Daily Brief

GEB-Bench introduces a benchmark for evaluating models on abstract structural motifs, revealing a significant gap in cross-voice mapping. Twelve models were tested, showing that while they excel in recognizing structures within a single voice, they struggle to transfer this understanding across different representations, with errors aligning more with formal geometries than perceptual ones.

The study presents a foraging agent utilizing active inference to dynamically allocate interoceptive precision, resulting in over double the survival rate during learning phases compared to uniform-precision agents (0.414 vs 0.199). This selective attention enhances both planning and perception, with significant implications for artificial intelligence in resource allocation.

Kimi K3, an open-weight model, is now available in GitHub Copilot, offering advanced coding capabilities at competitive pricing. It is accessible across various platforms including Visual Studio Code and Jetbrains, but requires activation for Copilot Business and Enterprise users.

The Right Reset (RR) method enhances prefix-removal probing in causal language models, achieving a 47.7% recovery rate of original records compared to 25.9% with BGE embeddings. This technique minimizes local output disruption across six models and shows that context dependence can signal boundaries effectively.

DataRx introduces a missingness-aware sampling method that enhances safety in large language model fine-tuning. By selecting safety-critical examples based on high-dimensional hidden representations, it reduces the attack success rate of Llama3-8B-Instruct from 59.23% to 13.70% with just 1% additional safety samples. This method aims to inspire further research into data-centric defenses.

This study reveals that the effect of repeated tokens in language models is position-dependent, with adjacent repetitions showing a plateau in prediction probability, while displaced repetitions exhibit an inverted-U effect. The findings are consistent across 13 models, including multilingual tests in Spanish, Chinese, German, and French, indicating that readout position significantly influences prediction outcomes.

Tesla and SpaceX will invest $16.8 billion to build the 'Terafab' chip factory in Texas, aiming to create over 100 million square feet of manufacturing space and employ at least 3,000 people. This facility will produce advanced chips to meet the growing demand for computing power in future technologies, including AI and robotics.

Cloudflare introduces Kitesurf, a new agent-first browser running on Workers, optimized for AI tasks with significantly lower CPU and memory usage compared to Chromium, enabling broader access for AI agents.