BlueLM-2.5-3B Technical Report

Baojiao Xiong; Boheng Chen; Chengzhi Wang; Daxiong Luo; Dongsheng Xu; Dongyang Liu; Fangyuan Li; Fan Yang; Fei Teng; Feng Wang

REVIEW 6 major objections 5 minor 81 references

BlueLM-2.5-3B Technical Report

T0 review · 6 major / 5 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 BlueLM-2.5-3B claims a 2.9B dense multimodal model can switch between fast and deep reasoning modes.

desk verdict A serious small-model systems report with a genuinely useful thinking-mode switch, but the headline benchmark parity rests on self-reported, contamination-sensitive numbers that need artifacts before they convince me. read the letter →

arxiv 2507.05934 v1 pith:PJHK6O7W submitted 2025-07-08 cs.AI

Baojiao Xiong , Boheng Chen , Chengzhi Wang , Daxiong Luo , Dongsheng Xu , Dongyang Liu , Fan Yang , Fangyuan Li

show 53 more authors

Fei Teng Feng Wang Fukang Qin Fuquan Peng Guanxin Tan Guozhi Wang Haibo Yu Haohao Gao Heng Liu Hongbo Yang Hongjian Zou Houzheng Shen Hu Meng Huan Li Hui Tan Jiali Chen Jianzhao Chen Jinliang Zhu Kai Wang Lei Wu Liangbing Liu Liuyang Bian Liyan He Long Liu Peiwen Li Penggang Shi Qi Ding Rui Hu Shuai Cao Shuai Ren Shuang Peng Teng Xie Weiji Chen Weilin Xiang Weixin Wu Xi Yin Xiaoxin Chen Xu Chen Yafei Wen Yan Hu Yanzhou Yang Yina Xie Yinghao Chen Yixuan Liao Yu Geng Yuanjiang Ouyang Yuanzhuo Yang Yuehua He Yushuai Peng Zhaoxiong Wang Zheng Wang Zhibo Zhou Ziyang Wu

This is my paper

classification cs.AI

keywords multimodallargelanguagemodelsedgedeploymentthinkingmodetokenbudgetcontrolreinforcementlearningdataefficiencyreasoningvision-languagemodel

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 aims to establish that a small dense multimodal model can serve as an edge-deployable assistant without sacrificing either fast response mode or deep reasoning. It presents BlueLM-2.5-3B, a 2.9B-parameter multimodal model that toggles between a thinking mode and a non-thinking mode through a control token, with explicit control over how many tokens the model may spend thinking. If the claims hold, a single small model could run on phones or vehicles and in thinking mode match a 4B text-only reasoning model on text benchmarks while trailing a 16B multimodal model by only about 5% on multimodal benchmarks. In non-thinking mode, it is reported to outperform a similarly sized vision-language model on most multimodal benchmarks, and the whole recipe reportedly uses less total training data than comparable models.

What carries the argument

The load-bearing mechanism is a special control token, written [|BlueThink|], placed at the end of the query to switch the model between a long chain-of-thought thinking mode and a fast non-thinking mode. Around that core, the training recipe stacks a multi-phase pipeline: pure-text pretraining of a 2.5B backbone via pruning and distillation from a 7B teacher, joint text-image pretraining with 40% pure-text data, a reasoning-enhanced stage with long and short chain-of-thought synthetic data, a fast-decay long-context stage, and then a hybrid reinforcement learning stage. Within the RL stage, a dynamic length penalty drawn from a per-group comparison the authors call Group Overlong encourages the model to compress its reasoning so it stays accurate under a tight token budget. This combination of mode control, length-aware RL, and data curation is what the paper argues lets a dense 2.9B model close the gap to far larger systems.

What would settle it

Run the released checkpoint, if it becomes available, on AIME24 and LiveCodeBench-v1 under the stated sampling settings, then audit whether near-duplicates of test problems appear in the described training data; if the scores drop materially or any overlap surfaces, the central comparison claim is falsified.

Watch

Extended reading notes

Core claim

The paper claims that a 2.9B-parameter dense multimodal model can simultaneously offer a fast non-thinking mode and a long-reasoning thinking mode, with the user controlling the reasoning budget through a dedicated token. The authors report that in thinking mode the model is comparable to a 4B-parameter text-only reasoning model on text benchmarks and trails a much larger 16B MoE multimodal reasoning model by only about 5% on average across ten multimodal benchmarks. In non-thinking mode, it is said to outperform a similarly scaled vision-language model on most multimodal metrics and across the text benchmarks shown. The report also claims especially strong accuracy at constrained token budgets under 8K tokens on AIME24 and LiveCodeBench-v1, and a measured thinking-switch failure rate below one in a million uses. The contribution is framed as the first 3B-scale multimodal model with both thinking and non-thinking modes plus explicit token-budget control, achieved with 17.1T total pre-training tokens, about 23% less than a comparable model.

Load-bearing premise

The load-bearing premise is that the reported benchmark numbers are clean, complete, and measured under conditions identical to those of the baselines, since the report supplies no externally verified decontamination certificate and no released artifacts.

Editorial extensions

If this is right

A single 2.9B dense model can serve both fast-response and deliberate-reasoning workloads, with the user choosing the mode by adding or omitting a control token.
In thinking mode, the model's text-only benchmark scores are comparable to a 4B-parameter reasoning model while using about 60% of that model's parameters.
Across ten multimodal benchmarks in thinking mode, the reported gap to a 16B MoE reasoning model averages about 5%, and the model beats a 72B multimodal model on two math-vision suites.
Under constrained token budgets below 8K tokens, the model reports large accuracy gains over both baselines on AIME24 and LiveCodeBench-v1, which matters for latency-limited edge deployment.
The recipe uses substantially less total pre-training data than comparable open models, suggesting the gains come from data curation and training design rather than raw scale.

Reading between the lines

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

Editorial inference: If the reported switch reliability of fewer than one failure per million uses replicates outside the authors' own data distribution, the end-of-query control-token pattern is a portable design that other small dense models could adopt without architectural change.
Editorial inference: The Group Overlong length penalty and the long-CoT data selection are separable interventions; an ablation holding data fixed while toggling the penalty would isolate how much of the under-8K-token accuracy gain comes from RL length control.
Editorial inference: The data-efficiency claim would become stronger or weaker with an explicit scaling-curve comparison at matched token budgets across training, which the report does not provide.
Editorial inference: If these results replicate, they would push edge vision-language design toward dense models with explicit mode control rather than mixture-of-experts architectures, since close parity is achieved at a lower parameter count and lower inference complexity.

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

Here's my take: this is a serious small-model engineering report, and the switchable thinking/non-thinking mode with explicit token-budget control is a genuinely useful capability for edge deployment. The headline parity claims with Qwen3-4B and Kimi-VL are plausible but not proven; the numbers rest on self-reported evaluation, small high-stakes test sets, and no released artifacts.

What's actually new: the [|BlueThink|] control token plus length-penalty RL gives users a mode switch and a token budget at 2.9B dense scale. The training recipe—SigLIP2 + AnyRes, 40% pure text in joint pretraining, pruning+distillation from a 7B, curriculum-style reasoning data, and hybrid RL with a dynamic length penalty—is a coherent and informative combination. The data-efficiency claim (17.1T total vs 22.1T for Qwen2.5-VL-3B) is worth taking seriously, and the infrastructure sections are concrete (context parallelism, async RL, load balancing). The citation pattern is fair: they credit Qwen, Kimi, and the RL systems they build on.

Where it frays: no weights, code, or data are released, so all benchmark results are self-reported. The decontamination description (Sections 3.7.4/3.7.8) is qualitative—no hash lists, overlap counts, or benchmark-version details—and the reasoning-eval conclusions are sensitive: AIME24 and AIME25 are 30 problems each, so one memorized or leaked problem is worth about 3.3 points. The reported AIME24 gap to Qwen3-4B is exactly one problem, and the AIME25 edge is one problem, so the "comparable to Qwen3-4B" claim hinges on a couple of items. There's also an in-house GUI benchmark (ScreenSpot vivo) with no public description, and the "<1 PPM" thinking-switch reliability number has no methodology behind it. Token accounting is inconsistent: Section 3.3 lists 950B + 450B = 1.4T reasoning tokens, while Section 2.2.3 says the reasoning-enhanced stage used 2.5T, and Section 3.6 claims 3.3T across stages. Those differ by a wide margin and the report doesn't reconcile them.

Who it's for: people building small MLLMs for phones and robots, and anyone working on training recipes for efficient multimodal reasoning. It deserves a serious referee, but the referee should ask for artifacts, a detailed contamination report, and fixed token accounting. As it stands, treat the benchmark tables as promising but unverified.

Referee Report

6 major / 5 minor

Summary. The manuscript presents BlueLM-2.5-3B, a 2.9B-parameter dense multimodal LLM developed by vivo AI Lab. The authors describe a three-stage pretraining pipeline (pure-text distillation, joint text-image pretraining with 40% text, reasoning-enhanced and long-context stages), a post-training pipeline with SFT and hybrid RL, infrastructure engineering, and an evaluation on 20+ benchmarks. The headline claims are that the model supports both thinking and non-thinking modes through the [BlueThink] control token, that thinking mode is comparable to Qwen3-4B on text benchmarks and within roughly 5% of Kimi-VL-A3B-16B on multimodal benchmarks, and that this is achieved with substantially less total pretraining data than Qwen2.5-VL-3B and Qwen3-4B.

Significance. If the empirical claims hold, the contribution is significant for edge deployment of small MLLMs. The paper's strengths are the unusually detailed data-curation and infrastructure descriptions, the explicit mode-switch mechanism with a length-penalty objective aimed at token-budget control, and the breadth of benchmarks considered. However, the manuscript does not release model weights, benchmark prompts, or evaluation harnesses; all central comparisons are self-reported, and several quantitative claims are internally inconsistent or lack measurement detail. The soundness of the headline parity claims therefore cannot be assessed at the level required for a research contribution until the contamination and uncertainty issues below are addressed.

major comments (6)

[§3.7.4, §3.7.8, §3.3, Tables 4-7] The central reasoning claims (AIME24 73.3 vs 76.7 for Qwen3-4B; AIME25 66.7 vs 65.6; LiveCodeBench-v1 81.5) are separated by at most one or two items on 30-problem tests, yet the decontamination evidence is only qualitative. The manuscript states that rigorous deduplication is performed but reports no overlap counts, hash lists, benchmark-version identifiers, or held-out checks for the teacher-generated data described in Section 3.3 or for the web- and PDF-derived corpora. Since one leaked AIME problem changes the score by about 3.3 points, the parity claim cannot be distinguished from test-set memorization without released artifacts or quantitative contamination reporting.
[§5, Tables 4-8] All benchmark numbers are single-point estimates with no error bars, no repeated sampling, no significance tests, and no released evaluation harness. The comparable-performance claims are based on differences of one to a few benchmark items on small sets. The manuscript only states that OpenCompass was used; it does not report decoding parameters, numbers of samples, or per-benchmark prompts. Please provide confidence intervals or repeated-sample statistics, at least for the 30-item AIME sets, and state the evaluation protocol for each score.
[§2.2.3, §3.3, §3.6, Table 3] The token counts for the reasoning-enhanced stage are internally inconsistent. Section 2.2.3 reports 2.5T tokens (1.3T textual + 1.2T multimodal); Section 3.3 reports approximately 950B pre-training tokens plus 450B synthetic tokens, totaling 1.4T; Section 3.6 and the Introduction state that 3.3T tokens of reasoning-enhanced data were used across these stages. Although the 4T + 2.5T + 1.3T entries in Table 2 do sum to the 7.8T MLLM pretraining number in Table 3, the descriptions in Sections 2.2.3, 3.3, and 3.6 cannot all be correct. Since Section 3.6 uses these totals to support the data-efficiency claim, a single consistent breakdown by stage and by data type must be provided.
[§3.5.1] The claim that the [BlueThink] switch has a failure rate below 1 PPM is not supported by any definition of failure, sample size, or testing protocol. This quantitative reliability claim is used to justify the mode-switch design and the practical usability of the model. Please report the evaluation set, the measurement procedure, and the observed failure count.
[Table 8] The GUI comparison rests in part on ScreenSpot vivo, an in-house Chinese GUI grounding benchmark that is neither released nor described. The table provides no sample size, task format, annotation protocol, or metric definition for this benchmark. Without this information, the statement that BlueLM-2.5-3B outperforms Qwen2.5-VL-7B and UI-TARS-7B on this benchmark cannot be independently checked. Please release the benchmark or replace it with public GUI benchmarks for the headline GUI claims.
[§2.3.5, Eq. (1)] The text says that the length difference is capped as DeltaL = max(500, DeltaL). The max operator places a floor, not a cap, on the length difference, which contradicts the stated purpose of preventing large reward fluctuations. If the intended operation is min(500, DeltaL), please correct the equation and any downstream derivation; if the max is intentional, please explain why a floor is needed and how the resulting denominator is used in the length reward.

minor comments (5)

[§2.2.1] The model name LLL-base-3B appears to be a typo for LLM-base-3B.
[§2.3.3, §3.5.1] The thinking-mode control token is typeset inconsistently across the paper, for example [ |BlueThink|], [|BlueThink|], and [BlueThink]; please unify the notation.
[§2.2.1, §3.1, Table 3] The pure-text pretraining token counts are given as 6T plus 3T in Section 2.2.1, 6.3T plus 3T in Section 3.1, and 9.3T in Table 3; please reconcile these numbers.
[References] References [54] and [55] both cite the Kimi-VL Technical Report; consolidate them to avoid duplicate citation.
[§5.1] The description of the in-house assessment model for long-chain reasoning reports speed and cost improvements but gives no scoring-agreement metric; please either provide the consistency number or state that it is unavailable.

Circularity Check

0 steps flagged · score 0.0 of 10

No definitional circularity: benchmark comparisons rest on external public evaluations.

full rationale

The paper's central claims are empirical benchmark scores, not derivations from the model's own equations or fitted parameters. Comparisons to Qwen3-4B, Kimi-VL-A3B-16B, and Qwen2.5-VL-72B are measured on public benchmarks (AIME24/25, LiveCodeBench, MMMU, MathVista, etc.) using OpenCompass, so the reported numbers are not equivalent to the training objective by construction. The length-penalty mechanism in Eq. (1) does reward shorter correct responses, but the accuracy-vs-token-budget curves in Figure 5 are measured outcomes; nothing in the reward formula forces a particular accuracy level, so the efficiency result is not a disguised restatement of the RL objective. Self-citations (BlueLM 2023, BlueLM-V-3B, PreSelect) are used for background or to describe methods employed, and none is invoked as a uniqueness theorem or as the sole justification for a contested premise. The remaining concerns identified in the manuscript — self-reported evaluations, qualitative decontamination descriptions, the in-house ScreenSpot vivo benchmark, and the small size of AIME sets — are verification and contamination risks rather than definitional circularity, and per the provided rules they belong to a correctness pass, not a circularity finding. No equation, fitted parameter, or citation chain reduces a prediction to its own input, so the appropriate score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The central claim rests on a set of hand-chosen training and evaluation choices plus assumptions about benchmark cleanliness. The length penalty α and cap, the data mixture ratio, and the think-token placement are selected by the authors; none are derived from first principles.

free parameters (6)

Length penalty base coefficient α = 0.2
Hand-chosen in §2.3.5 to balance inference length vs accuracy; used in Eq. (1).
Length difference cap ΔL = 500
Set in §2.3.5 to prevent drastic reward fluctuations; affects the 'Group Overlong' length reward.
Format and repetition penalty coefficients = 0.1
Rule-based reward multipliers in §2.3.5, chosen by hand for structured outputs.
Pure-text ratio in joint pre-training = 40%
Chosen in §2.2.2 and §3.2 to mitigate text capability degradation; a key data mixture decision.
Evaluation sampling temperature/top-p/top-k = 0.6 / 0.95 / 20
Set in §5, stated as consistent with Qwen3-4B thinking config; affects all reported scores.
SFT epochs for long CoT vs standard CoT = 9 / 3
Chosen in §2.3.1 to balance training and overfitting; no ablation shown.

assumptions (5)

domain assumption Logits-based distillation from a 7B teacher improves a pruned 3B student over training from scratch
Invoked in §2.2.1 to justify the LLM-base-3B initialization; based on prior results ([36]) and claimed in-house experiments, not re-derived here.
domain assumption Rule-based verifiable rewards for math/code also work for multimodal reasoning when the final answer is text
Stated in §2.3.5; the text reward model checks final output without image information, an assumption that may fail for answers requiring image context.
domain assumption OpenCompass evaluations with matched sampling settings yield fair comparisons to baselines
Assumed in §5; baseline numbers are treated as comparable, but the paper does not show it ran all baselines or that their reported numbers were produced under identical conditions.
ad hoc to paper Placing [|BlueThink|] at the end of the query yields stable mode activation
Empirical finding in §2.3.3, not derived; the paper gives no formal analysis of activation stability.
ad hoc to paper The length penalty schedule (linear decay with a floor) maintains training stability
Described in §2.3.5; the schedule is a design choice asserted to balance token efficiency and convergence.

invented entities (2)

[|BlueThink|] control token
purpose: Enables switching between thinking and non-thinking modes
Introduced in §2.3.1 and §3.5.1; its reliability (claimed <1 PPM failure) has no external falsifiable test, and the token's behavior is only demonstrated on the authors' own model.
ScreenSpot vivo (in-house Chinese GUI grounding benchmark)
purpose: Used to claim GUI superiority in Table 8
A proprietary benchmark introduced in §5.2.3; not public, so results cannot be checked or compared by independent groups.

how reviews work

0 comments

Cite this review

Pith. "Pith review of BlueLM-2.5-3B Technical Report." pith.science (2026). https://pith.science/paper/PJHK6O7W

@misc{pith2026250705934,
  author       = {Pith},
  title        = {Pith review of: BlueLM-2.5-3B Technical Report},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PJHK6O7W}},
  note         = {Machine review of arXiv:2507.05934}
}

read the original abstract

We present BlueLM-2.5-3B, a compact and unified dense Multimodal Large Language Model (MLLM) designed for efficient edge-device deployment, offering strong general-purpose and reasoning capabilities. To the best of our knowledge, this is the first 3B-scale MLLM to support both thinking and non-thinking modes, while also enabling explicit control over thinking token budget. BlueLM-2.5-3B is developed through diversified data curation, key data resampling, hybrid heterogeneous reinforcement learning, and a high-performance training infrastructure. Our model achieves superior multimodal capacity while preserving competitive pure-text performance with only 2.9 billion parameters. We conduct comprehensive evaluations across a broad range of multimodal and text-only benchmarks. In thinking mode, BlueLM-2.5-3B achieves comparable performance to Qwen3-4B on text-only benchmarks, and trails the larger Kimi-VL-A3B-16B by only about 5% on average across multimodal evaluations. In non-thinking mode, it outperforms Qwen2.5-VL-3B on the majority of multimodal benchmarks. Additionally, BlueLM-2.5-3B exhibits exceptional data efficiency. All of the aforementioned performance is achieved with substantially less total training data than Qwen2.5-VL-3B and Qwen3-4B. We hope our work contributes to the advancement of high-performance, on-device MLLMs and provides meaningful insights to the research community.

Figures

Figures reproduced from arXiv: 2507.05934 by the authors.

**Figure 2.** Model architecture of BlueLM-2.5-3B. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** ViT Inference Latency vs Token Number. Adapter Module To align visual representations with the language model input space, we utilize a two-layer multilayer perceptron (MLP) as the adapter module(also referred to as projection module). The MLP maps the image token space to the embedding space of the LLM. Additionally, a 2×2 spatial downsampling is applied to reduce the token length and enhance computational efficien… view at source ↗

Figures from the paper (10 more)

**Figure 4.** Figure 4: Training stages of BlueLM-2.5-3B. 2.2.3 Reasoning-Enhanced Stage To further enhance the model’s reasoning capabilities, we implemented reasoning-focused continued pre-training, targeting both textual and multimodal reasoning tasks. For the textual modality, we enriched…

**Figure 5.** Figure 5: Cumulative accuracy vs. Token Budget on the AIME2024 (left) and LiveCodeBench-v1 (right) benchmarks [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

**Figure 6.** Figure 6: Data Pipeline of BlueLM-2.5-3B. 3.7.1 Data Synthesis To enhance dataset coverage, particularly in data-scarce domains, we employ open-source models, including the latest versions of Qwen and DeepSeek, to generate synthetic data with controlled authenticity and consiste…

**Figure 7.** Figure 7: Multi-Sample Concatenation. Multiple image–text pairs are interleaved into a single composite sample per [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

**Figure 8.** Figure 8: Long-Context Training. Combined image and text embeddings are split into N equal-length subsequences [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]

**Figure 9.** Figure 9: Example of BlueLM-2.5-3B solving mathematical OCR problems with reasoning. [PITH_FULL_IMAGE:figures/full_fig_p034_9.png]

**Figure 10.** Figure 10: Example of BlueLM-2.5-3B solving bar-chart OCR comprehension problems. [PITH_FULL_IMAGE:figures/full_fig_p035_10.png]

**Figure 11.** Figure 11: Example of BlueLM-2.5-3B solving visual perception problems with reasoning. [PITH_FULL_IMAGE:figures/full_fig_p036_11.png]

**Figure 12.** Figure 12: Example of BlueLM-2.5-3B solving complex text reasoning problems. [PITH_FULL_IMAGE:figures/full_fig_p037_12.png]

**Figure 13.** Figure 13: Example of BlueLM-2.5-3B solving GUI agent problems. [PITH_FULL_IMAGE:figures/full_fig_p038_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

81 extracted references · 23 canonical work pages

[1]

Phi-3 technical report: A highly capable language model locally on your phone

Marah Abdin et al. “Phi-3 technical report: A highly capable language model locally on your phone”. In: arXiv preprint arXiv:2404.14219 (2024)

arXiv 2024
[2]

Phi-4 technical report

Marah Abdin et al. “Phi-4 technical report”. In: arXiv preprint arXiv:2412.08905 (2024)

arXiv 2024
[3]

Qwen2.5-VL Technical Report

Shuai Bai et al. Qwen2.5-VL Technical Report. 2025. arXiv: 2502.13923 [cs.CV]. url: https://arxiv.org/ abs/2502.13923

arXiv 2025
[4]

Paligemma: A versatile 3b vlm for transfer

Lucas Beyer et al. “Paligemma: A versatile 3b vlm for transfer”. In: arXiv preprint arXiv:2407.07726 (2024)

arXiv 2024
[5]

π0: A vision-language-action flow model for general robot control, 2024

Kevin Black et al. “ π0: A vision-language-action flow model for general robot control, 2024”. In: URL https://arxiv. org/abs/2410.24164 ()

arXiv 2024
[6]

BlueLM: An Open Multilingual 7B Language Model

BlueLMTeam. BlueLM: An Open Multilingual 7B Language Model . https://github.com/vivo- ai- lab/ BlueLM/blob/main/BlueLM_technical_report.pdf. Technical Report, vivo AI Lab. 2023

2023
[7]

Rt-2: Vision-language-action models transfer web knowledge to robotic control

Anthony Brohan et al. “Rt-2: Vision-language-action models transfer web knowledge to robotic control”. In: arXiv preprint arXiv:2307.15818 (2023)

arXiv 2023
[8]

Are We on the Right Way for Evaluating Large Vision-Language Models?

Lin Chen et al. “Are We on the Right Way for Evaluating Large Vision-Language Models?” In: arXiv preprint arXiv:2403.20330 (2024)

arXiv 2024

Show all 81 references

[9]

Evaluating Large Language Models Trained on Code

Mark Chen et al. Evaluating Large Language Models Trained on Code . 2021. arXiv: 2107.03374 [cs.LG]. url: https://arxiv.org/abs/2107.03374

2021 arXiv
[10]

Training Verifiers to Solve Math Word Problems

Karl Cobbe et al. “Training Verifiers to Solve Math Word Problems”. In: arXiv preprint arXiv:2110.14168 (2021)

2021 arXiv
[11]

Open R1: A fully open reproduction of DeepSeek-R1

Hugging Face. Open R1: A fully open reproduction of DeepSeek-R1 . Jan. 2025. url: https://github.com/ huggingface/open-r1

2025
[12]

Helix: A vision-language-action model for generalist humanoid control

AI Figure. “Helix: A vision-language-action model for generalist humanoid control”. In: Figure AI News (2024)

2024
[13]

Athene-70b: Redefining the boundaries of post-training for open models, July 2024

Evan Frick et al. “Athene-70b: Redefining the boundaries of post-training for open models, July 2024”. In: URL https://huggingface.com/Nexusflow/Athene-70B (2024)

2024
[14]

AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning

Wei Fu et al. AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning
[15]

The Llama 3 Herd of Models

Aaron Grattafiori et al. The Llama 3 Herd of Models . 2024. arXiv: 2407 . 21783 [cs.AI]. url: https : //arxiv.org/abs/2407.21783

2024 arXiv
[16]

HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models

Tianrui Guan et al. HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models . 2024. arXiv: 2310 . 14566 [cs.CV]. url: https : //arxiv.org/abs/2310.14566

2024 arXiv
[17]

Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model

Jingcheng Hu et al. Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model . 2025. arXiv: 2503.24290 [cs.LG]. url: https://arxiv.org/abs/2503.24290

2025 arXiv
[18]

LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code

Naman Jain et al. “LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code”. In: ArXiv abs/2403.07974 (2024). url: https://arxiv.org/abs/2403.07974

2024 arXiv
[19]

MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs

Ziheng Jiang et al. MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs . 2024. arXiv: 2402.15627 [cs.LG]. url: https://arxiv.org/abs/2402.15627

2024 arXiv
[20]

Experiment with gemini 2.0 flash native image generation, March 2025

Kat Kampf and Nicole Brichtova. “Experiment with gemini 2.0 flash native image generation, March 2025”. In: URL https://developers. googleblog. com/en/experiment-with-gemini-20-flash-native-image-generation/. Accessed (2025), pp. 04–27

2025
[21]

A diagram is worth a dozen images

Aniruddha Kembhavi et al. “A diagram is worth a dozen images”. In: European conference on computer vision. Springer. 2016, pp. 235–251

2016
[22]

Openvla: An open-source vision-language-action model

Moo Jin Kim et al. “Openvla: An open-source vision-language-action model”. In:arXiv preprint arXiv:2406.09246 (2024)

2024 arXiv
[23]

NuminaMath

Jia LI et al. NuminaMath. [https://huggingface.co/AI- MO/NuminaMath- CoT](https://github.com/ project-numina/aimo-progress-prize/blob/main/report/numina_dataset.pdf) . 2024

2024
[24]

Let’s Verify Step by Step

Hunter Lightman et al. “Let’s Verify Step by Step”. In: arXiv preprint arXiv:2305.20050 (2023)

2023 arXiv
[25]

Ring Attention with Blockwise Transformers for Near-Infinite Context

Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring Attention with Blockwise Transformers for Near-Infinite Context. 2023. arXiv: 2310.01889 [cs.CL]. url: https://arxiv.org/abs/2310.01889

2023 arXiv
[26]

Improved Baselines with Visual Instruction Tuning

Haotian Liu et al. “Improved Baselines with Visual Instruction Tuning”. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 . IEEE, 2024, 27 BlueLM-2.5-3B pp. 26286–26296. doi: 10.1109/CVPR52733.2024.02484. url: htt...

2024
[27]

Regmix: Data mixture as regression for language model pre-training

Qian Liu et al. “Regmix: Data mixture as regression for language model pre-training”. In: arXiv preprint arXiv:2407.01492 (2024)

2024 arXiv
[28]

MMBench: Is Your Multi-modal Model an All-around Player?

Yuan Liu et al. “MMBench: Is Your Multi-modal Model an All-around Player?” In: arXiv:2307.06281 (2023)

2023 arXiv
[29]

OCRBench: on the hidden mystery of OCR in large multimodal models

Yuliang Liu et al. “OCRBench: on the hidden mystery of OCR in large multimodal models”. In: Science China Information Sciences 67.12 (Dec. 2024). issn: 1869-1919. doi: 10.1007/s11432-024-4235-6 . url: http://dx.doi.org/10.1007/s11432-024-4235-6

2024 doi
[30]

SCP-116K: A High-Quality Problem-Solution Dataset and a Generalized Pipeline for Automated Extraction in the Higher Education Science Domain

Dakuan Lu et al. SCP-116K: A High-Quality Problem-Solution Dataset and a Generalized Pipeline for Automated Extraction in the Higher Education Science Domain . 2025. arXiv: 2501.15587 [cs.CL]. url: https://arxiv.org/abs/2501.15587

2025 arXiv
[31]

Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts

Pan Lu et al. “Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts”. In: arXiv preprint arXiv:2310.02255 (2023)

2023 arXiv
[32]

Bluelm-v-3b: Algorithm and system co-design for multimodal large language models on mobile devices

Xudong Lu et al. “Bluelm-v-3b: Algorithm and system co-design for multimodal large language models on mobile devices”. In: Proceedings of the Computer Vision and Pattern Recognition Conference . 2025, pp. 4145– 4155

2025
[33]

DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Scaling RL

Michael Luo et al. DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Scaling RL . https://pretty- radio- b75.notion.site/DeepScaleR- Surpassing- O1- Preview- with- a- 1- 5B- Model- by- Scaling- RL- 19681902c1468005bed8ca303013a4e2. Notion Blog. 2025

2025
[34]

An empirical model of large-batch training

Sam McCandlish et al. “An empirical model of large-batch training”. In: arXiv preprint arXiv:1812.06162 (2018)

2018 arXiv
[35]

A Theory on Adam Instability in Large-Scale Machine Learning

Igor Molybog et al. A Theory on Adam Instability in Large-Scale Machine Learning . 2023. arXiv: 2304.09871 [cs.LG]. url: https://arxiv.org/abs/2304.09871

2023 arXiv
[36]

Compact language models via pruning and knowledge distillation

Saurav Muralidharan et al. “Compact language models via pruning and knowledge distillation”. In: Advances in Neural Information Processing Systems 37 (2024), pp. 41076–41102

2024
[37]

Computer-using agent: Introducing a universal interface for ai to interact with the digital world

OpenAI. “Computer-using agent: Introducing a universal interface for ai to interact with the digital world”. In: (2025). url: https://openai.com/index/computer-using-agent

2025
[38]

GPT-4o System Card

OpenAI et al. GPT-4o System Card . 2024. arXiv: 2410.21276 [cs.CL]. url: https://arxiv.org/abs/2410. 21276

2024 arXiv
[39]

Vlp: Vision language planning for autonomous driving

Chenbin Pan et al. “Vlp: Vision language planning for autonomous driving”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 2024, pp. 14760–14769

2024
[40]

Yarn: Efficient context window extension of large language models

Bowen Peng et al. “Yarn: Efficient context window extension of large language models”. In: arXiv preprint arXiv:2309.00071 (2023)

2023 arXiv
[41]

MIA-Bench: Towards Better Instruction Following Evaluation of Multimodal LLMs

Yusu Qian et al. MIA-Bench: Towards Better Instruction Following Evaluation of Multimodal LLMs . 2025. arXiv: 2407.01509 [cs.CV]. url: https://arxiv.org/abs/2407.01509

2025 arXiv
[42]

UI-TARS: Pioneering Automated GUI Interaction with Native Agents

Yujia Qin et al. “UI-TARS: Pioneering Automated GUI Interaction with Native Agents”. In:CoRR abs/2501.12326 (2025). doi: 10.48550/ARXIV.2501.12326 . arXiv: 2501.12326. url: https://doi.org/10.48550/arXiv. 2501.12326

2025 doi arXiv
[43]

GPQA: A Graduate-Level Google-Proof Q&A Benchmark

David Rein et al. GPQA: A Graduate-Level Google-Proof Q&A Benchmark . 2023. arXiv: 2311.12022 [cs.AI]. url: https://arxiv.org/abs/2311.12022

2023 arXiv
[44]

Deepseekmath: Pushing the limits of mathematical reasoning in open language models

Zhihong Shao et al. “Deepseekmath: Pushing the limits of mathematical reasoning in open language models”. In: arXiv preprint arXiv:2402.03300 (2024)

2024 arXiv
[45]

HybridFlow: A Flexible and Efficient RLHF Framework

Guangming Sheng et al. “HybridFlow: A Flexible and Efficient RLHF Framework”. In: Proceedings of the Twentieth European Conference on Computer Systems . EuroSys ’25. ACM, Mar. 2025, pp. 1279–1297. doi: 10.1145/3689031.3696075. url: http://dx.doi.org/10.1145/3689031.3696075

2025
[46]

Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism

Mohammad Shoeybi et al. Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism. 2020. arXiv: 1909.08053 [cs.CL]. url: https://arxiv.org/abs/1909.08053

2020 arXiv
[47]

Predictive data selection: The data that predicts is the data that teaches

Kashun Shum et al. “Predictive data selection: The data that predicts is the data that teaches”. In: arXiv preprint arXiv:2503.00808 (2025)

2025 arXiv
[48]

Roformer: Enhanced transformer with rotary position embedding

Jianlin Su et al. “Roformer: Enhanced transformer with rotary position embedding”. In: Neurocomputing 568 (2024), p. 127063. 28 BlueLM-2.5-3B

2024
[49]

Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them

Mirac Suzgun et al. Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them . 2022. arXiv: 2210.09261 [cs.CL]. url: https://arxiv.org/abs/2210.09261

2022 arXiv
[50]

MathScale: Scaling Instruction Tuning for Mathematical Reasoning

Zhengyang Tang et al. MathScale: Scaling Instruction Tuning for Mathematical Reasoning . 2024. arXiv: 2403.02884 [cs.CL]. url: https://arxiv.org/abs/2403.02884

2024 arXiv
[51]

Gemini robotics: Bringing ai into the physical world

Gemini Robotics Team et al. “Gemini robotics: Bringing ai into the physical world”. In: arXiv preprint arXiv:2503.20020 (2025)

2025 arXiv
[52]

Gemma 3 Technical Report

Gemma Team et al. Gemma 3 Technical Report. 2025. arXiv: 2503.19786 [cs.CL]. url: https://arxiv. org/abs/2503.19786

2025 arXiv
[53]

Kimi k1.5: Scaling Reinforcement Learning with LLMs

Kimi Team et al. “Kimi k1.5: Scaling Reinforcement Learning with LLMs”. In: CoRR abs/2501.12599 (2025). doi: 10.48550/ARXIV.2501.12599 . arXiv: 2501.12599 . url: https://doi.org/10.48550/arXiv.2501. 12599

2025 doi arXiv
[55]

Kimi-vl technical report

Kimi Team et al. “Kimi-vl technical report”. In: arXiv preprint arXiv:2504.07491 (2025)

2025 arXiv
[56]

Drivevlm: The convergence of autonomous driving and large vision-language models

Xiaoyu Tian et al. “Drivevlm: The convergence of autonomous driving and large vision-language models”. In: arXiv preprint arXiv:2402.12289 (2024)

2024 arXiv
[57]

Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features

Michael Tschannen et al. “Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features”. In: arXiv preprint arXiv:2502.14786 (2025)

2025 arXiv
[58]

Measuring multimodal mathematical reasoning with math-vision dataset

Ke Wang et al. “Measuring multimodal mathematical reasoning with math-vision dataset”. In: arXiv preprint arXiv:2402.14804 (2024)

2024 arXiv
[59]

Self-Consistency Improves Chain of Thought Reasoning in Language Models

Xuezhi Wang et al. Self-Consistency Improves Chain of Thought Reasoning in Language Models . 2023. arXiv: 2203.11171 [cs.CL]. url: https://arxiv.org/abs/2203.11171

2023 arXiv
[60]

Yubo Wang et al.MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark
[61]

DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Under- standing

Zhiyu Wu et al. DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Under- standing. 2024. arXiv: 2412.10302 [cs.CV]. url: https://arxiv.org/abs/2412.10302

2024 arXiv
[62]

Data selection for language models via importance resampling

Sang Michael Xie et al. “Data selection for language models via importance resampling”. In: Advances in Neural Information Processing Systems 36 (2023), pp. 34201–34227

2023
[63]

Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments

Tianbao Xie et al. “Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments”. In: Advances in Neural Information Processing Systems 37 (2024), pp. 52040–52094

2024
[64]

Drivegpt4: Interpretable end-to-end autonomous driving via large language model

Zhenhua Xu et al. “Drivegpt4: Interpretable end-to-end autonomous driving via large language model”. In: IEEE Robotics and Automation Letters (2024)

2024
[65]

Qwen3 technical report

An Yang et al. “Qwen3 technical report”. In: arXiv preprint arXiv:2505.09388 (2025)

2025 arXiv
[66]

MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models

Longhui Yu et al. MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models . 2024. arXiv: 2309.12284 [cs.CL]. url: https://arxiv.org/abs/2309.12284

2024 arXiv
[67]

DAPO: An Open-Source LLM Reinforcement Learning System at Scale

Qiying Yu et al. DAPO: An Open-Source LLM Reinforcement Learning System at Scale . 2025. arXiv: 2503. 14476 [cs.LG]. url: https://arxiv.org/abs/2503.14476

2025 arXiv
[68]

RLAIF-V: Open-Source AI Feedback Leads to Super GPT-4V Trustworthiness

Tianyu Yu et al. “RLAIF-V: Open-Source AI Feedback Leads to Super GPT-4V Trustworthiness”. In: arXiv preprint arXiv:2405.17220 (2024)

2024
[69]

Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback

Tianyu Yu et al. “Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback”. In: arXiv preprint arXiv:2312.00849 (2023)

2023 arXiv
[70]

Mm-vet: Evaluating large multimodal models for integrated capabilities

Weihao Yu et al. “Mm-vet: Evaluating large multimodal models for integrated capabilities”. In: International conference on machine learning . PMLR. 2024

2024
[71]

Scaling Relationship on Learning Mathematical Reasoning with Large Language Models

Zheng Yuan et al. Scaling Relationship on Learning Mathematical Reasoning with Large Language Models
[72]

Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi

Xiang Yue et al. “Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 2024, pp. 9556–9567

2024
[73]

Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?

Yang Yue et al. “Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?” In: arXiv preprint arXiv:2504.13837 (2025)

2025 arXiv
[74]

Instruction-Following Evaluation for Large Language Models

Jeffrey Zhou et al. “Instruction-Following Evaluation for Large Language Models”. In: ArXiv abs/2311.07911 (2023). url: https://arxiv.org/abs/2311.07911. 29 BlueLM-2.5-3B

2023 arXiv
[75]

InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models

Jinguo Zhu et al. InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models. 2025. arXiv: 2504.10479 [cs.CV]. url: https://arxiv.org/abs/2504.10479. 30 BlueLM-2.5-3B A Appendix A.1 Contributors We would like to express our sincere gratitude...

2025 arXiv
[79]

**估算 √ 11的范围**: - 已知32 = 9和42 = 16。- 因此， √ 11位于3和4之间，即3 < √ 11 < 4。
[80]

34 BlueLM-2.5-3B Input Image Prompt: 回答图中选择题 Response: <think>

** 比较各点的位置**：- 点P位于2和3之间，显然小于 √ 11。- 点Q位于3和4之间，且更接近3。由于 √ 11 ≈ 3.3166，点Q正好落在这一区间内。- 点R位于4附近，明显大于 √ 11。综上所述，表示实数 √ 11的点可能是 Q Figure 9: Example of BlueLM-2.5-3B solving mathematical OCR problems with reasoning. 34 BlueLM-2.5-3B Input Image Prompt: 回答图中选择题 Response: <think>. . ...
[81]

** 姐妹总数**：- 如果爱丽丝有M 个姐妹（不包括她自己），那么整个家庭中就有M + 1 个姐妹（包括爱丽丝）。
[82]

37 BlueLM-2.5-3B T ask: 帮我搜索头戴式的蓝牙耳机，并筛选出价格区间在0到300的。 Figure 13: Example of BlueLM-2.5-3B solving GUI agent problems

** 兄弟视角下的姐妹数**：- 每个兄弟都会看到全部的家庭女性成员，因为这些女性都不是他/她的兄弟。- 所以，无论哪个兄弟看待，他都会看到M + 1 个姐妹。三、示例验证假设：- 爱丽丝有2 个兄弟（N = 2）；- 她有3 个姐妹（M = 3，不包括她自己）。那么：- 整个家庭中共有M + 1 = 4 个姐妹（包括爱丽丝）；- 每个兄弟也会看到这4 个姐妹。四、最终答案爱丽丝的兄弟有 M + 1 个姐妹。 Figure 12: Example of BlueLM-2.5-3B solving complex text reasoni...
[2023]

url: https://arxiv.org/abs/2308.01825

arXiv: 2308.01825 [cs.CL]. url: https://arxiv.org/abs/2308.01825

arXiv
[2024]

url: https://arxiv.org/abs/2406.01574

arXiv: 2406.01574 [cs.CL]. url: https://arxiv.org/abs/2406.01574

arXiv
[2025]

url: https://arxiv.org/abs/2505.24298

arXiv: 2505.24298 [cs.LG]. url: https://arxiv.org/abs/2505.24298

arXiv

Pith tools

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

[1] [1]

Phi-3 technical report: A highly capable language model locally on your phone

Marah Abdin et al. “Phi-3 technical report: A highly capable language model locally on your phone”. In: arXiv preprint arXiv:2404.14219 (2024)

arXiv 2024

[2] [2]

Phi-4 technical report

Marah Abdin et al. “Phi-4 technical report”. In: arXiv preprint arXiv:2412.08905 (2024)

arXiv 2024

[3] [3]

Qwen2.5-VL Technical Report

Shuai Bai et al. Qwen2.5-VL Technical Report. 2025. arXiv: 2502.13923 [cs.CV]. url: https://arxiv.org/ abs/2502.13923

arXiv 2025

[4] [4]

Paligemma: A versatile 3b vlm for transfer

Lucas Beyer et al. “Paligemma: A versatile 3b vlm for transfer”. In: arXiv preprint arXiv:2407.07726 (2024)

arXiv 2024

[5] [5]

π0: A vision-language-action flow model for general robot control, 2024

Kevin Black et al. “ π0: A vision-language-action flow model for general robot control, 2024”. In: URL https://arxiv. org/abs/2410.24164 ()

arXiv 2024

[6] [6]

BlueLM: An Open Multilingual 7B Language Model

BlueLMTeam. BlueLM: An Open Multilingual 7B Language Model . https://github.com/vivo- ai- lab/ BlueLM/blob/main/BlueLM_technical_report.pdf. Technical Report, vivo AI Lab. 2023

2023

[7] [7]

Rt-2: Vision-language-action models transfer web knowledge to robotic control

Anthony Brohan et al. “Rt-2: Vision-language-action models transfer web knowledge to robotic control”. In: arXiv preprint arXiv:2307.15818 (2023)

arXiv 2023

[8] [8]

Are We on the Right Way for Evaluating Large Vision-Language Models?

Lin Chen et al. “Are We on the Right Way for Evaluating Large Vision-Language Models?” In: arXiv preprint arXiv:2403.20330 (2024)

arXiv 2024