scaling to a smaller model in the same family.'
minor comments (5)
  1. [Figure 2 caption] Typo: 'CASHEW ASHEWC -RL' should be 'CASHEW and CASHEW-RL'.
  2. [Eq. (14)] The KL regularization term is written as 'αKLKL[πθ∥πref]', which is confusing. Use α_KL · KL(·) or a clear product notation.
  3. [Section 4, Grounded Aggregation] The sentence 'encouraging it to anchor the new trajectory τ(t+1)i to anchor its reasoning in verified evidence' contains a redundant phrase ('anchor ... to anchor').
  4. [Appendix B.2] The example JSON uses 'visual_key' singular in one candidate but 'visual_keys' in others; make the format consistent.
  5. [Section 5.3, Eq. (12)] The lower bound max(˜pt,1/J) is not discussed. With J=4, this imposes a minimum penalty of 0.25·β·Ntok even on easy, always-solved prompts. Please justify this choice or show its effect in ablations.

Circularity Check

2 steps flagged

CASHEW's inference-time loop is not circular, but CASHEW-RL's headline gains on ScienceQA and NExT-QA are compromised by an unstated train/eval overlap, and its visual-key reward target is teacher-generated rather than independent.

specific steps
  1. fitted input called prediction [Appendix B Table 4; Section 6.2 Tables 1, 2, and Table 8]
    "Table 4: CASHEW-RL training dataset distribution for SFT and RL stages. Dataset SFT (# Samples) RL (# Samples) ... ScienceQA (Lu et al., 2022) 2,000 10,000 ... NExT-QA (Xiao et al., 2021) 600 2,500 ... Table 1: + CASHEW-RL (T= 3) 97.8(+4.9) ... Table 8: CASHEW-RL improves ScienceQA accuracy from 69.5% to 95.7% (+26.2 percentage points)."

    The RL training corpus explicitly contains 10,000 ScienceQA and 2,500 NExT-QA instances, and the evaluation tables report CASHEW-RL accuracy on ScienceQA and NExT-QA. The paper never states that these training and evaluation instances are disjoint. If they are not disjoint, the reward R_acc = 1[a = y*] in Eq. (8) is optimized on the very answers used in the reported gains, so the headline CASHEW-RL improvements on those two benchmarks measure memorization/overfitting rather than learned aggregation. The limitations section only mentions perception-model dependency and does not address split separation.

  2. other [Section 5.1 (RL Data) and Eq. (9)]
    "RL Data. We curate a 200k-instance corpus from the same source datasets. Each instance contains the original annotation (x, q, y*), three diverse candidate trajectories {τi}3 i=1 generated using Qwen3-VL-30B-Thinking, and a shared set of visual keys extracted from the candidates and verified by Grounding DINO. ... Let G denote the ground-truth visual-key set derived from dataset annotations, and let K be the model-predicted key set. Rkey = (1−α) |K∩G|/|G| + α |K∩G|/|K|."

    The 'ground-truth visual-key set' G in Eq. (9) is not an independent annotation. Section 5.1 says the visual keys are 'extracted from the candidates' that were themselves 'generated using Qwen3-VL-30B-Thinking.' Thus the grounding reward Rkey rewards the trained aggregator for rediscovering the same teacher's key sets, so the evidence-selection signal partly reduces to teacher imitation. The claim that RL improves visual grounding is therefore partially self-referential: the reward target is produced by the same model family that generates the candidate trajectories, not by an externally verified key annotation.

full rationale

The paper's core CASHEW pipeline has no circular derivation: Eq. (5) aggregates sampled trajectories conditioned on externally verified visual evidence from a frozen Grounding DINO detector, and its claims are tested against external benchmarks and compared with self-consistency/self-synthesizer baselines. The self-citations (e.g., VidHalluc, Kulkarni-Fazli) are related-work pointers, not load-bearing uniqueness theorems. The main circularity risk is in CASHEW-RL's evaluation: Table 4 places ScienceQA and NExT-QA in the RL training corpus, while Tables 1, 2, and 8 report CASHEW-RL accuracy on those same benchmarks, with no stated disjoint split; if the same instances are used, R_acc in Eq. (8) is optimized on the eval answers, so the reported +4.9 and +26.2 point gains on those benchmarks measure memorization, not learned aggregation. An additional definitional weakness is that the 'ground-truth visual-key set' G in Eq. (9) is constructed from the same teacher (Qwen3-VL-30B-Thinking) that generated the candidate trajectories, so the grounding reward partly reduces to teacher imitation. I also note the abstract/intro gains differ (+23.6/+8.1 vs +26.2/+9.1), a reporting inconsistency rather than circularity. Because most benchmarks and the inference-time CASHEW results are independent, the circularity is partial.

Axiom & Free-Parameter Ledger

14 free parameters · 6 axioms · 0 invented entities

The central claim rests on several free hyperparameters, a detector-as-verifier assumption, and possibly overlapping train/eval data. The most important non-standard assumption is that RL training on ScienceQA/NExT-QA does not leak into their evaluation; this is unstated. Otherwise, the approach uses standard GSPO and external grounding models.

free parameters (14)
  • Grounding DINO confidence threshold δg = 0.35
    Eq. 3: objects are considered present only if G(I, ok)>δg; threshold chosen by hand, directly controls which evidence survives verification.
  • Population size N = 8
    Number of candidate trajectories per iteration; selected via ablation across {4,...,16} (Fig. 3, Appendix D).
  • Iteration count T = 3
    Aggregation depth; selected via ablation T∈{1,2,3,4} (Fig. 3, Appendix D).
  • Group size K/M = 4
    Candidates aggregated per synthesis step; fixed as stated in Section 6.
  • Reward weight wacc = 1.0
    Answer-correctness weight in Eq. 7.
  • Reward weight wkey = 0.35
    Evidence-overlap weight in Eq. 7.
  • F1 balancing coefficient α = 0.5
    Precision-recall balance in Eq. 9.
  • Length penalty coefficient β = 0.001
    Strength of difficulty-aware length penalty in Eq. 12.
  • EMA decay γ = 0.9
    Smoothing of empirical solve rate in Eq. 11.
  • GSPO temperature λ = 1.0
    Sharpness of within-group softmax weights in Eq. 13.
  • KL coefficient α_KL = 0.02
    KL regularization against SFT reference policy in Eq. 14.
  • Decoding temperature = 0.8
    Sampling temperature for candidate trajectories (Section 6).
  • top_p = 0.95
    Nucleus sampling probability (Section 6).
  • Number of rollouts J = 4
    Rollouts per prompt for solve-rate estimate and GSPO group (Eq. 10, Section 6).
axioms (6)
  • domain assumption Grounding DINO confidence scores are a valid proxy for whether objects mentioned in reasoning actually occur in the visual input.
    Invoked in Eq. 3 for every trajectory at every iteration; if the detector misses small or video objects, verification silently discards valid evidence.
  • domain assumption The ground-truth visual-key set G in Rkey is available and reliable for the benchmarks and RL instances.
    Eq. 9 assumes G exists; in Section 5.1 the keys are extracted from teacher reasoning and verified by Grounding DINO, so 'ground truth' may be teacher-mediated.
  • domain assumption RL training on ScienceQA and NExT-QA does not overlap the evaluation instances of those benchmarks.
    Table 4 includes 10k ScienceQA and 2.5k NExT-QA training samples; Tables 1-2 report those benchmarks. The paper does not state split/exclusion.
  • standard math GSPO update and KL regularization from Zheng et al. (2025) are accepted background.
    Eq. 14 uses the GSPO objective and SFT reference policy without derivation.
  • domain assumption Exact-match on benchmark answers is a sufficient measure of reasoning quality for reward optimization.
    Racc in Eq. 8 uses exact match; it ignores partial correctness and may reward guessing on multiple-choice benchmarks.
  • ad hoc to paper The difficulty-aware length penalty's EMA solve rate approximates true task difficulty.
    Eqs. 10-12 define R_len from J rollouts and EMA; this is a design choice specific to this paper.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CASHEW: Stabilizing Multimodal Reasoning via Iterative Trajectory Aggregation." pith.science (2026). https://pith.science/paper/AC6Y7XZT

@misc{pith2026260108010,
  author       = {Pith},
  title        = {Pith review of: CASHEW: Stabilizing Multimodal Reasoning via Iterative Trajectory Aggregation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AC6Y7XZT}},
  note         = {Machine review of arXiv:2601.08010}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Vision-language models achieve strong performance across a wide range of multimodal understanding and reasoning tasks, yet their multi-step reasoning remains unstable. Repeated sampling over the same input often produces divergent reasoning trajectories and inconsistent final predictions. To address this, we introduce two complementary approaches inspired by test-time scaling: (1) CASHEW, an inference-time framework that stabilizes reasoning by iteratively aggregating multiple candidate trajectories into higher-quality reasoning traces, with explicit visual verification filtering hallucinated steps and grounding reasoning in visual evidence, and (2) CASHEW-RL, a learned variant that internalizes this aggregation behavior within a single model. CASHEW-RL is trained using Group Sequence Policy Optimization (GSPO) with a composite reward that encourages correct answers grounded in minimal yet sufficient visual evidence, while adaptively allocating reasoning effort based on task difficulty. This training objective enables robust self-aggregation at inference. Extensive experiments on 13 image understanding, video understanding, and video reasoning benchmarks show significant performance improvements, including gains of up to +26.2 percentage points on ScienceQA and +9.1 percentage points on EgoSchema.

Figures

Figures reproduced from arXiv: 2601.08010 by Chaoyu Li, Fei Tao, Pooyan Fazli.

Figure 1
Figure 1. Figure 1: CASHEW enables robust reasoning through visually grounded iterative aggregation. Unlike stan￾dard vision-language models that rely on single-path reasoning and are prone to hallucinations, CASHEW ag￾gregates multiple reasoning trajectories with explicit visual verification. CASHEW-RL further internalizes this aggregation behavior via reinforcement learning. 2025a), and inconsistent across inference runs (L… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the CASHEW and CASHEW-RL frameworks. Left: CASHEW performs test-time iterative aggregation by generating a population of candidate trajectories from a frozen VLM, verifying object-level claims with Grounding DINO, and synthesizing subsets into refined trajectories over multiple iterations to produce a consolidated trajectory τ ∗ . Right: CASHEW-RL extends this framework via post-training with G… view at source ↗
Figure 3
Figure 3. Figure 3: Performance across CASHEW population (N) for different values of T. All results with fixed K = 4 [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Prompt templates for different stages of [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Prompt used for supervised fine-tuning (SFT) in [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Prompt templates for Self-Synthesizer in ablation study. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7 [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

12 extracted references · 6 linked inside Pith

  1. [1]

    In <think></think>, write a single coherent reasoning chain that compares and aggregates the candidate answers, and discards incorrect or unsupported claims

  2. [2]

    Only include objects that provide useful visual evidence

    In <visual_keys></visual_keys>, output a Python-style list of objects from the visual input that are most relevant for answering the question. Only include objects that provide useful visual evidence

  3. [3]

    object_1

    In <answer></answer>, provide one concise and correct final answer. Output Format: <think> Your reasoning chain here. </think> <visual_keys> ["object_1", "object_2", ...] </visual_keys> <answer> Your final answer here. </answer> Figure 5: Prompt used for supervised fine-tuning (SFT) in CASHEW-RL. The prompt enforces a structured output format consisting o...

  4. [4]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Car- roll L

    Training vision-language process reward models for test-time scaling in multimodal reasoning: Key insights and lessons learned.arXiv preprint arXiv:2509.23250. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Car- roll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Mil...

    arXiv 2022

  5. [5]

    InProceedings of the Conference on Empirical Methods in Natural Language Process- ing (EMNLP)

    PlanGEN: A multi-agent framework for gener- ating planning and reasoning trajectories for complex problem solving. InProceedings of the Conference on Empirical Methods in Natural Language Process- ing (EMNLP). Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn

  6. [7]

    InProceedings of the Thirty- Seventh Annual Conference on Neural Information Processing Systems (NeurIPS)

    Reflexion: Language agents with verbal re- inforcement learning. InProceedings of the Thirty- Seventh Annual Conference on Neural Information Processing Systems (NeurIPS). Weiqin Wang, Yile Wang, and Hui Huang. 2025a. Ranked voting based self-consistency of large lan- guage models. InProceedings of the Annual Meet- ing of the Association for Computational...

    Pith/arXiv arXiv 2023

  7. [8]

    InProceedings of the Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track (NeurIPS)

    Longvideobench: A benchmark for long- context interleaved video-language understanding. InProceedings of the Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track (NeurIPS). Penghao Wu and Saining Xie. 2024. V*: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference o...

    2024

  8. [9]

    question_id

    Just enough thinking: Efficient reasoning with adaptive length penalties reinforcement learning. arXiv preprint arXiv:2506.05256. Junbin Xiao, Xindi Shang, Angela Yao, and Tat- Seng Chua. 2021. Next-qa:next phase of question- answering to explaining temporal actions. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CV...

    Pith/arXiv arXiv 2021

  9. [2017]

    something something

    The "something something" video database for learning and evaluating visual common sense. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, I...

    Pith/arXiv arXiv 2022

  10. [2023]

    InProceedings of the Thirty-Seventh Annual Conference on Neural Information Processing Systems (NeurIPS)

    Direct preference optimization: Your language model is secretly a reward model. InProceedings of the Thirty-Seventh Annual Conference on Neural Information Processing Systems (NeurIPS). Ziyao Shangguan, Chuhan Li, Yuxuan Ding, Yanan Zheng, Yilun Zhao, Tesca Fitzgerald, and Arman Cohan. 2025. Tomato: Assessing visual temporal reasoning capabilities in mult...

    Pith/arXiv arXiv 2025

  11. [2024]

    Mme: A comprehensive evaluation bench- mark for multimodal large language models.arXiv preprint arXiv:2306.13394. Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, Peixian Chen, Yanwei Li, Shaohui Lin, Sirui Zhao, Ke Li, Tong Xu, Xiawu Zheng, Enhong Chen, Caifeng Shan, and 2 other...

    Pith/arXiv arXiv 2025

  12. [2025]

    Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mo- hammad Shoeybi, and Song Han

    Longperceptualthoughts: Distilling system-2 reasoning for system-1 perception.arXiv preprint arXiv:2504.15362. Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mo- hammad Shoeybi, and Song Han. 2024. Vila: On pre- training for visual language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Haotian Liu, Chu...

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.