Beyond SFT-to-RL: Pre-alignment via Black-Box On-Policy Distillation for Multimodal RL

• URL: http://arxiv.org/abs/2604.28123v2
• Date: Fri, 01 May 2026 18:31:46 GMT
• Title: Beyond SFT-to-RL: Pre-alignment via Black-Box On-Policy Distillation for Multimodal RL
• Abstract summary: PRISM is a three-stage pipeline that mitigates drift between SFT and RLVR.<n>PRISM casts alignment as a black-box, response-level adversarial game between the policy and a Mixture-of-Experts discriminator.<n> Experiments on Qwen3-VL show that PRISM consistently improves downstream RLVR performance across multiple RL algorithms.
• Score: 32.91626063934283
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: The standard post-training recipe for large multimodal models (LMMs) applies supervised fine-tuning (SFT) on curated demonstrations followed by reinforcement learning with verifiable rewards (RLVR). However, SFT introduces distributional drift that neither preserves the model's original capabilities nor faithfully matches the supervision distribution. This problem is further amplified in multimodal reasoning, where perception errors and reasoning failures follow distinct drift patterns that compound during subsequent RL. We introduce PRISM, a three-stage pipeline that mitigates this drift by inserting an explicit distribution-alignment stage between SFT and RLVR. Building on the principle of on-policy distillation (OPD), PRISM casts alignment as a black-box, response-level adversarial game between the policy and a Mixture-of-Experts (MoE) discriminator with dedicated perception and reasoning experts, providing disentangled corrective signals that steer the policy toward the supervision distribution without requiring access to teacher logits. While 1.26M public demonstrations suffice for broad SFT initialization, distribution alignment demands higher-fidelity supervision; we therefore curate 113K additional demonstrations from Gemini 3 Flash, featuring dense visual grounding and step-by-step reasoning on the hardest unsolved problems. Experiments on Qwen3-VL show that PRISM consistently improves downstream RLVR performance across multiple RL algorithms (GRPO, DAPO, GSPO) and diverse multimodal benchmarks, improving average accuracy by +4.4 and +6.0 points over the SFT-to-RLVR baseline on 4B and 8B, respectively. Our code, data, and model checkpoints are publicly available at https://github.com/XIAO4579/PRISM.

Related papers

• Stop Thinking, Start Looking: Efficient Post-Training for Multimodal Document Question Answering via Reasoning-Free Alignment [2.8730465903425877]
  We introduce Perception-RFT, a training framework that applies Group Relative Policy Optimization to multimodal document QA.<n>We find that reasoning-enabled models suppress their reasoning traces during training, converging to direct perception-based policies at the 4B parameter scale.<n>We show that an early SFT$rightarrow$RL transition achieves comparable precision with 65% less training data.
  arXiv  Detail & Related papers  (2026-07-16T07:43:39Z)
• Bridging SFT and RL: Dynamic Policy Optimization for Robust Reasoning [10.4404175363648]
  Post-training paradigms for Large Language Models (LLMs) face a fundamental dilemma.<n>SFT provides stability (low variance) but suffers from high fitting bias, while RL enables exploration (low bias) but grapples with high gradient variance.<n>Existing unified optimization strategies often employ naive loss weighting, overlooking the statistical conflict between these distinct gradient signals.<n>We propose textbfDYPO (Dynamic Policy Optimization), a unified framework designed to structurally mitigate this conflict.
  arXiv  Detail & Related papers  (2026-04-10T03:42:16Z)
• Controllable Exploration in Hybrid-Policy RLVR for Multi-Modal Reasoning [88.42566960813438]
  CalibRL is a hybrid-policy RLVR framework that supports controllable exploration with expert guidance.<n>CalibRL increases policy entropy in a guided manner and clarifies the target distribution.<n>Experiments across eight benchmarks, including both in-domain and out-of-domain settings, demonstrate consistent improvements.
  arXiv  Detail & Related papers  (2026-02-22T07:23:36Z)
• Trust-Region Adaptive Policy Optimization [82.09255251747818]
  Post-training methods play an important role in improving large language models' (LLMs) complex reasoning abilities.<n>We introduce TRAPO, a framework that interleavesSupervised Fine-Tuning (SFT) and Reinforcement Learning (RL) within each training instance.<n>Experiments on five mathematical reasoning benchmarks show that TRAPO consistently surpasses standard SFT, RL, and SFT-then-RL pipelines.
  arXiv  Detail & Related papers  (2025-12-19T14:37:07Z)
• More Than the Final Answer: Improving Visual Extraction and Logical Consistency in Vision-Language Models [74.10138874771852]
  We propose PeRL-VL (Perception and Reasoning Learning for Vision-Language Models), a decoupled framework that separately improves visual perception and textual reasoning on top of RLVR.<n>For perception, PeRL-VL introduces a VLM-based description reward that scores the model's self-generated image descriptions for faithfulness and sufficiency.<n>For reasoning, PeRL-VL adds a text-only Reasoning SFT stage on logic-rich chain-of-thought data, enhancing coherence and logical consistency independently of vision.
  arXiv  Detail & Related papers  (2025-12-13T23:06:18Z)
• Enhancing the Outcome Reward-based RL Training of MLLMs with Self-Consistency Sampling [90.87033586963828]
  Outcome-reward reinforcement learning (RL) is a common and increasingly significant way to refine the step-by-step reasoning of multimodal large language models (MLLMs)<n>We propose Self-Consistency Sampling (SCS) to correct this issue.<n>Based on Qwen2.5-VL-7B-Instruct, SCS improves accuracy by up to 7.7 percentage points on six multimodal benchmarks with negligible extra computation.
  arXiv  Detail & Related papers  (2025-11-13T18:59:57Z)
• The Path Not Taken: RLVR Provably Learns Off the Principals [85.41043469428365]
  We show that sparsity is a surface artifact of a model-conditioned optimization bias.<n>We mechanistically explain these dynamics with a Three-Gate Theory.<n>We provide a parameter-level characterization of RLVR's learning dynamics.
  arXiv  Detail & Related papers  (2025-11-11T18:49:45Z)
• Directional Reasoning Injection for Fine-Tuning MLLMs [51.53222423215055]
  Multimodal large language models (MLLMs) are rapidly advancing, yet their reasoning ability often lags behind that of strong text-only counterparts.<n>Existing methods to bridge this gap rely on supervised fine-tuning over large-scale multimodal reasoning data or reinforcement learning.<n>We propose Directional Reasoning Injection for Fine-Tuning (DRIFT) to solve this problem.
  arXiv  Detail & Related papers  (2025-10-16T18:06:46Z)
• UniAPL: A Unified Adversarial Preference Learning Framework for Instruct-Following [12.924923059340395]
  We argue that post-training alignment is fundamentally a unified Preference Learning problem.<n>UniAPL implements a single-stage unified training objective, jointly learning from mixed batches of SFT and preference data.
  arXiv  Detail & Related papers  (2025-09-29T17:53:09Z)
• MMR1: Enhancing Multimodal Reasoning with Variance-Aware Sampling and Open Resources [113.33902847941941]
  Variance-Aware Sampling (VAS) is a data selection strategy guided by Variance Promotion Score (VPS)<n>We release large-scale, carefully curated resources containing 1.6M long CoT cold-start data and 15k RL QA pairs.<n> Experiments across mathematical reasoning benchmarks demonstrate the effectiveness of both the curated data and the proposed VAS.
  arXiv  Detail & Related papers  (2025-09-25T14:58:29Z)
• SFT or RL? An Early Investigation into Training R1-Like Reasoning Large Vision-Language Models [39.551767637896404]
  This work revisits the dominant supervised fine-tuning (SFT) then reinforcement learning (RL) paradigm for training Large Vision-Language Models (LVLMs)<n>We show that SFT can significantly undermine subsequent RL by inducing pseudo reasoning paths'' imitated from expert models.<n>We introduce VLAA-Thinking, a new multimodal dataset designed to support reasoning in LVLMs.
  arXiv  Detail & Related papers  (2025-04-10T16:54:05Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.