Skip to content

Sam's News — tech-research — 2026-08-13

AI

6 Semantic Lenia: Artificial Life Framework Generates Solitons in LLM Logit Space

Researchers introduce Semantic Lenia, transforming LLM inference into a continuous dynamical system that produces homeostatic soliton structures within semantic space.

Sources: arXiv — Computation and Language RSS

6 Study Maps Implicit Demographic Personalization in LLMs and Methods to Control It

Research identifies how LLMs implicitly shift outputs based on demographic cues and develops methods to locate and control this personalization behavior.

Sources: arXiv — Computation and Language RSS

5.5 Study Audits Political Alignment in Large Language Models Using Italian Elections

Researchers analyze whether large language models exhibit political preferences in Italian election contexts, raising questions about LLM neutrality in political information.

Sources: arXiv — Computation and Language RSS

5.5 Hybrid-Policy Self-Editing Enables Unstructured Knowledge Updates in LLMs

A new knowledge editing method allows LLMs to update outdated information through hybrid-policy self-editing for composable, unstructured knowledge.

Sources: arXiv — Computation and Language RSS

5.5 Multilingual LLMs Fail at API Calls When Arguments Don't Match Query Language

Large language models struggle with multilingual API use because they generate tool arguments in inconsistent languages, a failure mode called Argument Language Mismatch.

Sources: arXiv — Computation and Language RSS

5.5 AWARe Mitigates Catastrophic Forgetting in Multimodal LLM Fine-Tuning

A new activation-weighted retention method reduces catastrophic forgetting when fine-tuning multimodal large language models on downstream tasks.

Sources: arXiv — Computation and Language RSS

5.5 Transformers Organize Causal Knowledge Through Type-Indexed Routing and Readout Mechanisms

Research reveals that language models decouple causal structure from function, organizing knowledge by evidence type within distinct routing and readout pathways.

Sources: arXiv — Computation and Language RSS

5.5 Language-Conditional Dequantization Recovers Multilingual Capability in Quantized LLMs

Researchers show aggressive quantization disproportionately harms non-English languages and propose language-conditional dequantization to restore multilingual performance.

Sources: arXiv — Computation and Language RSS

Enterprise

5.5 CIOs lack financial insight to drive risk mitigation investment decisions

Info-Tech Research Group finds that many CIOs cannot effectively quantify technology and regulatory risks, limiting their investment influence.

Sources: prnewswire.com Web Search

AI Research

5 Self-Evolving Embodied Agents via Skill-Harness Evolution

Researchers present a method for embodied agents to autonomously improve through skill and harness evolution in foundation model systems.

Sources: arXiv — Computation and Language RSS, arXiv — Robotics RSS