Sam's News — tech-research — 2026-08-21¶

AI¶

7.5 Google DeepMind partners with game studios to advance AI gameplay research¶

Google DeepMind is partnering with game studios to advance AI gameplay research, building on 15 years of work from Atari to EVE Online. The collaboration will prototype AI systems including SIMA 2, an agent that plays and learns alongside users, and Genie 3 for generating interactive worlds.

Partnership with game developers announced August 21, 2026
Research spans 15 years from Atari to EVE Online
Developing SIMA 2 agent and Genie 3 for world generation

Sources: Google DeepMind AI Web Searched, DeepMind Blog RSS

6.5 Fine-Tuning Long-Context Language Models with Sparse Attention¶

A new fine-tuning method enables efficient long-context inference in transformers through key-value cache selection and compression.

Sources: arXiv — Computation and Language RSS

Security¶

7 Inadvertent Context Leakage in Language Models¶

Research published on arXiv reveals that sensitive data stored in language model context windows creates hidden security vulnerabilities, enabling secret reconstruction from benign model outputs even when direct extraction is refused. Across eight proprietary models tested, 2-digit secrets were reconstructed with near-perfect accuracy and 4-digit secrets at 82% match rates.

2-digit in-context secrets reconstructed with near-perfect accuracy
4-digit secrets recovered at 82% exact match rate
More capable models leak more information
Developed novel adaptive black-box attack technique
RL-trained adversary extracted full Social Security Numbers in production-style agent test

Sources: arXiv AI Web Searched, arXiv — Machine Learning RSS, arXiv — Cryptography and Security RSS

6 HARP: Adaptive CVE Prioritization System¶

New method ranks cybersecurity vulnerabilities based on operational preferences rather than fixed scoring.

Sources: arXiv — Computation and Language RSS, arXiv — Cryptography and Security RSS

AI Security¶

6.5 TempJail: Temporal Jailbreak Attack on Vision-Language Models¶

Research demonstrates video-based jailbreak attacks against large vision-language models via subtitle manipulation.

Sources: arXiv — Computation and Language RSS, arXiv — Computer Vision RSS

AI Research¶

6 PEA-DPO: Perception-Enhanced Alignment for Multimodal Models¶

Research extends Direct Preference Optimization to multimodal large language models for improved human alignment.

Sources: arXiv — Computation and Language RSS, arXiv — Computer Vision RSS

ML Architecture¶

6 Asymmetric Attention Heads in Transformer Models¶

A study proposes asymmetric attention mechanisms that allocate different context spans to individual heads in transformer models based on their functional roles.

Sources: arXiv — Computation and Language RSS

AI Applications¶

6 Using LLM Hallucinations to Generate Scientific Hypotheses¶

Researchers explore a multi-agent architecture that leverages language model hallucinations as a creative tool to generate testable scientific hypotheses.

Sources: arXiv — Computation and Language RSS

6 Conversational AI Reduces Intergroup Bias in Immigration Attitudes¶

An experiment demonstrates that conversational AI employing common identity framing can reduce us-versus-them boundaries and increase pro-immigrant helping behavior.

Sources: arXiv — Computation and Language RSS

AI Behavior¶

6 Longitudinal Analysis: LLM Creativity Convergence Over Three Years¶

A three-year study examines whether large language models are becoming increasingly similar in their performance on open-ended creative tasks.

Sources: arXiv — Computation and Language RSS