Research Briefing — June 15, 2026

This paper provides theoretical justification for the empirical success of non-Euclidean matrix optimizers like Muon, proving they achieve optimal sample complexity under heavy-tailed gradients while Euclidean methods incur dimension-dependent costs. It explains why Muon outperforms Adam-style methods for transformer training.

This paper introduces a diagnostic framework that models LLM behaviors (refusal, jailbreak, sycophancy) as low-rank subspaces in activation space, showing that interventions on one behavior propagate asymmetrically to others. It explains why targeted safety interventions cause unintended side-effects, which matters for designing reliable safety controls.

Introduces Compliance Asymmetry, a bidirectional diagnostic comparing whether LLMs follow helpful nudges versus misleading nudges. Across 9 models and 972,000 responses, finds models selectively resist harmful nudges on factual questions but comply with both directions nearly equally on moral questions, revealing directional blindness in moral judgment.

Every Eval Ever introduces the first shared schema and community-crowdsourced repository for AI evaluation results, standardizing how evaluations are represented in a unified JSON document. It tackles fragmentation and inconsistency across leaderboards, harnesses, and papers.

Proposes NULLs (Natively Unlearnable LLMs), a model class with shared backbone neurons and sparsely activated per-source sinks so that source-specific information concentrates in sinks and can be disabled at deployment to unlearn data. Resolves the tension between joint learning and clean removal.

Uses an LLM's representational geometry to predict which concept combinations it will fail to compose, attributing failures to feature interference when concept encodings are near each other rather than orthogonal. Demonstrated across programmatic, multihop, and multilingual tasks.

Investigates whether the same-origin policy remains effective in agentic browsers, showing the agent itself can act as a cross-origin data flow channel violating SOP. Introduces SOPBench and finds existing agentic browsers frequently violate SOP.

OdysSim builds foundation models for simulating human behavior, addressing the homogeneous overly-agreeable register that helpfulness training induces. It introduces the SOUL taxonomy, a 21.4M-interaction corpus, a benchmark, and a training recipe for behavioral simulation.

GauntletBench is a web-based benchmark evaluating AI agent generalisation in unfamiliar professional applications (3D modeling, circuit design, flight analysis) focusing on temporal perception, graphical understanding, and 3D reasoning. It probes underexplored agent capabilities where current benchmarks are saturated.

Introduces Muon^p, an optimizer generalizing Muon by using fractional spectral-power updates to interpolate between Muon and gradient descent. Proves fractional powers cannot be computed by fixed polynomial iterations and derives matrix-multiplication-only approximations.

Challenges the assumption that all residual stream activations suit sparse decomposition, hypothesizing a dense low-rank component unsuitable for sparse representation. Adds a small rank-r linear bottleneck parallel to sparse autoencoders, reducing dense latent count by up to 84 percent on Gemma-2-2B while improving reconstruction.

SuperThoughts compresses pairs of consecutive chain-of-thought tokens into single latent representations and decodes two tokens per step via multi-token prediction, doubling reasoning throughput while preserving discrete supervision. Evaluated on math reasoning benchmarks with Qwen models.