Research Briefing — May 7, 2026

Proves a fundamental impossibility triangle for long-context models: no model can simultaneously achieve per-step efficiency independent of sequence length, compact state size, and recall proportional to sequence length. Uses information-theoretic tools (Data Processing Inequality, Fano's Inequality) to establish the bound.

Introduces Design Conductor 2.0, a multi-agent system powered by April 2026 frontier models that autonomously builds an LLM inference accelerator (VerTQ) with TurboQuant support in 80 hours — handling 80x larger tasks than its predecessor which built a RISC-V CPU.

Demonstrates 'AI authority laundering' where adversarial perturbations to images cause VLMs to give confident, authoritative responses about wrong inputs. Shows standard CLIP attacks transfer to production VLMs including GPT and Claude.

Derives analytical decomposition of RLHF optimization gradient into standard policy gradient and parameter-steering term, showing standard iterative RLHF suffers 'alignment collapse' as policy exploits RM blind spots. Proposes corrections based on the Stackelberg game formulation.

Develops methods for aligning LLMs in 'fuzzy' domains where verification is difficult, using online natural language feedback from experts. The approach iteratively optimizes against proxy reward signals while collecting fresh expert supervision and updating the proxy reward model.

Provides theoretical analysis of in-context learning (ICL) for nonlinear regression, showing how attention mechanisms can construct nonlinear features like polynomial or spline bases. Establishes a framework analyzing end-to-end in-context nonlinear regression beyond the linear setting.

Introduces TRIBE v2, a tri-modal (video, audio, language) foundation model predicting human brain activity across diverse conditions. Uses 1,000+ hours of fMRI across 720 subjects and enables in-silico experimentation for neuroscience.

Proves a fundamental predictive-causal gap: across 2695 neural network configurations, optimal predictive encoders track the environment rather than the system of interest. Shows this is structural (not an optimization artifact) and worsens with dimensionality.

Uncovers a mechanism underlying safety degradation during LLM fine-tuning: benign fine-tuning causes parameters to cumulatively drift toward danger-aligned directions. Proposes a sample-level risk scoring method to identify high-risk training samples before fine-tuning.

Reveals that single-position activation intervention achieves 0% task transfer across all 28 layers of Llama-3.2-3B despite 100% probing accuracy, demonstrating task encoding is fundamentally distributed. Multi-position intervention at demonstration output tokens achieves up to 90% transfer.

Shows that the power distribution (target of power sampling) unifies sampling, self-reward KL-regularized RL, and self-distillation for LLMs. Demonstrates that local approximations cannot reproduce sequence-level power without suffix information, providing theoretical clarity on RL vs. sampling debates.