Extrapolation Merging: Keep Improving With Extrapolation and Merging

Large Language Models (LLMs) require instruction fine-tuning to perform different downstream tasks. However, the instruction fine-tuning phase still demands significant computational resources and labeled data, lacking a paradigm that can improve model performance without additional computational power and data. Model merging aims to enhance performance by combining the parameters of different models, but the lack of a clear optimization direction during the merging process does not always guarantee improved performance. In this paper, we attempt to provide a clear optimization direction for model merging. We first validate the effectiveness of the model extrapolation method during the instruction fine-tuning phase. Then, we propose Extrapolation Merging, a paradigm that can continue improving model performance without requiring extra computational resources or data. Using the extrapolation method, we provide a clear direction for model merging, achieving local optimization search, and consequently enhancing the merged model's performance. We conduct experiments on seven different tasks, and the results show that our method can consistently improve the model's performance after fine-tuning.

SuperGPQA: Scaling LLM Evaluation across 285 Graduate Disciplines

Large language models (LLMs) have demonstrated remarkable proficiency in mainstream academic disciplines such as mathematics, physics, and computer science. However, human knowledge encompasses over 200 specialized disciplines, far exceeding the scope of existing benchmarks. The capabilities of LLMs in many of these specialized fields-particularly in light industry, agriculture, and service-oriented disciplines-remain inadequately evaluated. To address this gap, we present SuperGPQA, a comprehensive benchmark that evaluates graduate-level knowledge and reasoning capabilities across 285 disciplines. Our benchmark employs a novel Human-LLM collaborative filtering mechanism to eliminate trivial or ambiguous questions through iterative refinement based on both LLM responses and expert feedback. Our experimental results reveal significant room for improvement in the performance of current state-of-the-art LLMs across diverse knowledge domains (e.g., the reasoning-focused model DeepSeek-R1 achieved the highest accuracy of 61.82% on SuperGPQA), highlighting the considerable gap between current model capabilities and artificial general intelligence. Additionally, we present comprehensive insights from our management of a large-scale annotation process, involving over 80 expert annotators and an interactive Human-LLM collaborative system, offering valuable methodological guidance for future research initiatives of comparable scope.

SKIL: Semantic Keypoint Imitation Learning for Generalizable Data-efficient Manipulation

Real-world tasks such as garment manipulation and table rearrangement demand robots to perform generalizable, highly precise, and long-horizon actions. Although imitation learning has proven to be an effective approach for teaching robots new skills, large amounts of expert demonstration data are still indispensible for these complex tasks, resulting in high sample complexity and costly data collection. To address this, we propose Semantic Keypoint Imitation Learning (SKIL), a framework which automatically obtain semantic keypoints with help of vision foundation models, and forms the descriptor of semantic keypoints that enables effecient imitation learning of complex robotic tasks with significantly lower sample complexity. In real world experiments, SKIL doubles the performance of baseline methods in tasks such as picking a cup or mouse, while demonstrating exceptional robustness to variations in objects, environmental changes, and distractors. For long-horizon tasks like hanging a towel on a rack where previous methods fail completely, SKIL achieves a mean success rate of 70\% with as few as 30 demonstrations. Furthermore, SKIL naturally supports cross-embodiment learning due to its semantic keypoints abstraction, our experiments demonstrate that even human videos bring considerable improvement to the learning performance. All these results demonstrate the great success of SKIL in achieving data-efficint generalizable robotic learning. Visualizations and code are available at: https://skil-robotics.github.io/SKIL-robotics/.

RoboHorizon: An LLM-Assisted Multi-View World Model for Long-Horizon Robotic Manipulation

Efficient control in long-horizon robotic manipulation is challenging due to complex representation and policy learning requirements. Model-based visual reinforcement learning (RL) has shown great potential in addressing these challenges but still faces notable limitations, particularly in handling sparse rewards and complex visual features in long-horizon environments. To address these limitations, we propose the Recognize-Sense-Plan-Act (RSPA) pipeline for long-horizon tasks and further introduce RoboHorizon, an LLM-assisted multi-view world model tailored for long-horizon robotic manipulation. In RoboHorizon, pre-trained LLMs generate dense reward structures for multi-stage sub-tasks based on task language instructions, enabling robots to better recognize long-horizon tasks. Keyframe discovery is then integrated into the multi-view masked autoencoder (MAE) architecture to enhance the robot's ability to sense critical task sequences, strengthening its multi-stage perception of long-horizon processes. Leveraging these dense rewards and multi-view representations, a robotic world model is constructed to efficiently plan long-horizon tasks, enabling the robot to reliably act through RL algorithms. Experiments on two representative benchmarks, RLBench and FurnitureBench, show that RoboHorizon outperforms state-of-the-art visual model-based RL methods, achieving a 23.35% improvement in task success rates on RLBench's 4 short-horizon tasks and a 29.23% improvement on 6 long-horizon tasks from RLBench and 3 furniture assembly tasks from FurnitureBench.

SOLAMI: Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters

Human beings are social animals. How to equip 3D autonomous characters with similar social intelligence that can perceive, understand and interact with humans remains an open yet foundamental problem. In this paper, we introduce SOLAMI, the first end-to-end Social vision-Language-Action (VLA) Modeling framework for Immersive interaction with 3D autonomous characters. Specifically, SOLAMI builds 3D autonomous characters from three aspects: (1) Social VLA Architecture: We propose a unified social VLA framework to generate multimodal response (speech and motion) based on the user's multimodal input to drive the character for social interaction. (2) Interactive Multimodal Data: We present SynMSI, a synthetic multimodal social interaction dataset generated by an automatic pipeline using only existing motion datasets to address the issue of data scarcity. (3) Immersive VR Interface: We develop a VR interface that enables users to immersively interact with these characters driven by various architectures. Extensive quantitative experiments and user studies demonstrate that our framework leads to more precise and natural character responses (in both speech and motion) that align with user expectations with lower latency.

Distractor-free Generalizable 3D Gaussian Splatting

We present DGGS, a novel framework addressing the previously unexplored challenge of Distractor-free Generalizable 3D Gaussian Splatting (3DGS). It accomplishes two key objectives: fortifying generalizable 3DGS against distractor-laden data during both training and inference phases, while successfully extending cross-scene adaptation capabilities to conventional distractor-free approaches. To achieve these objectives, DGGS introduces a scene-agnostic reference-based mask prediction and refinement methodology during training phase, coupled with a training view selection strategy, effectively improving distractor prediction accuracy and training stability. Moreover, to address distractor-induced voids and artifacts during inference stage, we propose a two-stage inference framework for better reference selection based on the predicted distractor masks, complemented by a distractor pruning module to eliminate residual distractor effects. Extensive generalization experiments demonstrate DGGS's advantages under distractor-laden conditions. Additionally, experimental results show that our scene-agnostic mask inference achieves accuracy comparable to scene-specific trained methods. Homepage is \url{https://github.com/bbbbby-99/DGGS}.

SRA-MCTS: Self-driven Reasoning Augmentation with Monte Carlo Tree Search for Code Generation

Large language models demonstrate exceptional performance in simple code generation tasks but still face challenges in tackling complex problems. These challenges may stem from insufficient reasoning and problem decomposition capabilities. To address this issue, we propose a reasoning-augmented data generation process, SRA-MCTS, which guides the model to autonomously generate high-quality intermediate reasoning paths. This creates a positive feedback loop, enabling continuous improvement. Our method operates entirely through the model itself without requiring additional supervision. By synthesizing natural language reasoning paths and translating them into executable code, the approach ensures analytical accuracy and enhances the success rate in solving complex tasks. Experimental results show that, even without additional supervisory signals, our method achieves performance improvements across different model scales, demonstrating the significant potential of self-improvement in small models. Furthermore, the method remains robust when traditional Chain-of-Thought (CoT) approaches exhibit performance degradation, with notable improvements observed in diversity metrics such as pass@10. We encourage further exploration of reasoning processes within training data to enhance the ability of language models to address complex problems.

Reviving Dormant Memories: Investigating Catastrophic Forgetting in Language Models through Rationale-Guidance Difficulty

Although substantial efforts have been made to mitigate catastrophic forgetting in continual learning, the intrinsic mechanisms are not well understood. In this paper, we discover that when a forgetting model passively receives an externally provided partial appropriate rationale, its performance on the forgotten task can be restored. Furthermore, by simply adding a task-agnostic prefix to the original instruction, the forgetting model can actively generate an appropriate rationale to reach the correct answer. These findings suggest that the model does not actually ``forget'' the task knowledge; instead, the degraded performance can be attributed to the failure of the original instructions in guiding the model to generate the appropriate rationales. Based on this insight, we propose the Rationale-Guidance Difficulty metric to evaluate how effectively a given instruction guides the model in generating appropriate rationales. We apply this metric to optimize the allocation of replay data in replay-based continual learning algorithm. Experimental results demonstrate that our data allocation method effectively mitigates catastrophic forgetting and maintains better model plasticity simultaneously across models.

PSPO*: An Effective Process-supervised Policy Optimization for Reasoning Alignment

Process supervision enhances the performance of large language models in reasoning tasks by providing feedback at each step of chain-of-thought reasoning. However, due to the lack of effective process supervision methods, even advanced large language models are prone to logical errors and redundant reasoning. We claim that the effectiveness of process supervision significantly depends on both the accuracy and the length of reasoning chains. Moreover, we identify that these factors exhibit a nonlinear relationship with the overall reward score of the reasoning process. Inspired by these insights, we propose a novel process supervision paradigm, PSPO*, which systematically outlines the workflow from reward model training to policy optimization, and highlights the importance of nonlinear rewards in process supervision. Based on PSPO*, we develop the PSPO-WRS, which considers the number of reasoning steps in determining reward scores and utilizes an adjusted Weibull distribution for nonlinear reward shaping. Experimental results on six mathematical reasoning datasets demonstrate that PSPO-WRS consistently outperforms current mainstream models.

METEOR: Evolutionary Journey of Large Language Models from Guidance to Self-Growth

Model evolution enables learning from feedback to refine experiences and update skills, transforming models from having no domain knowledge to becoming domain experts. However, there is currently no unified and effective method for guiding this evolutionary process. To address this gap, we propose the Meteor method, which includes three training phases: weak-to-strong data distillation, iterative training, and self-evolution strategies. Each phase maximizes the model's inherent domain capabilities, allowing it to autonomously refine its domain knowledge and enhance performance. Experiments demonstrate that our approach significantly improves accuracy, completeness, relevance, coherence, and reliability across domain-specific tasks.