Research Briefing

Introduces Trojan-Speak, an adversarial fine-tuning method that bypasses Anthropic's Constitutional Classifiers with 99+% evasion while maintaining <5% capability degradation, compared to >25% in prior work. Uses curriculum learning and GRPO-based hybrid RL.

Challenges the conventional understanding that integer multiplication is hard for neural networks due to long-range carry chain dependencies. Shows this is a 'mirage' of representation choice—a 2D outer-product grid collapses operations to 3x3 local neighborhoods, enabling a 321-parameter neural cellular automaton to achieve perfect length generalization up to 683x training range.

Proposes a framework for predicting when CoT training will degrade monitorability. Models post-training as RL where reward decomposes into output-dependent and CoT-dependent terms, classifiable as aligned, orthogonal, or in-conflict.

Investigates whether suppressing LLM self-attribution of mental states (via safety fine-tuning) degrades Theory of Mind capabilities. Finds these are behaviorally and mechanistically dissociable, but notes safety-tuned models under-attribute mind to non-human animals and suppress spiritual beliefs.

Introduces the '2-datapoint reduced density matrix' from quantum chemistry to study phase transitions during deep learning training. Derives spectral heat capacity for early warning of phase transitions and participation ratio for measuring reorganization dimensionality.

Analyzes how AI model producers can game arena-style rankings (like Chatbot Arena) by submitting multiple model variants to artificially boost their top models' rankings. Establishes theoretical conditions for exploitation and proposes algorithmic solutions to make rankings manipulation-resistant.

Continuing our coverage from yesterday, Introduces MazeBench (110 procedurally generated mazes) and reveals that multimodal models solve mazes by converting to text grids and enumerating paths (1,710-22,818 tokens) rather than genuine visual planning. GPT-5.4 solves 91% but fails on hard mazes.

Shows that LLMs systematically fail when salient surface cues conflict with unstated feasibility constraints. Introduces the Heuristic Override Benchmark (HOB) with 500 instances demonstrating that no model exceeds 24% accuracy under strict evaluation, revealing fundamental reasoning limitations.

Large-scale field experiment at Alibaba assessing generative AI assistant impact on customer service workers. Finds AI improves service speed but agents retain discretion to adopt, modify, or ignore suggestions.

Proposes Tucker Attention as a generalized view of attention approximation methods (GQA, MLA), using Tucker decomposition to provide a unified framework for understanding and constructing parameter-efficient attention mechanisms.