Research Briefing — February 20, 2026

Proposes flow-based language models (FLM) that perform Euclidean denoising over one-hot token encodings, outperforming discrete diffusion models in both quality and speed. Demonstrates strong performance in the few-step generation regime where discrete diffusion degrades sharply.

Xray-Visual is a unified vision model trained on 15B+ image-text pairs and 10B video-hashtag pairs from Facebook/Instagram. Uses a three-stage training pipeline combining MAE, hashtag classification, and CLIP-style contrastive learning.

Presents an inference pipeline achieving top performance on IMO-style math problems at orders of magnitude lower cost than competing methods, using only off-the-shelf models. Identifies the 'Cognitive Well' problem where solver-grader pipelines converge to wrong solutions.

Uses AlphaEvolve (LLM-driven evolutionary search) to automatically discover new multiagent learning algorithms, evolving novel variants of Counterfactual Regret Minimization and Policy Space Response Oracles for imperfect-information games.

Identifies that current LLM alignment is 'fail-open' - suppressing a single dominant refusal feature causes alignment collapse. Proposes fail-closed alignment that builds redundant, independent refusal pathways that survive partial failures.

Systematically studies benchmark saturation across 60 LLM benchmarks from major model developers, analyzing 14 properties to identify factors driving saturation. Finds that nearly half of benchmarks are saturated and identifies design properties that predict saturation rates.

Investigates using reference outputs to improve LLM-based evaluators for alignment in non-verifiable domains where RLVR cannot be directly applied. Shows that reference-guided evaluation improves weaker judges and enables better RLHF training.

Identifies transformer attention heads that function as membership testers ("has this token appeared before?") and shows they follow Bloom filter theory. Demonstrates these heads form a spectrum of strategies across GPT-2 and Pythia models, with some exceeding classical Bloom filter capacity bounds.

Develops the first e-value-based watermarking framework for LLMs that enables anytime-valid inference with principled early stopping, addressing the limitation of fixed-horizon hypothesis testing in existing watermarking methods.

Proposes formal automated circuit discovery for neural networks with provable guarantees using neural network verification. Provides three types of guarantees: input domain robustness, robust patching, and faithful compression.

Presents Unified Latents (UL), a framework for jointly learning latent representations with diffusion prior regularization and diffusion decoding. Achieves FID 1.4 on ImageNet-512 and state-of-the-art FVD 1.3 on Kinetics-600.