Converts image-level rewards into explicit intermediate targets for diffusion-model denoising via an on-policy self-distillation loop. Constructs bounded positive/negative targets around anchors from reward gradients, fits those targets with finite updates, and refreshes a behavior policy by EMA—improving aligned performance across backbones while reducing GPU hours.
AI Model2026
GGUF-quantized, refusal-removed build of Qwen3.8-Flash-Next for llama.cpp that provides multimodal (image+text), reasoning and tool-calling capabilities; released for security research and red-teaming under the Apache-2.0 license.
AI Model2026
GGUF-quantized build of Qwen3.8-Flash-Next for image-text-to-text inference and local deployment. Ships with Unsloth Dynamic 3.0 quantization, thinking-mode controls (preserve_thinking, reasoning_effort), and native long-context support (262k, extensible to 1M with YaRN).
AI Model2026
Integrates a pretrained vision–language model with a BEV perception head and a Planning Expert to provide 3D perception, driving VQA and motion planning for autonomous driving while keeping the base VLM architecture unchanged.
Performs causal, bounded‑memory streaming 3D reconstruction by caching KV features from only the preceding 11 frames, predicting a per‑frame point map and adjacent relative pose, and composing these local predictions into a global trajectory; includes a lightweight rotation refiner and composition‑aware loss to limit drift.
Turns a flow-matching image generator's self-exploration into dense, per-step supervision without a pretrained teacher; it branches the student's next-state into stochastic SDE candidates, scores them against a deterministic self-reference, and applies an advantage-weighted pull–push velocity regression with reward-level fusion for multi-objective alignment.
Provides a real-scale 3D Hong Kong sandbox to evaluate whether multimodal LLM agents can turn local street-view perception into sustained spatial action, supporting closed-loop first-person interaction, an interactive map, and controlled tests of grounding, long-range navigation, and robustness.
Rewrites physical scenes as executable world programs (e.g., MuJoCo scene descriptions) and uses an agentic abductive loop to propose, execute, render, verify, and iteratively refine those programs from videos or text. Verified executable worlds supply scalable physical supervision for training vision–language models.
Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models
Proposes VLAct, a representation-centric continued pre-training method for Vision-Language-Action models that preserves VLM priors and enforces cross-embodiment action semantics to turn limited robot trajectories into transferable visual-action representations; shows strong gains and sample efficiency on multiple VLA benchmarks using modest compute.
AI Agent Papers2026
Autonomous multimodal GUI agent that executes natural-language interface tasks across mobile apps, web domains, and desktop OS. Expands environment coverage (170+ multilingual apps, 4,000+ web domains), uses function-grounded task generation and keypoint-based multi-model verification to produce reliable RL rewards for real-world deployment.
Proposes GenFirst, a generation-before-reconstruction end-to-end training strategy for latent generative models that avoids latent collapse by prioritizing generative objectives and then progressively strengthening reconstruction, validated with strong gFID/GenEval results on ImageNet-256 and text-to-image tasks.
AI Model2026
An experimental multimodal model that adds visual understanding to DeepSeek-V4-Flash: accepts text+image inputs and returns text analyses. Improves vision-dependent agent workflows while maintaining comparable text-only performance; released under an MIT license on Hugging Face.