Sam's News — tech-research — 2026-09-03¶

Neuroscience¶

8 Fruit fly connectome completion reshapes neuroscience approaches¶

Researchers at HHMI's Janelia Campus completed a comprehensive map of all neurons in the fruit fly brain, fundamentally transforming neuroscience methodology.

Largest brain connectome map completed to date
Male fruit fly brain fully mapped including central nervous system
Collaboration between Google Research and HHMI Janelia
Results published with accessible connectomics website and dataset

Sources: Phys.org RSS, Live Science RSS, Google Research AI Web Searched

Research¶

6.5 DiDrive: Risk-aware diffusion framework for safe offline reinforcement learning in autonomous driving¶

A new arXiv paper presents DiDrive, a hierarchical diffusion-based approach for safe offline reinforcement learning that mitigates distribution shift and OOD actions in autonomous driving.

Sources: arXiv — Machine Learning RSS, arXiv — Robotics RSS

6.5 AI Agents Reshape Consensus Formation in Human Groups¶

Study demonstrates that LLM agent participation in human groups measurably changes how consensus is formed.

Sources: arXiv — Computation and Language RSS

6.5 Lawyer Agents Mitigate Guilty Bias in Legal Judgment Prediction¶

LLM agents trained as legal defenders reduce systemic guilty bias in automated legal judgment prediction systems.

Sources: arXiv — Computation and Language RSS

6.5 Fusing LLMs with Ontology Rankers for Rare Disease Diagnosis¶

Hybrid approach combines LLM differential diagnosis generation with ontology rankers for improved rare disease detection.

Sources: arXiv — Computation and Language RSS

6 SafeEvolve: Co-evolutionary safety alignment for LLM-based agents¶

A new arXiv paper proposes SafeEvolve, a method for jointly optimizing language model policies and system harnesses to improve safety alignment in agent behavior.

Sources: arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS

6 SEAL: Shared expert alignment for safety in mixture-of-experts language models¶

A new arXiv paper introduces SEAL, a method for reinforcing global safety constraints across mixture-of-experts architectures through alignment of shared expert modules.

Sources: arXiv — Machine Learning RSS, arXiv — Cryptography and Security RSS

6 Interpretable Symptom Vectors for Depression Detection via LLMs¶

Research applies large language models to capture diverse depression symptom profiles from patient narratives.

Sources: arXiv — Computation and Language RSS

6 Cognitive Alignment Between Humans and Reasoning Models¶

Study shows large reasoning models exhibit behavioral alignment with humans on abductive reasoning tasks.

Sources: arXiv — Computation and Language RSS

6 EqGrid: Energy-Efficient LLM Agents for Peer-to-Peer Energy Markets¶

Framework deploys compute-efficient LLM agents on edge devices to address energy poverty in distributed energy systems.

Sources: arXiv — Computation and Language RSS