Also Worth Noting - 2026-06-22

Also Worth Noting

02 [Eval] StylisticBias: A Few Human Visual Cues Drive Most Social Biases in MLLMs A small set of styling signals, not identity categories, accounts for most measurable social bias in multimodal models. StylisticBias holds identity constant across 500 photorealistic base faces and generates roughly 25,000 images by varying one attribute at a time, things like clothing and grooming, so appearance effects can be measured cleanly. That design separates what auditors actually need to fix from what they have been inadvertently confounding. Teams running MLLM bias audits should start with styling cues before broader demographic comparisons. link

03 [Theory] Multi-Vector Embeddings are Provably More Expressive than Single Vector Embeddings The accuracy gap between ColBERT-style multi-vector retrieval and dense single-vector retrieval is not a tuning artifact. This is the first formal proof that no single-vector inner-product model can replicate Chamfer-similarity multi-vector scoring on certain query distributions, regardless of dimension. The expressiveness ceiling is structural, not a matter of scaling up embedding size. Retrieval engineers choosing between dense and multi-vector architectures now have theoretical grounding to match their system constraints against, not just empirical benchmarks. link

04 [Training] Multi-Turn Reflective Masking Elicits Reasoning in Mask Diffusion Models Mask diffusion models can do chain-of-thought reasoning without regenerating the full sequence at each step. Instead of discarding a previous answer and starting from scratch, the masking mechanism selectively re-masks only the tokens that need revision, then resamples them, cutting compute per reasoning iteration while preserving answer quality. Autoregressive models revise by appending; this approach revises by local overwrite, which more closely mirrors how iterative correction actually works. Teams exploring non-autoregressive reasoning pipelines should treat this as a concrete efficiency baseline. link

05 [Agent] Tmax: A simple recipe for terminal agents A 9B-parameter open model hits 27% on Terminal-Bench 2.0, outperforming larger closed models from prior work. Tmax gets there through RL on a curated shell-task dataset, and the full recipe including data generation and training code is public, making it the first reproducible open baseline for terminal agents. The gap between open and proprietary in this category has been wide; Tmax closes it at a parameter count teams can actually deploy. If you are building or evaluating terminal-using agents, this is now the reference starting point. link

06 [Open-source] Scaling Linear Mode Connectivity and Merging to Billion Parameter Pretrained Transformers Existing linear mode connectivity methods optimize the interpolation path from only one model endpoint, and that asymmetry is what breaks them at scale. This framework applies functionality-preserving weight transformations from both endpoints simultaneously, aligning functionally equivalent solutions before interpolation begins. The result is successful LMC-based merges on billion-parameter pretrained transformers, a scale where previous methods fail. Teams merging fine-tuned checkpoints for multi-task or continual learning deployments now have a scalable method that matches the models they actually ship. link