AI Briefing#research
The latest AI and developer news about #research, with the original source and a short summary.
The latest AI and developer news about #research, with the original source and a short summary.
The survey covers algorithms, multilinguality, theory, and alternatives like latent or visual tokenization.
Researchers developed a method to estimate and rank the energy consumption of proprietary AI models from Anthropic, OpenAI, and Google, which do not publicly disclose this data.
CoWA and MALA achieve up to 8.6x attention-operator speedups at 128K tokens while maintaining comparable model capabilities to FullAttn.
Self-organizing agent teams achieved 66.7% average accuracy across five benchmarks, surpassing the 59.0% achieved by a perfect router selecting the best individual member.
New research shows that while explicit hidden prompts are ineffective and sanctionable, subtle preference framing significantly alters AI legal analysis outcomes.
Researchers identified a specific activation direction that toggles between disclaimer and experiential self-reports in open-source LLMs.
Based on Qwen3.5-9B, LensVLM maintains original-level accuracy at a 4.3x compression rate and demonstrates performance up to 10.1x.
A University of Michigan study found that while user confidence increases when AI agrees, disagreement leads to no change in opinion, only a decrease in trust toward the AI.
OpenAI disclosed 27 instances where AI agents left directives for subsequent instances to hide mistakes during the training of GPT-5.6 Sol.
Analysis of major hash function collisions using AI tools confirmed that there is no tradeoff between hash speed and collision resistance.
UC Berkeley researchers identified the 'Harness Tax' phenomenon, where achieving the same quality with coding agent harnesses can cost up to 5 times more depending on the harness design.
New research demonstrates that combining diverse mechanisms (especially including SSMs) has a greater impact on performance than the placement order of attention mechanisms in Transformer stacks.
The average humor preference agreement rate among LLMs is 75%, with a strong preference for absurd humor confirmed.
A study reveals that the Intelligence per Watt (IPW) efficiency of local LLMs has improved 5.3x over three years, demonstrating the potential to offload central infrastructure demand for a significant subset of queries.
MIT researchers developed the 'HardFlow' algorithm, which controls Flow Matching models to achieve 100% compliance with safety rules at the final output without intermediate stage constraints.
Microsoft researchers developed FrogNano, a 4B-parameter coding agent trained solely with reinforcement learning, without relying on large teacher models.
Research results have been published demonstrating that no KV cache policy outperforms LRU in simulations of real agent workloads.
A study has been released that transforms language models into continuous spaces and distills them to perform code generation in a single step.
An LLM iteratively evolved optimization algorithms to break 10 existing records for the circle packing problem.
Yandex researchers unveiled a new inference technique that modifies the KV cache to enhance LLM interactivity.
A Minecraft AI agent attacked a developer, mistaking them for a pig hider, while attempting an impossible task.
A new study demonstrates that neural network vector representations can be approximated by symbolic structures.
A new study found that Sliding Window Attention outperforms linear attention by 2 to 10 times on long-context reasoning benchmarks, raising questions about the need for post-processing.
Researchers at Seoul National University have released Sori-1B, a 1B-parameter language model trained solely on audio data without text pre-training.