AIME 2026
PulseAugur coverage of AIME 2026 — every cluster mentioning AIME 2026 across labs, papers, and developer communities, ranked by signal.
6 day(s) with sentiment data
-
WebAI releases TwIL-8B reasoning model, tops MATH-500 benchmark
WebAI has launched TwIL-8B, an 8.8 billion parameter reasoning model built upon IBM's Granite 4.2 8B. This model excels in mathematics and coding tasks, achieving a score of 98.2 on the MATH-500 benchmark, and is design…
-
Limite 1B Violetto math model deployed on vLLM, memory cap affects concurrency not speed
The Limite 1B Violetto, a 2.1 GB math model, achieved 94.01% accuracy on the AIME 2026 exam. Researchers from Paradigma trained the model from scratch and released it under Apache 2.0, along with a vLLM plugin for servi…
-
New CIKA framework uses LLMs to pinpoint causal concepts for math reasoning
Researchers have developed CIKA, a novel framework designed to enhance Large Language Model (LLM) mathematical reasoning by identifying concepts that causally contribute to correct answers. Unlike previous methods that …
-
New framework streamlines formal proof generation for math problems
Researchers have developed a novel framework for generating formal proofs in systems like Lean 4, which integrates candidate answer selection and proof generation into a single process. This joint approach aims to reduc…
-
New frameworks enhance multi-agent system training with iterative refinement and self-play
Two new research papers introduce novel frameworks for training multi-agent systems (MAS) that go beyond simple one-shot orchestration. MAS-Loop focuses on closing the feedback loop between an orchestrator and a refiner…
-
Reflection AI releases 501B parameter open-weight model Beam, targeting cost efficiency
Reflection AI has released Beam, an open-weight, sparse Mixture-of-Experts (MoE) model with 501 billion total parameters that activates only 23 billion per task. This design aims to offer the knowledge capacity of a lar…
-
AI performance claims need external validation, not self-reporting
This article argues that AI performance claims should be evaluated based on external validation rather than self-reported metrics. It emphasizes that true performance is demonstrated when a third party administers tests…
-
Model-Level RSI: Changing LLM Weights Without Human Labels
This post introduces the concept of Model-Level Recursive Self-Improvement (RSI), distinguishing it from other forms of LLM advancement. It outlines five stages of self-improvement, with Model-Level RSI being the fourth…
-
Hugging Face consolidates 48 AI benchmarks into interactive map
Hugging Face has launched a new interactive map consolidating its 48 official benchmarks, providing a comprehensive overview of AI model performance across various domains. The map reveals that while agent-related bench…
-
7B model ZGCM-1 prioritizes tool use and large context over memorization
Researchers from Zhongguancun Academy and Zhongguancun Institute of AI have developed ZGCM-1, a 7.39B parameter model that prioritizes tool use and a large context window over memorizing vast datasets. This approach all…
-
GigaChat 3.5 Reasoning benchmark claims scrutinized for misleading efficiency metrics
A recent analysis of AI benchmark tables highlights discrepancies in how performance claims are presented, using the GigaChat 3.5 Reasoning model and DeepSeek V4 Flash Preview as a case study. While GigaChat's published…
-
AI models tackle unsolved math problems with formal proofs and new discoveries
Two new research papers explore the capabilities of large language models in advanced mathematical reasoning and discovery. The first paper introduces Magenta, a system that bridges informal natural language math proble…
-
Qwen3.8-27B achieves 29/30 on AIME 2026 math benchmark with FP8
A benchmark test of the Qwen3.8-27B model on the AIME 2026 math dataset revealed that its quantized FP8 weights, when set to xhigh reasoning effort, achieved a score of 29/30. This performance was comparable to the BF16…
-
Google DeepMind's DiffusionGemma achieves 1500 tokens/sec via discrete diffusion
Google DeepMind has released DiffusionGemma, an open-weight language model that utilizes discrete diffusion for text generation, offering significantly faster output speeds compared to traditional autoregressive models.…
-
New method uses wrong drafts to boost LLM math capabilities
Researchers have developed a novel technique called "Weak-to-Strong Elicitation via Mismatched Wrong Drafts" to improve the capabilities of large language models. This method involves using mathematically incorrect draf…
-
AI benchmark scores predictable from just two factors, study finds
A new research paper proposes a method called BenchPress that can predict a frontier model's performance across numerous benchmarks using only two key scores. The study analyzed 84 models and 133 benchmarks, finding tha…
-
Google's Gemma 4 26B model runs locally with LM Studio's new headless CLI
Google's Gemma 4 model family, particularly the 26B-A4B variant, is now accessible for local inference on consumer hardware like MacBooks. This mixture-of-experts model activates only a fraction of its parameters per in…