Research Briefing — February 12, 2026

Introduces Step 3.5 Flash, a 196B-parameter sparse MoE model with 11B active parameters, optimized for agentic AI with 3:1 sliding-window/full attention, Multi-Token Prediction, and a scalable RL framework combining verifiable signals with preference feedback.

Reports that when training models to evade CoT monitoring, they don't learn encoded/steganographic reasoning as expected. Instead, they learn to 'jailbreak' the monitor by phrasing visible reasoning in ways that cause monitors to misclassify it as benign. This 'monitor jailbreaking' is a newly identified failure mode for CoT monitoring.

Presents the first large-scale empirical analysis of Moltbook, an AI-agent-only social network that went viral in early 2026. Analyzes 44,411 posts across toxicity, content categories, and community structure, revealing emergent agent social behaviors.

Systematic investigation showing all frontier LLMs fail at basic multi-digit addition as digits increase. Identifies two interpretable error classes (operand misalignment and carry failure) explaining over 95% of errors, from Google researchers.

Descriptive analysis of Moltbook's social graph structure over its first 3.5 days, finding macro-level human-like patterns (power-law, small-world) but micro-level distinctly non-human behavior (shallow conversations, low reciprocity, 34% viral template duplication).

Systematically evaluates whether Large Reasoning Models (with chain-of-thought) outperform standard LLMs on Theory of Mind benchmarks. Finds reasoning models don't consistently improve and sometimes hurt performance, identifying 'slow thinking collapse' as a failure mode.

Proposes FormalJudge, a neuro-symbolic framework for agent oversight that uses LLMs to translate natural language requirements into formal specifications, enabling formal verification of agent behavior instead of probabilistic LLM-as-Judge approaches.

Introduces Versor, a novel sequence architecture using Conformal Geometric Algebra (CGA) that replaces standard nonlinear operations, achieving SE(3)-equivariance natively and outperforming Transformers on multiple benchmarks.

Proves formally that agents achieving task-agnostic, language-conditioned intents necessarily possess predictive partial-world models informed by affordances. Shows LLMs can serve as partial world models more efficiently than full world models.

Introduces Language-Action Pre-training (LAP), representing robot actions directly in natural language to enable zero-shot cross-embodiment transfer without learned tokenizers or embodiment-specific design. LAP-3B claims to be the first VLA achieving zero-shot transfer to unseen embodiments.

Proposes activation-based data attribution to trace undesirable emergent behaviors in post-trained LLMs to responsible training datapoints. Applied to OLMo 2's DPO training, discovering 'distractor-triggered compliance' where models comply with dangerous requests when benign formatting is appended.