The paper “LLMs Improving LLMs” introduces AutoTTS, an environment where a Claude‑based explorer agent automatically searches test‑time scaling policies, achieving up to 69.5% token savings and superior accuracy on unseen models, all for $39.9 and 160 minutes without any LLM calls during evaluation.
The article analyzes how the PRISM framework redesigns test‑time scaling for discrete diffusion language models by replacing costly Best‑of‑N sampling with a three‑stage hierarchical search, local branching via partial remasking, and self‑verified feedback, achieving large accuracy gains on math and code benchmarks while cutting inference compute by up to four‑fold.
The SimpleTES framework decomposes trial‑and‑error into three scalable dimensions—Concurrency, Length, and Candidates—enabling test‑time scaling that lets open‑source models outperform closed‑source rivals across 21 diverse scientific benchmarks, from LASSO regression to quantum circuit compilation.
The paper introduces Squeeze Evolve, a validator‑free multi‑model evolutionary framework that orchestrates diverse large language models to break the performance ceiling of any single model, delivering up to 23‑point accuracy improvements and 1.4‑3.3× cost reductions across math, vision, and scientific benchmarks.
Alibaba’s new Qwen3‑Max‑Thinking model adds inference‑time scaling and adaptive tool use, delivering large gains on math, coding, and agent benchmarks while remaining closed‑source, and it offers drop‑in OpenAI‑compatible API access at the cost of higher latency and token usage.
The paper introduces RoboMonkey, a framework that applies a generate‑and‑verify paradigm and test‑time scaling to Vision‑Language‑Action models, showing that increasing sampling and verification at inference dramatically reduces action error across multiple VLA architectures, and presents scalable verifier training, synthetic data augmentation, and efficient deployment strategies.
DiMo-GUI is a plug‑and‑play framework that dramatically improves multimodal large language models' ability to locate GUI elements by using a hierarchical dynamic visual reasoning loop and modality‑aware optimization, achieving up to double the performance on high‑resolution GUI benchmarks without any additional training data.
The EvoSearch method introduced by HKUST and Kuaishou’s KuaLing team leverages test‑time scaling to dramatically improve diffusion‑based image and video generation without training, using evolutionary search along the denoising trajectory, achieving state‑of‑the‑art results on SD2.1, Flux‑1‑dev and other models.
EvoSearch, a test‑time evolutionary search method, dramatically improves image and video generation by increasing inference compute without extra training, outperforming existing scaling techniques on diffusion and flow models while maintaining robustness and diversity across multiple benchmarks.
The article explains the Generator‑Verifier Gap (GVG)—the asymmetry where verifying a solution is far cheaper than generating it—covers its origin, its impact on test‑time scaling for large language models, reinforcement‑learning approaches, and how the concept can shape agent architectures and AI product strategy.
This article surveys the latest research on large language model reasoning, highlighting test‑time scaling methods, chain‑of‑thought variants, and novel inference‑time techniques that boost performance while exposing trade‑offs, costs, and future directions for AI developers.
The article critically examines whether the pre‑training Scaling Law still applies to Grok 3, compares its compute usage and model size with DeepSeek and OpenAI models, evaluates the cost‑effectiveness of pre‑training, RL and test‑time scaling, and explores how these insights shape future large‑language‑model development strategies.
This article analyses the training tricks behind OpenAI's o1 model, explaining test/inference‑time scaling laws, post‑training techniques, process‑supervised reward models (PRM), various inference‑time search methods, data‑collection pipelines, and the trade‑offs between allocating compute to pre‑training versus inference.
