P1-VL: Vision-Language Models for Physics

• P1-VL is an open-source family of vision-language models designed for Olympiad-level physics, merging visual diagrams with text to capture spatial and physical constraints.
• It employs a multimodal backbone with a frozen image encoder and advanced curriculum reinforcement learning to align symbolic reasoning with physical reality.
• Agentic augmentation via PhysicsMinions refines inference by iteratively parsing, critiquing, and verifying solutions, enhancing the model's structured problem-solving capability.

Searching arXiv for recent and related papers on P1-VL, HiPhO, and PhysicsMinions to ground the article in the latest literature. P1-VL, short for Physics-1 Vision-Language, is an open-source family of vision-LLMs engineered for advanced scientific reasoning in Olympiad-level physics. It is designed around the premise that physics at this level cannot be reduced to symbolic manipulation alone: diagrams are often constitutive rather than illustrative, encoding boundary conditions, spatial symmetries, and other constraints absent from text. In this formulation, multimodal perception is treated as a prerequisite for maintaining physical consistency with the laws governing the universe. The system combines a high-capacity multimodal backbone with Curriculum Reinforcement Learning and test-time Agentic Augmentation, and is evaluated primarily on HiPhO, a benchmark composed of 13 recent physics Olympiad exams from 2024–2025 (Luo et al., 10 Feb 2026).

1. Problem setting and conceptual scope

P1-VL is situated at what its defining paper describes as the transition from symbolic manipulation to science-grade reasoning. Physics is used as the critical test anchor because it requires models to bind abstract logic to physical reality rather than merely continue text patterns. The paper’s central claim is that this binding is inherently multimodal: at Olympiad level, figures may encode essential constraints that are not recoverable from the written statement alone (Luo et al., 10 Feb 2026).

This framing distinguishes P1-VL from systems that treat visual input as auxiliary evidence. A common misconception in text-centric reasoning benchmarks is that the diagram is optional context. P1-VL is explicitly built against that assumption. Its motivating examples are problems in which geometry, alignment, boundary conditions, and symmetry relations are carried by the image rather than the prose. This suggests that the model is intended not simply as a generic VLM applied to physics, but as a system whose training and inference procedures are organized around the visual-logical gap specific to scientific problem solving.

The paper further positions Olympiad physics as a stringent evaluation regime for general-purpose physical intelligence. That claim is tied to the scoring setting: HiPhO mirrors human grading and assigns partial credit for reasoning, so the target is not merely answer selection but structured, physically coherent solution generation (Luo et al., 10 Feb 2026).

2. Model family and architectural organization

The released family contains two variants: P1-VL-30B-A3B, described as approximately 30 B language-model parameters with a 3 B-expert MoE adapter, and P1-VL-235B-A22B, described as approximately 235 B language-model parameters with a 22 B-expert MoE adapter. Both variants are built around a backbone vision-language encoder in which a frozen pre-trained image encoder, exemplified in the paper by Qwen3-VL, is connected to a large Transformer LLM through linear projection layers that bridge high-dimensional image embeddings into the decoder stack (Luo et al., 10 Feb 2026).

The language decoder is a large Transformer LM in either the 30 B or 235 B configuration and supports a “Thinking” prefine tuned instruction style. During reinforcement learning, only the language-model and MoE layers are trainable; the vision encoder and projection layers remain frozen in order to preserve pre-trained alignment. That design choice makes the architecture asymmetric: visual representation is inherited and stabilized, while physics-specific post-training is concentrated in the generative and expert-routing components.

This organization is technically significant because it separates multimodal grounding from domain-specialized reasoning. The paper does not present this as a purely architectural novelty in isolation; rather, the backbone is the substrate on which the specialized post-training pipeline operates. A plausible implication is that P1-VL’s reported gains derive less from replacing the visual front end than from aligning a strong existing VLM stack to the demands of scientific reasoning under physics-specific reward signals and curricula.

3. Curriculum Reinforcement Learning

P1-VL formulates physics problem solving as a Markov decision process,

$\mathcal{M}=(\mathcal S,\mathcal A,P,r),$

where $\mathcal S$ consists of text and image context plus generated tokens, $\mathcal A$ is the next-token vocabulary, and the reward is binary at the sub-answer level, aggregated as

$R=\frac{1}{N}\sum_{i=1}^{N} r_i.$

The training objective is

$$J(\pi_\theta)=\mathbb E_{\tau\sim\pi_\theta}\left[\sum_t r(s_t,a_t)\right].$$

Optimization is performed with Group Sequence Policy Optimization (GSPO), using length-normalized importance ratios

$$\rho_i(\theta)=\exp\Bigl(\tfrac1{|y_i|}\sum_{t=1}^{|y_i|}\log\frac{\pi_\theta(y_{i,t}|x,y_{i,<t})}{\pi_{\theta_{\rm old}(\cdot)}}\Bigr),$$

together with sequence-level clipped gradients. To stabilize off-policy updates, the method introduces Sequence-Level Masked Importance Sampling (Seq-MIS),

$$\nabla_\theta J_{\rm Geo\text{-}MIS} =\mathbb E_{\tau\sim\pi_{\rm old}} \Bigl[ \mathbb I(\rho(\tau)^{1/T}\le C)\,\rho(\tau) \sum_{t=0}^T\nabla_\theta\log\pi_\theta(a_t|s_t)A^\pi(s_t,a_t) \Bigr].$$

These are the paper’s core RL mechanisms for post-training (Luo et al., 10 Feb 2026).

The curriculum itself is stage-wise and difficulty-controlled. Difficulty is defined through an empirical zero-shot pass rate,

$D(x,y)=\tfrac1N\sum_{k=1}^N\mathbb I(\hat y_k=y).$

Training is partitioned into three stages that progressively tighten an upper-bound on difficulty, filtering out “trivial” samples with $D>0.7$ and “malformed” samples with $D=0$, while increasing rollout depth through larger group size and generation window. For P1-VL-30B-A3B, the schedule is: Stage 1 with group size 8, difficulty interval $\mathcal S$0, generation window 40 K, and learning rate $\mathcal S$1; Stage 2 with group size 8, difficulty interval $\mathcal S$2, and generation window 60 K; Stage 3 with group size 16, difficulty interval $\mathcal S$3, and generation window 80 K. For P1-VL-235B-A22B, Stage 1 uses group size 8, difficulty interval $\mathcal S$4, generation window 48 K, and learning rate $\mathcal S$5; Stage 2 uses group size 8, difficulty interval $\mathcal S$6, and generation window 54 K; Stage 3 uses group size 16, difficulty interval $\mathcal S$7, and generation window 54 K (Luo et al., 10 Feb 2026).

The ablation result reported for this component is specific: removing curriculum difficulty expansion stalls both response length and accuracy. Within the paper’s logic, the curriculum is therefore not a secondary optimization heuristic but part of the mechanism by which post-training remains stable while moving toward harder Olympiad reasoning.

4. Agentic Augmentation and inference-time self-verification

At inference time, P1-VL can be coupled with PhysicsMinions, described as a multi-studio agent loop. The pipeline contains a VisualStudio that parses the question into symbols, a LogicStudio that iteratively alternates solution drafting with introspective critique, and a ReviewStudio that verifies the candidate solution using sympy together with an LLM judge (Luo et al., 10 Feb 2026).

Operationally, the sequence is: the input consists of the question and images; VisualStudio parses the question into symbolic elements $\mathcal S$8; LogicStudio initializes a draft answer $\mathcal S$9; for each iteration up to a maximum, a Solver produces $\mathcal A$0, an Introspector critiques it, and the loop terminates if the critique is satisfactory, otherwise refinement continues; ReviewStudio then verifies the candidate through sympy and an LLM judge before returning the best answer. The paper characterizes this as mirroring the human pattern “reason, critique, refine.”

This inference strategy is separate from the base model weights. The benchmark results therefore report both standalone model performance and agent-augmented performance. Conceptually, the distinction matters: P1-VL is not only a trained VLM family but also an extensible reasoning system in which test-time control contributes materially to final capability. A plausible implication is that the architecture is intended as a platform for scientific reasoning workflows rather than a single monolithic forward pass.

5. Data curation and training strategy

The training corpus comprises 8,033 problems, split into 51% Olympiad material and 49% textbooks, including undergraduate sources and guides. Of these, 5,513 items are multimodal, with an average of 0.82 figures per question and up to 8 figures. When an item has no image, the procedure pads it with a blank $\mathcal A$1 image and an <image> token. The curation pipeline includes OCR correction, answer cross-validation using Gemini-2.5-Flash, Claude-3.7, and GPT-4o, followed by rule-based and LLM semantic filtering and then manual expert review (Luo et al., 10 Feb 2026).

The paper states that the vision encoder is frozen and that RL updates focus on language-model tokens. It also states that mixed text-only and image-text training yields no negative transfer. In the context of multimodal scientific reasoning, that claim is important because it addresses a recurrent concern that domain-specific multimodal specialization may erode language-only competence. Here, the reported result is the opposite: multimodal grounding does not degrade the text regime and can modestly improve transfer beyond physics.

The data composition also clarifies the intended capability profile. Because nearly half of the corpus is textbook material rather than Olympiad-only supervision, P1-VL is not trained exclusively for a narrow competition benchmark. The broader mixture suggests an attempt to couple formal problem-solving depth with wider scientific coverage, while the multimodal subset ensures that diagram-grounded reasoning is present at scale.

6. Benchmark results, generalization, and research significance

HiPhO aggregates 13 recent Olympiad theoretical exams, and its grading mirrors human scoring with partial credit for reasoning. On this benchmark, Gemini-3-Pro (high) records an average score of 42.7 with 13 gold medals; GPT-5.2 (high) records 40.6 with 12 gold and 1 silver; P1-VL-235B-A22B+PhysicsMinions records 40.9 with 12 gold and 1 silver; P1-VL-235B-A22B records 39.3 with 12 gold and 1 silver; P1-VL-30B-A3B records 35.0 with 9 gold and 4 silver; and the base Qwen3-VL-235B-A22B-Thinking records 33.9 with 10 gold and 3 silver (Luo et al., 10 Feb 2026).

Within that table, several claims are explicit. P1-VL-235B-A22B is reported as the first open-source vision-LLM to secure 12 gold medals and 1 silver. It ranks third overall among open and closed models, while its PhysicsMinions-augmented version rises to second globally, trailing only Gemini-3-Pro. Relative to its base Qwen3-VL-235B-A22B, RL post-training adds 5.4 average points. These results define the model family’s primary empirical identity.

Beyond HiPhO, the paper reports transfer to FrontierScience-Olympiad, a multidisciplinary text-only benchmark. GPT-5.2 scores 77.1 total; P1-VL-235B-A22B+PhysicsMinions scores 67.1; P1-VL-235B-A22B scores 64.3; and the text-only P1-235B-A22B scores 62.0. The multimodal model therefore yields a domain-transfer gain of 2.3 points over its text-only counterpart. The paper also reports that across ten math benchmarks, including AIME, HMMT, and IMO-Bench, and across five multimodal STEM tasks, including MMMU, EMMA-Mini, and MathVista-Mini, P1-VL consistently outperforms its base model; the examples given are a gain of 8.5 on AMOBench for the 235 B model and 3.4 on EMMA-Mini for the 30 B model (Luo et al., 10 Feb 2026).

The broader significance claimed for the release is threefold. First, it provides the community with the first open VLM specialized in physical reasoning at Olympiad rigor. Second, it lowers the barrier to multimodal scientific RL research and is presented as enabling extensions to chemistry, biology, embodied robot planning, and real-world experiment analysis. Third, it offers a blueprint—Curriculum RL plus Agentic Augmentation—toward “machine scientific discovery,” in which a model grounds symbolic reasoning in visual evidence and refines hypotheses iteratively. This suggests that P1-VL’s importance lies not only in its benchmark standing but also in the training-and-inference recipe it contributes to multimodal scientific reasoning research (Luo et al., 10 Feb 2026).