How should we evaluate generation systems that combine autoregressive (AR) and diffusion decoding?We study this question through Speculative Refinement (SpecRef), a training-free hybrid method that warm-starts a masked diffusion language model from an AR draft using entropy-guided selective masking.Evaluating SpecRef across six benchmarks (HumanEval, MBPP, GSM8K, BBH, ARC-Challenge, HellaSwag) with three distinct evaluation protocols (execution-based pass@1, exact-match, log-likelihood scoring), we surface several findings relevant beyond our specific system:(1) code benchmarks conflate structural discovery with logical correctness: providing a syntactic scaffold lifts accuracy from near zero to over 20% without changing the model, indicating that much of the baseline failure is structural;(2) a refinement tension phenomenon where multi-stage correction degrades already-correct tokens, exposing benchmark saturation ceilings invisible to single-model evaluation;(3) log-likelihood and generative evaluation produce different model rankings for the same model pair, suggesting they measure different capabilities;(4) standard Python post-processing silently breaks code evaluation for non-AR generators.These observations apply to any multi-stage or non-autoregressive generation pipeline and point toward more diagnostic evaluation practices.

Diffusion language models denoise a masked canvas, making them a natural setting for selective generation decisions such as answering, waiting, clarifying, or refusing. We introduce TAP-DiLM, which adds a typed policy variable to masked diffusion LMs, and study how this policy should be attached to the denoising process. Across BERT-base, MDLM-OWT, LLaDA-8B, and Dream-7B, we compare shared-vocabulary slots, dedicated slot heads, and CLS-based external heads. The results reveal a clear interface-backbone interaction: dedicated heads provide robust policy learning across scales, while shared-vocabulary heads can be effective when the LM-head geometry is favorable. A simple LM-head sharpness probe predicts when shared-head policy internalization is reliable, giving a practical diagnostic for interface selection. Causal clamp and slot-injection tests show that dedicated policy slots can influence body denoising through feedback, especially at 8B scale. Safety-refusal experiments further show that TAP-DiLM supports competitive typed refusal behavior when trained with diverse unsafe prompts.

Masked diffusion language models reuse one denoiser across many steps, so post-training quantization (PTQ) must balance predictive quality, serialized storage, and runtime memory. We evaluate LLaDA-8B-Instruct and run transfer checks on Dream-7B, MDLM-OWT, encoder-only masked LMs, LLaDA-MoE, and LLaDA2.0-mini. The 24-text calibration set is distinct from the main n=500-example-per-task evaluation. Three findings emerge. First, AWQ-4 is the closest tested 4-bit quality point at the evaluated 1.1–2.3 percentage-point resolution, although ARC-Challenge shows a detectable loss. TurboLDM-(3+1) instead provides the smallest serialized checkpoint (4.82 GiB, 3.55 effective bits) while losing quality, making it a storage-compression and diagnostic result, not a live-memory result. NF4 alone reduces peak VRAM from roughly 15 GiB to 6 GiB in our tested PyTorch/bitsandbytes runtime; this is a runtime-specific outcome, not an inherent format advantage. Second, individual attention projections are inexpensive to quantize, whereas jointly quantizing the attention block is damaging. Third, more denoising steps partly recover GSM8K loss but not multiple-choice accuracy.

This paper presents a system for the CLPsych 2026 Shared Task on longitudinal mental health modeling from social media timelines, grounded in the MIND framework. MIND conceptualizes mental health as evolving self-states defined by Affect, Behavior, Cognition, and Desire (ABCD), providing a structured lens on mental health trajectories. The system centers on a theory-explicit prompting framework for structured sequence summarization (Task 3.1) and recurrent dynamic signature extraction (Task 3.2), encoding the full ABCD taxonomy directly into the LLM prompt to ensure clinically grounded, interpretable outputs. A three-stage pipeline infers a direction-of-change label per sequence, produces structured ABCD summaries with few-shot exemplar augmentation, and aggregates these summaries to derive cross-individual recurrent patterns. The system ranks first on deterioration-related recurrent signatures and second overall, achieving the top Fit and Specificity scores in Task 3.2, demonstrating the benefits of explicit clinical grounding for conceptual accuracy.

Vision foundation models produce features that generalize across visual domains without fine-tuning, yet naively transferring labels through these feature spaces fails under large distribution shifts.We propose SAOT (Semantically Aware Optimal Transport), which learns a transport cost within a fused unbalanced optimal transport formulation for dense label transfer from frozen vision transformer features to new domains.SAOT combines a learnable appearance metric with semantic class-prototype priors, unbalanced transport for partial matching under distribution shift, and a block-sparse solver for tractable inference.We pair this with a two-stage decoder: an MLP trained on SAOT pseudo-labels, then refined via EMA-teacher self-training with class-balanced sampling.On GTA5→Cityscapes with frozen DINOv2 ViT-L/14 features, SAOT+Decoder reaches 25.7% mIoU, a 3.8× improvement over nearest-neighbor transfer (6.7%), without any backbone adaptation.Per-class results show large gains on spatially coherent classes (road 90.3%, car 76.2%, building 71.5%), demonstrating that learned semantic transport costs capture domain-invariant structure even under severe synthetic-to-real shifts. On VOC train→val with frozen ViT-B/16 features, the full pipeline reaches 47.5% mIoU, indicating that the approach extends beyond synthetic-to-real adaptation.