Normalizes each domain's teacher feedback spread during multi-teacher on-policy distillation so that no domain (e.g., instruction following) overwhelms others, improving student recovery of specialist skills and raising average scores across benchmarks.
Extrapolates RL-induced representation residuals into student hidden states during on-policy distillation: at each layer and token, it regresses the student beyond the teacher along the teacher’s RL-induced direction to improve stability and empirical performance.
Uses a multimodal model's own critiques as privileged context and applies on-policy self-distillation over diffusion sampling trajectories to internalize corrective guidance, improving text-to-image generation without an external teacher; shows measurable gains on GenEval and GenEval2.
Converts Turkish text to 48 kHz speech offline using a compact 5.6M DiT acoustic model plus a 3M vocoder (8.6M parameters, ~34 MB). Streams audio with very low latency (~3.86 ms first audio on an RTX 4090), supports fast batching and runs under an Apache‑2.0 license.
Combines joint distribution distillation from a video teacher with marginal (frame-level) distillation from an image teacher to improve few-step video generation. Introduces LatentBridge to align incompatible latents and Latent Variation Sampling to distribute frame supervision, boosting per-frame visual quality and semantic alignment while largely preserving motion.
Adapts LLM agents online by self-distilling verified execution trajectories into persistent LoRA weights during deployment to improve success and efficiency on long‑horizon tasks. Uses a frozen stable copy as a privileged teacher to predict hindsight next‑token distributions and filters invalid-action turns so experience consolidates without external solutions or memory retrieval.
Generates 5-second text- or image-conditioned videos with synchronized 44 kHz audio (including lip-sync) and built-in super-resolution to 1920×1080; available in Lite (3B) and Pro (29B) variants with code and checkpoints released under an MIT license.
Analyzes cross-tokenizer on-policy distillation (OPD) and shows that focusing supervision on strictly aligned token positions—or a student-selected top-16 subset of the shared vocabulary—retains most distillation gains, while adding span-level MSE supervision reduces downstream accuracy.
Creates programmable, real-time interactive code-based environments by separating deterministic simulator state from a shared neural video renderer so agents perceive, interact, and iteratively evolve via distilled playbooks; introduces Adversarial Forcing to distill a geometry-conditioned renderer for responsive visual feedback.
Uses large-scale, mixed-task reinforcement learning to drive self-improvement of multimodal foundation models. Key features include fully-asynchronous large-batch RL (1,568 prompts × 16 rollouts, up to 3.7B tokens/update, 1M context), groupwise agentic grading, and sparse MoE architectures.
Restores recoverability in on-policy distillation by having the teacher generate critiques of incorrect student rollouts, guide a critique-free student rewrite, and score the revised solution as privileged context; yields improved math-reasoning performance, largest gains on competition problems.
Runs the FLUX 3 Image Klein single-file checkpoint in ComfyUI for text-to-image generation and instruction-driven image editing with up to ten reference images, bounding-box layout control, and distilled/base variants; requires a prompt upsampler.