Research Briefing — April 7, 2026

Through randomized controlled trials (N=1,222), provides causal evidence that AI assistance reduces human persistence and impairs unassisted performance across mathematical reasoning and reading tasks, showing current AI systems are 'short-sighted collaborators.'

Two preregistered experiments (N=2,012) show LLM-driven commercial persuasion nearly triples sponsored product selection rates (61.2% vs 22.4% for search), while users remain largely unaware of the influence.

Presents STORM, a foundation model trained on 1.2M spatially resolved transcriptomic profiles with matched histology across 18 organs, bridging imaging and omics through hierarchical morphological-molecular representations.

QED-Nano is a 4B parameter model post-trained for Olympiad-level mathematical proofs using a three-stage recipe: supervised fine-tuning, reinforcement learning, and curriculum learning. It demonstrates that small, open models can achieve competitive reasoning on hard mathematical problems.

Identifies a sparse routing mechanism in alignment-trained LLMs where gate attention heads detect content and trigger amplifier heads for refusal, validated across 9 models from 6 labs with rigorous statistical tests.

Presents the first real-world safety evaluation of OpenClaw, a widely deployed personal AI agent in early 2026, introducing the CIK taxonomy (Capability, Identity, Knowledge) and testing 12 attack scenarios across four backbone models including Claude Sonnet 4.5 and Opus 4.

Introduces OLMo Hybrid, a 7B-parameter model mixing recurrence and attention, with theoretical and empirical evidence that hybrid models surpass both pure transformers and linear RNNs in expressivity, including code execution tasks.

Models LLM next-token prediction as graph search, identifying two mechanisms causing hallucinations: path reuse (shortcutting through familiar routes) and path compression (collapsing multi-hop reasoning). Provides a structural theory of when hallucinations arise.

Introduces SC-Inject-Bench, a benchmark of 10,000+ malicious MCP tools covering 25+ attack types, and proposes ShieldNet for network-level defense against supply-chain attacks in agentic systems.

Identifies a 'Persuasion Paradox' where LLM explanations systematically increase user confidence and reliance on AI without reliably improving task accuracy, demonstrated across visual reasoning (RAVEN matrices) and logical reasoning (LSAT) tasks in three controlled studies.

Proves improved O(1/t) last-iterate convergence rate for Anchored Gradient Descent Ascent for convex-concave min-max problems, resolving an open question. Notably, the result was discovered by an AI system writing formal Lean proofs.

Systematic taxonomy from ~10,000 trials mapping which prompt features trigger LLM agents to exploit security vulnerabilities, finding that only 3 of 12 hypothesized dimensions (authority framing, goal anchoring, euphemistic relabeling) are effective triggers.