The paper proposes MBD-LM, a Multi-Block approach designed to improve the parallel decoding performance of Block Diffusion Language Models (BD-LMs).
Google Research has unveiled TabFM, a model that processes tabular data without separate fine-tuning.
A new knowledge graph structure is proposed that treats relationships (Edges) as independent embedded documents rather than simple metadata.
The PACE framework accurately predicts performance on costly LLM agent benchmarks using only a small amount of data.
It covers the role and structure of Harness Engineering as a key element enabling AI's Recursive Self-Improvement (RSI).
We propose a new training objective to address the embedding condensation phenomenon that occurs in small language models.
Leanstral 1.5, an open-source model based on Lean 4 specialized for Formal Verification, has been released.
A deterministic folding engine that improves context management efficiency for AI agents has been released.
Current AI has launched Gap Map v0.1, which indexes the technical gaps and current status of the open source AI ecosystem.
An experimental Scatter sampler has been introduced to llama.cpp, which locally smooths the probability distribution to improve generation quality.
The number of discovered security vulnerabilities (CVEs) surged following the launch of Anthropic's Claude Mythos Preview.
A technique called CDD has been announced that recovers the exact original text of LLM fine-tuning data using only the difference in Logit values, without requiring access to weights.
Introduces a technique that dramatically reduces API costs by using images instead of text to cut down on LLM input token counts.
The Portuguese government has released Amalia, a 9B-parameter open-source LLM.
ReFreeKV, a threshold-free KV Cache compression technique that automatically adjusts the compression budget based on input data, is proposed.
A technique is proposed that maximizes the efficiency of Diffusion-based Speculative Decoding by predicting the optimal block size according to the input data.
ELDR routing, which accounts for expert locality in MoE models, cuts decode latency by up to 13.9%.
Amid growing concerns over profitability relative to the massive investments in the AI industry, the price per unit of AI service usage is falling.
AMD has unveiled Micro-World, an action-controllable interactive world model based on Wan2.1.
Hugging Face has launched a public dataset showing usage patterns of coding agents that use the Hub.
Anthropic has released a guide mapping intelligence levels between Claude Sonnet 5 and 4.6.
Google will discontinue the consumer version of Gemini Code Assist for GitHub on July 17.
Proposes a new paradigm that compiles natural language specifications into lightweight neural network artifacts to dramatically lower inference costs.
Palantir's CEO said that some US government clients are shifting to open-source AI.
AI Briefing