Tag
The paper introduces CIRCE, which uses a cheap LLM to extract a problem's constraint structure into a constraint graph that predicts LLM response correctness before inference, improving NLL and AUROC over existing baselines across seven reasoning benchmarks.
This paper proposes a lightweight evolutionary heuristic scheduler to optimize denoising trajectories in diffusion large language models, addressing failure modes like EOS Overflow and Proximal Bias, and outperforming baselines on reasoning and planning benchmarks.
This paper introduces an unsupervised metric based on attention dispersion to detect hallucinations in large language models, showing significant AUC improvements on mathematical reasoning benchmarks using the Qwen2.5 model family.
This paper proposes TASCO, a framework for test-time adaptation in LLMs that incorporates local stability into confidence-based optimization to improve reasoning accuracy and token efficiency without updating model parameters.
The paper investigates whether performance gains in on-policy self-distillation come from learning privileged reference information or recovering existing reasoning behavior, finding that the correct reference does not consistently benefit performance across various conditions.
Sapient Intelligence has released HRM-Text, a 1B parameter text generation model, trained on only 0.04 trillion tokens (costing approximately $1000), surpassing much larger models trained on 100-1000 times more data on multiple reasoning benchmarks, marking the beginning of a new paradigm for AI training.
This tweet announces Fast-Slow Training (FST), a new continual learning method that treats model parameters as slow weights and optimized context as fast weights, reportedly outperforming weights-only training on math, code, and general reasoning benchmarks.
AEvo is a meta-editing framework that improves iterative agentic search by separating proposal and evaluation into two roles and using accumulated memory to guide future search. It achieves a 26% relative gain over baselines and state-of-the-art results on open-ended optimization tasks.
This paper introduces LogiHard, a framework that uses combinatorial hardening to expose compositional failures in frontier LLMs, demonstrating significant accuracy drops in logical reasoning tasks.