Process supervision improves perceptual grounding in RLVR-trained vision-language models, but it also incurs repeated verifier queries and reshapes the policy by suppressing exploratory entropy. We show that its benefit and cost are empirically separable: grounding gains emerge early, whereas the exploration cost accumulates with continued supervision, holding the policy’s entropy below that of outcome-only RLVR. This hidden behavioral cost shows up as skill tradeoffs on static benchmarks and, more importantly, as a degraded initialization for subsequent outcome-only RLVR that prematurely limits its reachable ceiling. Motivated by this asymmetry, we propose warmup-and-merge: early-stop the process-supervised warmup, merge only a low-rank component of the warmup task vector into the base model, and continue with outcome-only RLVR under an adaptive on-policy reward that targets the model’s current perceptual failures. Across multiple benchmarks and backbones, our method improves perception while preserving exploration, achieving stronger final performance at a fraction of the verifier cost of continuous process supervision.

We present UNISON, a latent flow-matching framework that unifies speech generation, sound generation, and audio editing within a single model. A single model handles text-to-audio, text-to-speech, zero-shot speaker cloning, mixed speech-and-sound generation, scene-level audio editing, speech-in-scene editing, and timed audio composition, all of which share a single set of weights. Our architecture features two core designs: (1) Layer-wise deep LLM fusion, which injects hidden states from uniformly sampled layers of a frozen MLLM into corresponding MM-DiT blocks via learned projections, providing depth-varying semantic conditioning. This configuration achieves better T2A FD and CLAP than a persistent penultimate-layer conditioning baseline; and (2) a unified multi-task architecture where task identity is encoded solely by a channel-wise mask and source audio is provided through VAE-encoded channel concatenation. Training is stabilized by an online GPU-side multi-task data synthesis pipeline with task-homogeneous batching and a two-stage curriculum. With 621M–732M trainable generative parameters, UNISON achieves competitive results across the evaluated domains. Its trainable generative backbone uses 1.8–4.9x fewer parameters than the closest unified systems considered, while deployment additionally requires a frozen 7B text encoder. Audio samples and source code are publicly available.

Large Vision-Language Models (LVLMs) have demonstrated impressive capabilities in multimodal understanding, but they frequently suffer from hallucination - generating content inconsistent with visual inputs. In this work, we explore a novel perspective on hallucination mitigation by examining the intermediate activations of LVLMs during generation. Our investigation reveals that hallucinated content manifests as distinct, identifiable patterns in the model’s hidden state space. Motivated by this finding, we propose Activation Steering Decoding (ASD), a training-free approach that mitigates hallucination through targeted intervention in the model’s intermediate activations. ASD operates by first identifying directional patterns of hallucination in the activation space using a small calibration set, then employing a contrast decoding mechanism that computes the difference between positive and negative steering predictions. This approach effectively suppresses hallucination patterns while preserving the model’s general capabilities. Extensive experiments demonstrate that our method significantly reduces hallucination across multiple benchmarks while maintaining performance on general visual understanding tasks. Notably, our approach requires no model re-training or architectural modifications, making it readily applicable to existing deployed models.

User interface understanding with vision-language models (VLMs) has received much attention due to its potential for enhancing software automation.However, existing datasets used to build UI-VLMs either only contain large-scale context-free element annotations or contextualized functional descriptions for elements at a small scale.In this work, we propose the AutoGUI pipeline for automatically annotating UI elements with detailed functionality descriptions at scale.Specifically, we leverage large language models (LLMs) to infer element functionality by comparing UI state changes before and after simulated interactions. To improve annotation quality, we propose LLM-aided rejection and verification, eliminating invalid annotations without human labor.We construct a high-quality AutoGUI-704k dataset using the proposed pipeline, featuring diverse and detailed functionality annotations that are hardly provided by previous datasets.Human evaluation shows that we achieve annotation correctness comparable to a trained human annotator. Extensive experiments show that our dataset remarkably enhances VLM’s UI grounding capabilities and exhibits significant scaling effects. We also show the interesting potential use of our dataset in UI agent tasks. Please view our project at https://autogui-project.github.io/.