AI Briefing#optimization
The latest AI and developer news about #optimization, with the original source and a short summary.
The latest AI and developer news about #optimization, with the original source and a short summary.
The method uses simulated annealing on BF16 weights to fix reasoning loops in Qwen3.5-0.8B, requiring only $0.90 in GPU time and zero LLM tokens.
The framework runs hundreds of latent-reasoning agents on Qwen3 8B, 14B, and 32B models to reach scores that took other frameworks over an hour.
The 2016 paper by Karimi, Nutini, and Schmidt established that the Polyak–Łojasiewicz condition guarantees linear convergence for standard training methods, bridging a gap between theory and practice.
GPT-6 Astra improved the MATH-500 score of Qwen3.5-0.8B to 48.1% by modifying its weights without training.
MIT CSAIL researchers demonstrate that GPT-6 Astra can read and write neural network weights, improving Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% by editing 0.016% of its parameters to fix reasoning loops.
The method achieves 100% accuracy on 11 test items with a p50 latency of 22ms for dynamic options, or ~1.5ms for cached candidates.
Adaptive EVS reduces VLM input tokens by 80% and improves concurrent stream throughput by 46%
A new paper accepted at NeurIPS introduces adaptive representations for Functional Gradient Descent, ensuring convergence to the global minimizer and often outperforming neural networks by an order of magnitude.
Researchers from Apple and Georgia Tech introduce LIPPAX to mitigate client drift and improve convergence rates in federated variational inequalities.
Imp brings DSPy's declarative, self-improving language-model programming paradigm to Elixir, leveraging OTP for reliable, concurrent agent execution.
A new pull request introduces tiled matrix multiplication for k-quants in llama.cpp, leveraging VNNI instructions to achieve 3-7x speedups in CPU prompt processing.
The open-source library tool-prune has been released, reducing prompt tokens by 92% through pre-filtering of tool schemas.
Sakana AI proposed PC-ALM, which resolves the signal attenuation issue in existing Predictive Coding (PC), successfully training 1000-layer neural networks.
Catalyst has released a tool that measures and verifies key elements among 52 variables in compiled Rust models in just 1.37 seconds.
Nvidia has released SoL-Pi, an extension featuring four mechanisms to reduce token usage and latency for Pi agents.
An LLM iteratively evolved optimization algorithms to break 10 existing records for the circle packing problem.
The open-source library ParaRNN, which supports parallel training of nonlinear RNNs, has been released, achieving up to a 200x speedup.
A roundup of the latest AI research papers focusing on reducing inference costs, modularizing systems, and enhancing agent autonomy.
bitsandbytes developer Tim Dettmers has previewed a new quantization methodology, demonstrating high inference performance.
llama.cpp has improved prompt processing speed by up to 31% through vectorization using F16C intrinsics.
The new DD method ensures asymptotic alignment between training and test errors at each iteration step.
This article introduces an efficient technique that induces hallucinations instead of using structured outputs from LLMs, then performs actual classification via vector similarity.
An extension tool that adds an autonomous experimentation and optimization loop to the AI coding agent pi.
It proposes a federated learning optimization technique that applies second-order computation without increasing clients' communication load or privacy cost.