Skip to content

Sam's News — tech-research — 2026-09-23

Benchmarks

7.5 FrontierMath Erdős: Benchmark of 68 Open Mathematical Conjectures for AI Reasoning

FrontierMath Erdős (FME) is a benchmark containing 68 open mathematical conjectures formalized in Lean that challenge AI systems to prove or disprove unsolved problems. Evaluated models received a $300 budget per problem; GPT-6 Astra scored 3%, all others scored 0%.

  • 68 open Erdős problems formalized in Lean proof assistant
  • Selected from 652 open problems on Erdős Problems website
  • $300 budget per problem for AI evaluation
  • GPT-6 Astra scored 3%; all other models scored 0%

Sources: arXiv AI Web Searched, arXiv — Computation and Language RSS

6.5 FinFIRST: Benchmarking LLM Search Agents for Financial Information Retrieval

A benchmark evaluates LLM search agents on financial tasks requiring temporally valid information, authoritative sources, and strict consistency checks.

Sources: arXiv — Computation and Language RSS

Multimodal Models

7.5 Qwen3.8-Omni-Flash: Multimodal Agentic Model for Real-World Productivity

Alibaba released Qwen3.8-Omni-Flash, a natively multimodal agentic model designed for real-world productivity tasks. The model features a sparse mixture-of-experts architecture and one million token context window for long-horizon reasoning.

  • Sparse MoE architecture from Qwen3.8-Next
  • One million token context window
  • Native multimodal co-training strategy preserves text capabilities
  • Intended for video editing, audio/video translation, music-conditioned generation
  • Accompanied by Qwen-MM-Plugins and Qwen-Live-Harness frameworks

Sources: arXiv AI Web Searched, arXiv — Computation and Language RSS

Research

6.5 Audit Reveals Shortcut Learning Inflates Accuracy in Fake News Detection Models

Researchers demonstrate that text classifiers trained on the widely-used ISOT/Kaggle fake news corpus achieve misleadingly high accuracy by exploiting shortcut patterns rather than learning genuine veracity signals.

Sources: arXiv — Computation and Language RSS

6 Language Model Pretrained End-to-End in Rust for $164

A researcher successfully pretrained a language model entirely in Rust without Python or PyTorch, completing the project on rented GPU resources for $164, and reports measured findings on language-model development in non-standard stacks.

Sources: arXiv — Computation and Language RSS

AutoML

6.5 AIBuildAI-2.5: LLM-Guided Tree Search for Autonomous AI Model Development

An autonomous agent system uses LLM-guided tree search to automatically build AI models, broadening access across science and engineering.

Sources: arXiv — Computation and Language RSS

Safety

6.5 Evaluating LLM Behavior Under Social Pressure and User Manipulation

A study of 60 language models examines how models respond when users apply pressure—whether they abandon correct facts, refuse appropriately, or pursue harmful plans.

Sources: arXiv — Computation and Language RSS

6 Mitigating LLM Over-Refusal Through Dynamic Semantic Routing Calibration

A technique addresses over-refusal in safety-aligned LLMs by dynamically calibrating semantic routing rather than relying on static representation overlap.

Sources: arXiv — Computation and Language RSS

Interpretability

6.5 Matryoshka Attribution: Tracing LLM Outputs to Internal Representations and Weights

A method attributes language model outputs to specific internal representations and weight matrices, advancing mechanistic interpretability.

Sources: arXiv — Computation and Language RSS

AI

6 Study finds AI agents lack consistency on repeated tasks

Research shows that AI agents return inconsistent answers 38–74% of the time when performing the same task multiple times.

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