A 17-year-old high school student has successfully turned common algae into a biological altimeter that reached the stratosphere. Andrew's StratoSpore project combines spectral sensing with machine learning to measure altitude through algae fluorescence???a world first that could transform how we mo...
Read Full Article →
Shepherd is the first agent runtime to combine Lean-verified core operations with a Git-like execution trace, achieving 5x faster forking than Docker and >95% prompt-cache reuse on replay. This paper marks the moment agent infrastructure moved from empirical reliability to formal guarantees.
Pion introduces a spectrum-preserving optimization method that keeps singular values constant during LLM training. This approach promises more stable and efficient training, but faces adoption barriers in compute and memory.
Researchers have introduced a federated optimization framework that simultaneously ensures differential privacy and Byzantine robustness under weaker, more realistic assumptions than prior work. This removes the primary theoretical barrier to deploying secure FL in heterogeneous, adversarial environments.
This research brief analyzes the new formal framework for comparing LLM safety when no ground-truth benchmark exists, explaining what the contract actually requires and why most vendors will fail to meet it.
A study of 89K human comparisons shows that global LLM leaderboards are misleading, as top 50 models are statistically indistinguishable. The authors propose portfolio-based evaluation as a more accurate alternative.
A new arXiv paper proposes a carbon-taxed compression pipeline that reduces LLM inference emissions by up to 60% with minimal accuracy loss. The research provides the first empirical evidence that green AI and software engineering performance are not inherently in conflict.
Researchers used linear probes on two social-media datasets to show that LLMs encode rhetorical questions as distinct, early-emerging features. This has implications for safety, manipulation detection, and our understanding of how models represent speaker intent.
Researchers have introduced a method to peel back the layers of low-bit transformer networks, exposing silent training failures that aggregate metrics miss. The findings suggest that many widely-used frozen models may contain underoptimized layers, raising urgent questions about model trustworthiness.
A new arXiv paper argues that the fixed rotation manifold in Rotary Positional Embeddings (RoPE) is a wasted opportunity for Transformer expressivity. By making both temporal and semantic rotations learnable, the authors claim a new dimension of attention capacity is unlocked.
New research reveals LLMs possess an internal logical subspace that can be steered, potentially eliminating the need for external symbolic solvers. This finding reshapes the competitive landscape for reasoning-focused AI products.
rDPO introduces instance-specific rubrics for visual preference optimization, outperforming standard DPO on fine-grained reasoning tasks. This paper argues that off-policy perturbations are fundamentally inadequate for multimodal alignment.
Researchers show that supervised step-by-step reasoning training degrades performance. Their CRAFT framework, using a Reasoning Knowledge Graph, offers a superior path, exposing a critical flaw in current AI training methodologies.
Append the next batch without leaving this page.
