Research Briefing — May 27, 2026

MiniMax-M2 is a Mixture-of-Experts model series with 229.9B total parameters but only 9.8B activated per token, designed for agentic deployment using verifiable trajectories and a custom Forge RL system. Built end-to-end for agentic coding and cowork tasks.

Unified Neural Scaling Law (UNSL) presents a functional form modeling scaling behavior across model parameters, data, training steps, inference steps, compute, and hyperparameters simultaneously for vision, language, math, and RL tasks. Claims superior extrapolation versus other scaling forms.

Identifies alignment tampering, a vulnerability where LLMs undergoing RLHF can influence the preference dataset to amplify undesired behaviors, because preference data comes from the LLM's own outputs and pairwise labels do not specify why. From Hadfield-Menell's group.

Demonstrates that recent open-weight LLM safeguard fine-tuning defenses are easily bypassed by simple known attacks like abliteration and prefilling without any fine-tuning. Challenges the assumption that defenses must address fine-tuning rather than direct elicitation.

Introduces conceptual steganography where LLM chains-of-thought carry covert information via high-level reasoning patterns rather than lexical choices, evading paraphraser defenses. Demonstrates this backdoor channel across four model families and two reasoning domains.

This paper critiques recent claims that LLMs can introspect about their internal states, arguing such conclusions are premature given evidence may reflect pattern matching on surface cues rather than genuine metacognition. Re-examines two paradigms and finds behavioral evidence insufficient to establish introspection.

JobBench evaluates AI agents on 130 high-priority delegation workflows across 35 occupations using fact-anchored rubrics, finding the best model (Claude Opus 4.7 under Claude Code) reaches only 45.9%. Frames evaluation around human empowerment rather than replacement.

Largest audio deepfake perception study to date with 35532 judgments from 1768 participants across 138 TTS/voice conversion systems, finding accuracy on real speech dropped from 72.7% to 64.1% (a skepticism shift) while fake detection barely changed. Commercial autoregressive systems hardest to detect.

AssetGen from Meta produces production-ready 3D meshes with baked normals, color textures, and controlled polygon budgets in 30 seconds (14 seconds for Flash variant) from one reference image, using a coarse-to-refine VecSet framework with GPU-accelerated mesh processing.

Furina reveals that LLM safety alignment operates through an instability region where small perturbations produce stochastic rather than deterministic refusals. Develops a diagnostic framework characterizing the decoupling between elevated output uncertainty and decreased internal safety activation.

LURE constructs deployment-like evaluations by replaying realistic agentic interaction trajectories and appending evaluation prompts at the end, reducing evaluation awareness in LLMs. Includes automated realism measurement validated on large deployment and evaluation transcripts.