GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning

Training framework

• The training framework combines broad multimodal pre-training, domain-specific supervised fine-tuning, and RLCS, which uses curriculum learning and difficulty-aware sampling for cross-domain reasoning.Pre-training includes factual image-text pairs, interleaved academic image-text data, annotated documents and diagrams, instructional videos, and grounding data.Source: §1 · p. 2

Results

• Across 42 public benchmarks, GLM-4.5V achieves state-of-the-art performance on nearly all tasks and comparable or superior results on 22 versus Gemini-2.5-Flash; GLM-4.1V-9B-Thinking is competitive or superior to Qwen2.5-VL-72B on 29 benchmarks.GLM-4.5V advances comparable-size open-source VLMs by roughly 10% or more across a wide range of tasks.Source: §1 · p. 2

Reinforcement-learning findings

• Multi-domain reinforcement learning supports cross-domain generalization: training on one domain boosts others, while joint training yields greater improvements in each.The authors describe this as mutual facilitation across domains.Source: §1 · p. 2
• Selecting informative rollout problems dynamically is essential for efficiency and performance, while weak reward signals for even one capability can collapse unified multi-domain RL.The proposed strategies include RLCS, ratio-based Exponential Moving Average sampling expansion, and domain-specific reward systems.Source: §1 · pp. 2–3

Contributions

• The released model family includes 9B and 106B-A12B models, with GLM-4.5V and GLM-4.6V supporting native thinking and non-thinking modes for flexible performance-efficiency trade-offs.The authors also open-source GLM-4.1V-9B-Base and components including domain-specific reward systems.Source: §1 · pp. 2–3

Architecture

• The shared GLM-V architecture combines an AIMv2-Huge [9] vision encoder, an MLP adapter, and model-specific GLM language decoders [13].GLM-4.1V-Thinking and GLM-4.6V-Flash use GLM-4-9B-0414; GLM-4.5V and GLM-4.6V use GLM-4.5-Air. The encoder uses 3D convolutions for 2× video temporal downsampling and duplicates single images for consistency.Source: §2 · p. 3, Figure 2 · p. 4

Image resolution and position embeddings

• 2D-RoPE [44] and bicubic-interpolated absolute position embeddings adapt the ViT to variable resolutions, extreme aspect ratios, and images beyond 4K.Patch coordinates are normalized to a continuous grid and used to sample the original position-embedding table.Source: §2 · p. 3

Language-side spatial encoding

• 3D-RoPE in the LLM strengthens spatial understanding of multimodal contexts while preserving the original model’s text-related capabilities.Source: §2 · p. 3

Video temporal modeling

• Video inputs receive a timestamp-string token after each frame, conveying real-world timing and temporal distances to support temporal understanding and grounding.The design treats video frames as a temporally coherent sequence rather than as multi-image inputs.Source: §2 · p. 3, Figure 2 · p. 4

3 Pre-training

• On MathVista’s non-multiple-choice subset, GLM-4.1V-9B-Base significantly outperforms a similar-scale state-of-the-art pretrained base model in pass@k, which the authors say arguably sets the RL ceiling.Pre-training combines academic corpora, knowledge-rich interleaved image-text data, and pure text to preserve language capabilities.Source: §3 · p. 4

3.1 Pre-training Data

• Over 10 billion image-text pairs seed a caption-curation pipeline that filters low-quality and mismatched samples, balances concepts, and produces factual, information-rich recaptions.The curated originals and recaptioned data are merged at a predetermined ratio to balance broad world knowledge with descriptive depth.Source: §3.1 · p. 5
• Specialized web and academic-book pipelines extract high-quality interleaved image-text content to address noisy alignment and information density and strengthen foundational understanding and reasoning.The pipelines use contextual relevance filtering and high-knowledge-density image selection for web data, and collect and deeply parse over 100 million digitized books focused on domains including STEM.Source: §3.1 · pp. 5–6
• 220 million OCR images combine synthetic documents, natural-scene text with bounding boxes, and paired academic-document page renderings and structured markup.The three components target varied text layouts, authentic real-world text, and academic document recognition, respectively.Source: §3.1 · p. 6
• Grounding data covers natural images and GUIs, yielding 40 million natural-image annotations and over 140 million GUI question-answer pairs for referring-expression tasks.Natural-image boxes are generated with GLIPv2 and filtered to retain samples with at least two valid boxes; GUI data uses webpage screenshots and visible DOM elements with rendered boxes.Source: §3.1 · p. 6
• Video data combines diverse sources with fine-grained human annotation of actions and in-scene text, plus human-in-the-loop annotation of camera motion and shot composition.Integrity checks and multimodal embedding-based deduplication remove invalid files and semantically redundant video-text pairs.Source: §3.1 · pp. 6–7

Instruction tuning data.

• 50 million instruction-tuning samples span visual understanding, multimodal reasoning, documents, GUI operations, and UI coding, supporting the full-scenario RL pipeline.The data uses a task taxonomy and balanced sampling, synthetic complex-scenario augmentation, and manual and automated contamination reviews.Source: Instruction tuning data · p. 7

3.2 Training Recipe

• Training uses multimodal pre-training followed by long-context continual training, adding video, high-resolution imagery, and extended-context data in the later stage.The first stage runs for 120,000 steps at sequence length 8,192; continual training adds 10,000 steps at 32,768 tokens, and GLM-4.6V receives a further 2,000 steps to extend its context to 131,072 tokens.Source: §3.2 · p. 7

4 Supervised Fine-Tuning

• Unlike prior short-CoT SFT workflows [59] [20] [16], SFT aligns existing vision-language understanding to a more effective thinking and response style rather than injecting knowledge.This long-CoT bridge from pre-training to RL is intended to prime a stronger cold start and make subsequent RL more efficient and stable.Source: §4 · p. 7

4.1 Supervised Fine-Tuning Data

• The long-CoT dataset emphasizes verifiable tasks, adds non-verifiable visual question answering, is primarily Chinese and English, and filters examples to moderate difficulty.It is curated to train coherent, multi-step solutions in a standardized format for stable, scalable RL.Source: §4.1 · p. 8
• Reasoning responses use <think>/<answer> tags, with one boxed final answer for verifiable tasks; GLM-4.6V tool calls use an XML-like <tool_call> key/value structure.The tool protocol gives parsers discrete argument boundaries and supports deterministic mapping of intents to executable calls without tokenizer changes for each parameter key.Source: §4.1 · p. 8
• The cleaning pipeline enforces response formatting and removes inconsistent or noisy reasoning, mixed-language phrasing, and redundant thoughts because poor data can destabilize or collapse RL.The authors report that clean, consistent data improves RL convergence and overall performance.Source: §4.1 · p. 8, §4.2 · p. 9
• High-quality, informative examples sampled from RL checkpoints are fed back into the cold-start dataset to incorporate useful reasoning patterns into later RL rounds.This iterative enhancement is intended to improve dataset quality and challenge level.Source: §4.1 · p. 8

4.2 Training Recipe

• Clean, consistent cold-start reasoning data yields more stable RL convergence and higher overall performance than noisy data, although RL remains effective with noisy traces.Noisy traces can still provide useful guidance, but may contain formatting inconsistencies or repetitive patterns.Source: §4.2 · p. 9
• GLM-4.5V and GLM-4.6V mix thinking and non-thinking training data, using /nothink prompts to elicit empty reasoning; GLM-4.1V-Thinking is a pure thinking model.For non-thinking examples, reusing answers from thinking examples performed better than building a separately curated subset. The recipe uses full-parameter tuning with sequence lengths of 32,768 tokens (131,072 for GLM-4.6V) and global batch size 32, alongside text-only long-form examples to preserve language and reasoning abilities.Source: §4.2 · pp. 8–9

5 Reinforcement Learning: What Is Challenging and What Works

• After supervised fine-tuning, the authors combine RLVR and RLHF across multimodal capabilities, and warn that failure in any RL component can severely degrade efficiency or collapse training.The framework spans STEM, grounding, OCR, video, GUI agents, chart and document understanding, logical reasoning, and instruction following.Source: §5 · p. 9

5.1 Data Preparation

• Data preparation identifies verifiable tasks, curates assessable question-answer pairs, grades difficulty using passk evaluations and human labels, and pilots RL in each subdomain.Converting uniquely answered multiple-choice questions to fill-in-the-blank format is used to reduce random-guessing noise.Source: §5.1 · p. 10

5.2 Reward System

• A weak verifier in one capability can collapse unified VLM training across domains, even when STEM rewards are high; flawed multi-image QA rewards coincided with broad benchmark declines.The authors report reward noise and reward hacking in other image tasks, followed by stalled STEM reward growth and declines in STEM-related benchmarks.Source: §5.2 · p. 10, Figure 5 · p. 11
• The reward system replaces unreliable LLM-based answer extraction with dedicated box tokens and uses task-specific verification because answer formats and correctness criteria vary across domains.For example, “43” and “43.0” are equivalent in math but not OCR; coarse or incomplete verifiers can also be hacked to earn reward without improving task performance.Source: §5.2 · pp. 10–11

5.3 Reinforcement Learning with Curriculum Sampling (RLCS)

• After 200 training steps, over half of prompts achieved accuracy over 90%, motivating RLCS to continually favor informative mid-difficulty samples using offline and rollout-based difficulty tiers.RLCS combines pass@k results and human difficulty annotations, then down-samples trivial and currently too-challenging examples while boosting mid-range difficulties; the authors report accelerated improvement and consistent gains.Source: §5.3 · p. 12

5.3.1 Improving Effectiveness

• RL training enhancements include EMA-based oversampling for balanced rollouts and forced answer generation when reasoning is truncated.The authors also recommend larger batches, remove KL loss because it constrained capabilities, and report that cold-start checkpoints with different initial scores converged to nearly the same post-RL performance; cross-domain transfer was frequently observed.Source: §5.3.1 · p. 13

5.3.2 Improving Stability

• Removing entropy loss, sampling with top-p = 1, and averaging loss per sample improved training stability, while poor cold-start data or frequent format errors could destabilize training.Per-sample loss yielded more stable training despite no significant difference in mean reward; top-p = 1 eliminated garbled outputs that tended to appear in later iterations.Source: §5.3.2 · p. 14

5.4 Infrastructure

• Infrastructure balances sequence and compute loads across data-parallel ranks, packs variable-length samples into training micro-steps, and repacks samples to reduce wasted work.The repacking optimization halves forward-backward time; micro-steps use a fixed context_length of 32K.Source: §5.4 · p. 14

6.1 Evaluation Setting

• Evaluation covers 42 public benchmarks in eight categories, with a common toolchain, policies, and prompt templates and at least a 95% successful-request rate per model and benchmark.Image and video inputs have maximum expected lengths of 6,144 and 48,000 tokens, respectively; responses are capped at 8,192 tokens, and failed requests are excluded from scoring.Source: §6.1 · p. 16

6.2 Comparison to Other Advanced MLLMs

• Against a wide range of open-source state-of-the-art MLLMs, GLM-4.5V establishes a new state-of-the-art among open-source models across all benchmarks, demonstrating consistent superiority across a wide spectrum of multimodal tasks.Source: §6.2 · p. 16
• GLM-4.5V-Thinking surpasses comparable-size open-source models in General VQA and leads on the cited STEM, chart, document, coding, GUI-agent, video, spatial-reasoning, and grounding benchmarks.It leads on MMMU (Val), MMMU Pro, AI2D, MathVista, WeMath, ChartQAPro, ChartMuseum, MMLongBench-Doc, VideoMME, MMVU, OminiSpatial, and RefCOCO; OCRBench remains slightly behind the non-thinking version and Kimi-VL.Source: §6.2 · pp. 16–17
• GLM-4.1V-9B-Thinking outperforms Qwen2.5-VL-72B on 29 of 42 benchmarks, including MMStar, MUIRBENCH, MMMU Pro, and ChartMuseum.The paper presents this as a performance-efficiency trade-off for resource-constrained deployment.Source: §6.2 · p. 17

6.3 Investigating Cross-Domain Generalization in Reinforcement Learning

• Single-domain RL improves performance in other domains, and GUI-agent RL improves all evaluated domains, including general VQA without a corresponding training set.The experiments train from the SFT-stage checkpoint and compare four individual domains with a Mix-all setting at matched per-domain sample counts.Source: §6.3 · pp. 17–18
• Mix-all RL outperforms single-domain RL in STEM, OCR & Chart, and general VQA, but not in grounding or GUI-agent performance.The authors note that grounding and GUI-agent tasks may need more targeted or specialized multi-domain strategies.Source: §6.3 · p. 18, Figure 6 · p. 18
• Cross-domain results show reciprocal transfer between GUI-agent and grounding tasks and between OCR & Chart and GUI-agent tasks.The paper relates these effects to shared reliance on visual grounding and accurate text recognition, respectively.Source: §6.3 · p. 18

Limitations and Future Work

• Outcome-based rewards can reinforce flawed reasoning, motivating reward models that assess intermediate steps, detect hallucinations and logical inconsistencies, and reduce reward hacking on subjective tasks.Current reward models typically assess final outcomes without verifying intermediate reasoning; subjective tasks need strategies to prevent reward hacking.Source: §7 · pp. 18–19
• Minor setup changes caused substantial variation in reasoning depth or output style; improved rewards and cold-start data have stabilized later training, but RL still needs greater consistency and robustness.The remaining sensitivity points to challenges in large-scale RL optimization.Source: §7 · p. 19
• Cluttered images, occlusion, or ambiguous visual details can cause perceptual errors that lead to guesswork rather than grounded inference, so perception and reasoning must improve together.These components are described as intricately interconnected.Source: §7 · p. 19
• The authors propose investigating whether multimodal training, such as interpreting code in images, can improve text-only coding and how vision and language may reinforce general reasoning.They identify this as a promising research direction, not an established result.Source: §7 · p. 19
• As model capabilities improve, benchmarks should become more challenging and diagnostic, explicitly detecting failure modes such as shortcut reasoning and hallucination in reasoning chains.The authors note that many current benchmarks are nearing saturation or fail to identify critical errors.Source: §7 · p. 19

• The section demonstrates the model’s capabilities and advantages through qualitative examples.Source: A Qualitative Examples · p. 27

A.1 UI Code Generation

• Figure 7 shows a case of generating front-end code from a UI snapshot.Source: Figure 7 · p. 27

• Figure 8 presents a case demonstrating detailed, precise video description.Source: Figure 8 · p. 28

A.3 Video Description (in Chinese)

• A case illustrates the model’s ability to give detailed and precise descriptions of a video.Source: Figure 9 · p. 29

A.4 Video QA

• The video QA example illustrates answering a video-related question through perception, knowledge, and reasoning.Source: Figure 10 · p. 30

A.5 GUI Agent; A.6 Chart QA

• The GUI Agent and Chart QA examples illustrate GUI recognition and operation, and chart understanding and question answering.Source: Figure 11 · p. 31, Figure 12 · p. 32

A.7 Geolocation; A.8 OCR + Coding

• The Geolocation and OCR + Coding examples illustrate inferring geographic position from an image and recognizing, debugging, and correcting code.Source: Figure 13 · p. 33, Figure 14 · p. 34

A.9 Chemistry Problem Solving; A.10 Math Problem Solving

• The Chemistry and Math examples illustrate solving chemistry problems and complex math problems.Source: Figure 15 · p. 35, Figure 16 · p. 36

A.11 Long Document Understanding; A.12 Spatial Reasoning; A.13 Visual Grounding; A.14 Tool Using

• The remaining examples illustrate long-document interpretation, spatial reasoning, visual grounding, and tool use with visual reasoning.Source: Figure 17 · p. 37, Figure 18 · p. 38, Figure 19 · p. 39

B.1 Evaluation Protocol of VLM Coding

• For VLM coding evaluation, the protocol follows Design2Code’s direct setting without text augmentation or self-revision and uses GPT-o4-mini, rather than CLIP, to judge rendered HTML against reference UI screenshots.Accuracy is the proportion of test cases scoring at least 80; manual evaluation found scores at or above this threshold generally visually and functionally consistent, while resolution and font differences make a perfect score impractical.Source: §B.1 · p. 41

B.2 GUI Agent Instructions

• Reproducing the paper’s GUI Agent results on phones and computers requires following the model-specific prompts and settings in the provided instructions.Source: §B.2 · p. 41

B.3 Visual Grounding Instructions

• Visual grounding prompts should request a target’s bounding box; outputs use [x1,y1,x2,y2] coordinates normalized to image width or height and scaled by 1000.The target is specified by <expr>, and bracket styles may vary.Source: §B.3 · p. 41

C Experimental Results of GLM-4.1V-9B-Thinking

• GLM-4.1V-9B-Thinking sets a new state of the art on 23 of 28 benchmarks among models under 10B parameters.The comparison includes MiMo-VL, Kimi-VL, InternVL3 [72], and the Qwen-VL series [4].Source: C Experimental Results of… · p. 41, Table 3 · p. 42