[2508.15260] Deep Think with Confidence

kept by eddie jan 12

Deep Think with Confidence (DeepConf) uses model confidence signals to filter low-quality reasoning traces, achieving up to 99.9% accuracy on AIME 2025 with 84.7% fewer tokens.

Which AI Lies Best? A game theory classic designed by John Nash
A benchmark using the game "So Long Sucker" finds that simpler LLMs can outperform in complex scenarios and that models adjust their honesty based on opponent strength.
also on GPT-OSS, #benchmarks, #ai-research, #ai
9 Observations from Building with AI Agents
Building AI agents reveals nine lessons including using best models for prototyping, fine-tuning for stable tasks, static typing to reduce hallucinated code, and closed-loop prompt optimization.
also on Qwen 3
What (I think) makes Gemini 3 Flash so good and fast
Gemini 3 Flash likely uses a massive 1.2 trillion-parameter ultra-sparse MoE architecture, activating only 5–30 billion parameters per inference to deliver high intelligence at low cost, though it suffers from token bloat and a 91% hallucination rate on refusals.
also on #scaling-laws, #machine-learning, #ai-research, #ai
How To Scale Your Model
This online book teaches LLM scaling through roofline analysis, TPU/GPU hardware, Transformer math, and parallelism techniques to estimate training and inference costs.
also on #scaling-laws, #machine-learning, #ai-research
These Mathematicians Are Trying to Educate A.I. - The New York Times
Researchers are manually evaluating AI's failure to solve advanced math problems, revealing significant gaps in reasoning that automated benchmarks miss.
also on #benchmarks, #machine-learning, #ai-research