Skywork-R1V3: Open-Source Vision-Language Model

• Skywork-R1V3 is an open-source vision-language model that integrates reinforcement learning and connector modules for effective cross-modal alignment and visual reasoning.
• It employs a modular architecture combining a ViT-style image encoder, a large language decoder, and a dedicated connector to transform visual tokens.
• Empirical benchmarks reveal state-of-the-art accuracy with superior generalization, validated through reward model analyses and critical-token entropy measures.

Skywork-R1V3 is an open-source vision-LLM (VLM) designed for advanced visual reasoning, notable for integrating sophisticated reinforcement learning (RL) post-training protocols to transfer reasoning skills from LLMs into the visual domain. The architecture leverages a connector module for robust cross-modal alignment, introduces entropy-based reasoning capability indicators, and is supported by a reward-model interpretability analysis that illuminates both its internal mechanisms and alignment properties (Shen et al., 8 Jul 2025, Nadaf, 28 Apr 2026).

1. Model Architecture and Cross-Modal Alignment

Skywork-R1V3 employs a modular architecture built on three principal components:

• Vision Encoder: InternViT-6B-448px-V2.5, a ViT-style image encoder, maps the input image $I$ to a sequence of visual tokens $z_v$.
• Language Decoder: A 32B-parameter QwQ-32B LLM processes chat histories and visual features.
• Connector Module: A two-layer MLP with non-linearity and layer norm, implements the transformation

$$h_v = \mathrm{LN}(W_2\,\sigma(W_1\,z_v + b_1) + b_2),\quad h_v \in \mathbb{R}^{T \times d_\mathrm{model}}$$

and serves as the alignment bridge between visual token space and the LLM's token embedding space.

The cross-modal flow proceeds in three steps:

1. Visual input $I$ is encoded as tokens $z_v = \mathrm{ViT}(I)$.
2. The connector produces projected embeddings $h_v$ from $z_v$.
3. The decoder's cross-attention layer ingests both text tokens and $h_v$ (as keys/values $K_v = V_v = h_v$), facilitating effective joint processing.

Empirical analysis demonstrates that freezing the connector during RL post-training causes reward curve collapse, while freezing the image encoder yields minor performance impact, indicating the connector's centrality to multimodal reasoning performance.

2. Reinforcement Learning Post-Training and Critical-Token Entropy

Post-training RL in Skywork-R1V3 is central to reifying and enhancing visual reasoning:

• Objective Function:

$$\max_\theta J(\pi_\theta) = \mathbb{E}_{x\sim\mathcal{D},\,y\sim\pi_\theta(\cdot|x)}[r(x,y)]$$

• Reward Design:
  • Accuracy: $z_v$0 if the model's answer matches the ground truth, validated by a rule-based or LM verifier.
  • Format: $z_v$1 if the response adheres to the "> …" chain-of-thought template.
  • Final Reward:

  $z_v$2

Policy improvement uses both the PPO surrogate loss:

$z_v$3

and a group-normalized reward version (GRPO), which normalizes advantages across groups for variance control.

• Critical-Token Entropy $z_v$4:

$z_v$5

computed on the token after "<think>". RL checkpoints showing higher $z_v$6 consistently correspond to higher visual reasoning accuracy, and model selection tracks peaks in $z_v$7 rather than purely reward values.

3. Performance Benchmarks and Generalization

Skywork-R1V3 achieves state-of-the-art accuracy on several multimodal reasoning tasks, most notably on the MMMU benchmark:

System	MMMU (%)	MathVista (%)	LogicVista (%)	PhyX (%)
Skywork-R1V3-38B	76.0	77.1	59.7	52.8
GPT-4o	70.7	71.4	64.4	43.8
Claude 3.7 Sonnet	75.0	—	—	—
Open-source baseline	64.3	—	—	—
Entry-level human	≈76	—	—	—

Skywork-R1V3-38B, through connector tuning and RL post-training, matches entry-level human performance on MMMU and matches or exceeds top closed-source VLMs across several tasks, despite being entirely open-source.

Experiments indicate robust reasoning skill transfer from text-only models to visual domains, with RL finetuning yielding superior in-domain (77.2%) and out-of-domain (74.5%) generalization compared to SFT alone, which suffers a larger generalization gap (75.9% vs. 65.4%).

4. Mechanistic Interpretability: Reward Model Analysis

The reward-lens diagnostic suite reveals the internal dynamics of Skywork-R1V3’s reward model, clarifying the relationship between residual stream components, causal contributions, and reward head geometry (Nadaf, 28 Apr 2026):

• Reward Head Formula:

$z_v$8

where $z_v$9 is the learned reward vector.

• Component Attribution decomposes reward into per-layer/ per-component contributions ($$h_v = \mathrm{LN}(W_2\,\sigma(W_1\,z_v + b_1) + b_2),\quad h_v \in \mathbb{R}^{T \times d_\mathrm{model}}$$0), typically finding that final-layer MLPs dominate observationally: e.g., mlp_L31 (199/200 top-1 in helpfulness).

• Activation Patching reveals that early layers are causally critical: e.g., top δ_r for mlp_L0, mlp_L1, attn_L0.

• Faithfulness Gap:

Observational attribution ($$h_v = \mathrm{LN}(W_2\,\sigma(W_1\,z_v + b_1) + b_2),\quad h_v \in \mathbb{R}^{T \times d_\mathrm{model}}$$1) correlates negatively with causal effect ($$h_v = \mathrm{LN}(W_2\,\sigma(W_1\,z_v + b_1) + b_2),\quad h_v \in \mathbb{R}^{T \times d_\mathrm{model}}$$2), with mean Spearman ρ = –0.256. Final-layer MLPs have high apparent reward alignment but low necessity; early layers are load-bearing for reward causation.

• Concept Alignment: Abstract axes like agreement, verbosity, helpfulness, and formality all show significant projection onto $$h_v = \mathrm{LN}(W_2\,\sigma(W_1\,z_v + b_1) + b_2),\quad h_v \in \mathbb{R}^{T \times d_\mathrm{model}}$$3, e.g., cos(v_agreement, w_r/∥w_r∥) = 0.340, cos(v_helpfulness, w_r/∥w_r∥) = 0.290, and so on.

Hacking detector probes confirm strong penalization of sycophancy (d = –5.82), confidence (d = –2.08), and formatting (d = –0.95), indicating a policy disincentivized for these behaviors.

5. Training Strategies and Curriculum Learning

The Skywork-R1V3 training process integrates several strategy elements:

• Cold-start SFT employs 12K distilled “think-style” chains-of-thought from Skywork-R1V2 to bootstrap the reasoning skill transfer.

• Curriculum Learning: Two-stage protocol—Stage 1 samples medium-difficulty (K12 “Normal”), Stage 2 escalates to hard samples. Notably, distribution shift in Stage 2 diminishes generalization, leading to a validation accuracy drop, revealing sensitivity to curriculum sequencing.

• Out-of-domain Robustness: RL-finetuned models generalize more reliably to distributional shifts than SFT-only models, indicating RL is key to robust multimodal reasoning acquisition.

6. Alignment, Failure Modes, and Interpretability Tools

Reward-lens provides insights into failure modes and alignment dynamics:

• Distortion Index: Quantifies coverage gaps in reward-aligned quality dimensions.

• Misalignment Cascade: Skywork-R1V3 demonstrates tightly correlated misalignment axes (systemic-risk score 1.00 across six dimensions).

• Practical Findings:

  • Observational attribution must be supplemented with activation patching to validate layer necessity.
  • Early layers (attn_L0–L2, mlp_L0) are mechanistically critical; monitoring these provides early signals of misalignment cascades.
  • The uniform circuit structure across reward axes implies greater vulnerability to correlated failure modes compared to specialized multi-objective heads (e.g., ArmoRM).
• Hackability: Reward-lens flags that Skywork-R1V3’s reward model may be susceptible to policies exploiting agreement or verbosity axes unless proactively mitigated.

7. Limitations, Open Challenges, and Future Directions

Several open challenges and future research directions are identified:

• Hallucination and Chain-of-Thought Failures: False “I can’t see the image” responses and unreliable sustained CoT reasoning persist as performance bottlenecks.
• Proposed Extensions:
  • End-to-end RL finetuning for tool use, code execution, image cropping.
  • Unified vision-understanding and vision-generation tasks.
  • Physics-based grounding and embodied agent integration for richer interaction.
• Interpretability for Safety: Editing or regularizing the reward vector $$h_v = \mathrm{LN}(W_2\,\sigma(W_1\,z_v + b_1) + b_2),\quad h_v \in \mathbb{R}^{T \times d_\mathrm{model}}$$4, e.g., by amplifying honesty or penalizing specific concept axes, constitutes a promising avenue for safer model deployment.

Reward-lens’s negative φ–δ_r correlation is interpreted not as a tool defect but as a property arising from residual-stream redundancy and the intertwining of multiple misalignment dimensions. Continuous monitoring and targeted intervention—especially in early Transformers layers—are recommended for robust and aligned multimodal reasoning models (Shen et al., 8 Jul 2025, Nadaf, 28 Apr 2026).