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.
Improves test-time scaling of looped transformers by adaptively assigning extra recurrent iterations to tokens that benefit most. TaH2 is a post-training method that jointly trains an iteration decider with the backbone using lookahead depth supervision, boosting the accuracy–compute slope and peak accuracy on challenging benchmarks.
Models sequence generation by unmasking multiple tokens per denoising step and replaces a factorized reverse process with a mixture over discrete routing-based latents from an MoE backbone; improves few-step sampling quality without increasing active parameters.
Demonstrates that pretrained transformers typically use only ~1–3 lines of depth to follow reference chains, and that a task‑trained rank‑8 LoRA applied at one early layer (with all other weights frozen) can extend reference‑following to dozens or hundreds of lines while adding only a few ten‑thousand parameters.
Compresses recurrent states in linear-attention LLMs via spatial–temporal post-training quantization, allocating bits by error lifetime and per-row impact to preserve accuracy (6-bit ≈ FP32) while cutting serving memory up to 68.7%.
Turns solved protein structures into FoldingCorpus and Fold2Reason — a post-training recipe that supervises an LLM with discrete structural Q&A plus continuous 3D geometry to improve spatial, graph and scientific reasoning; reports consistent gains across 10 benchmarks.
Calibrates multi-reward reinforcement learning by adaptively upweighting infrequently active rewards per rollout batch, so sparse objectives provide stronger signals when they matter. Proposes an inverse-square-root density correction and shows faster learning on tool-calling and math-reasoning tasks.
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.
Analyzes how proposer–solver loops in self-evolving search agents can develop shared errors (co-cheating) that inflate internal rewards; introduces Multi-Sample Verification and CrossFit (cross-fitted scoring with partitioned sources) to reduce false agreement and improve downstream search performance.
Demonstrates that additive weighted aggregation in rubric-based reinforcement learning enables reward hacking—models improve rubric coverage but produce worse answers—and proposes ProRubric, an offline regrouping that aggregates checklist criteria into conjunctive protocol-level dimensions to raise appropriateness without losing coverage.
Couples a continuous latent diffusion trajectory with discrete token readouts so tokens are read from the latent at every step and fed back as a scaffold, reducing token-independence in parallel decoding and improving structured-task accuracy and LM perplexity.
Introduces a token-adaptive latent recurrence for diffusion language models that allocates computation per token during denoising. Key features: iterative latent refinement, discrete-feedback commits, and learned token-wise schedules to improve parallel decoding quality-efficiency trade-offs.