Tag
This paper identifies 'on-policy audit collapse,' a failure mode where refreshing a component verifier against full-outcome labels on improved-policy data causes the audit to systematically prefer worse verifiers. The authors derive an exact inversion threshold and propose repairing audits by preserving own-failure support in the refresh distribution.
NVIDIA 提出 Elastic Latent Recurrence (ELR) 训练配方，使循环（looped）语言模型在单次循环时达到与同规模非循环模型相当的精度，同时每次额外循环都能以更多推理计算换取精度提升，最终产出可在 1-4 循环间弹性的单一检查点 Koi-4B。
该论文研究了推理模型在不同思考预算下是否表现出更少的认知偏误，结果发现增加推理token并不会降低偏误，部分偏误甚至与人类方向相反，表明测试时推理不能保证决策理性。
该论文提出 InfiLoop，一种为循环 Transformer 设计的原生残差连接机制，通过内容加权与学习式时间衰减维持循环状态的运行摘要，使模型在数万次测试时迭代中持续提升推理性能，在 Sudoku-Extreme 上达到 97.9% 精度、ARC-AGI-2 上 13.6% pass@2。
A paper from Meta Superintelligence Labs shows that pre-trained LLMs equipped with a light inference harness can act as capable agents and even outperform RL post-trained counterparts when given a sufficient test-time sampling budget.
This workshop paper audits how a cumulative token cap falling inside a mathematical derivation affects test-time controllers, comparing strict (stop at cap) versus advisory (finish current attempt) stopping across 19,200 archived traces. It finds that budget curves must jointly report cap, realized cost, eligible candidates, and selector information, since candidate coverage and nominal budgets alone can mislead.
The paper proposes TAPS (Trajectory Adaptive Progress–Fluctuation Scheduler), which adaptively adjusts the step size of recurrent updates in looped transformers by decomposing terminal loss sensitivity into persistent progress and centered fluctuation, improving accuracy without retraining and delivering up to 1.56× wall-clock speedup when applied during training.
The paper demonstrates that test-time communication among AI agents can significantly outperform independent parallel attempts on challenging tasks like ARC-AGI-3, achieving state-of-the-art results in research-oriented domains.
The author shares a novel use case for Jev in building custom verifiers for AI agent harnesses, enabling scalable test-time compute by combining System One and System Two models.
OpenAI researcher Noam Brown discusses how multi-agent systems scale test-time compute through parallelization, highlights an experiment with 10,000 agents, and stresses the critical role of the underlying general model in driving AI progress.
State of Thought (SoT) proposes endogenous reasoning for LLMs by extracting internal dynamics-geometric states to selectively activate historical support, improving accuracy and efficiency across various reasoning tasks.
This paper introduces Elo-per-token analysis to study how LLM agents allocate test-time compute, revealing that agents initially outperform independent sampling but slow down over time, with parallel sessions offering performance gains.
OpenAI researcher Noam Brown predicts that within a year, AI accessible to everyone will be capable of solving Millennium Prize problems, citing advancements in test-time compute scaling.
This paper introduces a benchmark for revision propagation in conversationally generated artifacts using LLMs and evaluates cost-effective test-time compute methods, showing that parallel sampling with selection improves accuracy.
This paper presents a specialization pipeline for post-training language models to achieve gold-medal performance in coding competitions, demonstrating top scores on IOI benchmarks using techniques like supervised fine-tuning and reinforcement learning.
The paper proposes CASE, a dynamic selection combiner using a decodability criterion to predict when hidden-state selection outperforms majority voting in large language models, enhancing reliability on difficult questions.
The article summarizes eight inference-time techniques to improve LLM reasoning, based on research from Google, OpenAI, and Anthropic, with tradeoffs and practical notes.
The paper 'Thought-Level Beam Search for Reasoning' introduces Gambit, a method that actively reallocates test-time compute during generation by pruning low-scoring reasoning traces and branching from strong prefixes, and it has been accepted to COLM 2026.
This paper introduces an exam-style evaluation to study how reasoning models allocate a shared test-time compute budget across multiple questions. It finds that models fail to strategically ration compute, instead prioritizing questions by presentation order and ignoring value or difficulty.
An observation that Claude Opus 5's max thinking leads to performance degradation on ~20-30% of benchmarks compared to xhigh, contrary to the expectation that more test-time compute improves performance.