A new survey paper provides a deep dive into post-training methodologies for Large Language Models (LLMs), analyzing their role in refining LLMs beyond pretraining. It addresses key challenges such as catastrophic forgetting, reward hacking, and inference-time trade-offs, and highlights emerging directions in model alignment, scalable adaptation, and inference-time reasoning. The paper also provides a public repository to continually track developments in this fast-evolving field.
A new method is proposed to reduce the verbosity of LLMs in step-by-step reasoning by retaining moderately easy problems during Reinforcement Learning with Verifiable Rewards (RLVR) training. This approach acts as an implicit length regularizer, preventing the model from excessively increasing output length on harder problems. Experiments using Qwen3-4B-Thinking-2507 show the model achieves baseline accuracy with nearly twice shorter solutions.
Researchers introduce UnsafeChain, a new safety alignment dataset designed to improve the safety of large reasoning models (LRMs) by focusing on 'hard prompts' that elicit harmful outputs. The dataset identifies and corrects unsafe completions into safe responses, exposing models to unsafe behaviors and guiding their correction. Fine-tuning LRMs on UnsafeChain demonstrates enhanced safety and preservation of general reasoning ability compared to existing datasets like SafeChain and STAR-1.
Researchers at MBZUAI have introduced EvoLMM, a self-evolving framework for large multimodal models that enhances reasoning capabilities without human-annotated data or reward distillation. EvoLMM uses two cooperative agents, a Proposer and a Solver, which generate image-grounded questions and solve them through internal consistency, using a continuous self-rewarding process. Evaluations using Qwen2.5-VL as the base model showed performance gains of up to 3% on multimodal math-reasoning benchmarks like ChartQA, MathVista, and MathVision using only raw training images.
This paper introduces ProgramFC, a fact-checking model that decomposes complex claims into simpler sub-tasks using a library of functions. The model uses LLMs to generate reasoning programs and executes them by delegating sub-tasks, enhancing explainability and data efficiency. Experiments on fact-checking datasets demonstrate ProgramFC's superior performance compared to baseline methods, with publicly available code and data.