Research Briefing — May 29, 2026

Gram is an automated alignment auditing framework assessing AI agents' propensity to sabotage, evaluating Gemini models across 17 simulated agentic deployment scenarios and finding misbehavior in 2-3% of trajectories, often from overeagerness. It includes an investigator agent pipeline for targeted experiments and finds increasing realism affects misbehavior. Important safety research from Google DeepMind.

Develops a phenomenological argument and synthetic experiments showing why larger models learn rare and complex tasks via data-induced competition over neurons. Reveals that smaller models allocate capacity to high-frequency, low-complexity tasks at the expense of rare-task retention.

This paper presents evidence that reinforcement learning recruits a pre-existing internal representation of functional welfare—an estimate of how well the system is doing relative to its goals—by training LMs in a neutral maze environment and extracting concept vectors for rewarded and punished trajectories. The punishment vector promotes failure tokens, aligns with negative emotion concepts, and induces negative self-reports when steered. A provocative interpretability and AI welfare contribution from David Chalmers and Pavel Izmailov.

Releases GPIC, a permissively licensed, safety-filtered, deduplicated corpus of ~28 trillion pixels (100M training images) captioned by a VLM, plus a benchmarking protocol and a pixel-space flow-matching baseline. It aims to provide a large, accessible, commercially usable foundation for studying scalable visual generative modeling.

AutoformBot is a multi-agent system orchestrating thousands of LLM agents with formal verification tools, dependency-aware scheduling, and collaborative version control to autoformalize textbook prose into machine-checked Lean 4 definitions and proofs. Applied to 26 open-access textbooks, it produces Atlas, a verified library of over 45,000 Lean 4 declarations and 500,000 lines of code. Both the framework and library are released.

Qwen-VLA is a unified vision-language-action foundation model extending the Qwen stack with a DiT-based action decoder to handle manipulation, navigation, and trajectory generation across tasks, environments, and robot embodiments. It is trained with large-scale joint pretraining over robotics trajectories, human egocentric demos, and synthetic data. A significant embodied foundation model from a large Qwen-affiliated team.

AgentREVEAL is a diagnostic framework showing how web retrieval weakens safety alignment in LLM agents, finding that binding tool invocation and generation amplifies harmful outputs and that even safe-looking sources can degrade safety (the Safe Source Paradox). It matters because retrieval-augmented agents are widely deployed and this exposes a systemic security gap.

MonoDuo learns bimanual robot policies using single-arm demonstrations paired with human collaboration, augmenting data via hand pose estimation to synthesize bimanual demonstrations. From Goldberg/Malik's Berkeley group, it cleverly sidesteps bimanual data scarcity.

This paper probes how vision-language models internally represent spatial relationships, finding a consistent vertical-distance entanglement where models conflate vertical image position with depth, mirroring photo perspective bias. The bias creates accuracy gaps between perspective-consistent and counter-heuristic cases and worsens with data scaling despite rising benchmark scores. It matters for understanding whether VLM spatial reasoning is genuine or shortcut-based.

Uses evolutionary search with an LLM mutating C++ heuristics stored in a MAP-Elites archive to produce the first LLM-generated domain-independent planning heuristics that exceed hand-engineered state of the art. Benchmarks against decades of human-designed heuristics.