Language models (LMs) and their extension, vision-language models (VLMs), have achieved remarkable performance across various tasks. However, they still struggle with complex reasoning tasks that require multimodal or multilingual real-world knowledge. To support such capabilities, an external memory system that can efficiently provide relevant multimodal information is essential. Existing approaches generally concatenate image and text tokens into a long sequence as memory, which, however, may drastically increase context length and even degrade performance. In contrast, we propose using continuous memory-a compact set of dense embeddings-to more effectively and efficiently represent multimodal and multilingual knowledge. Our key insight is that a VLM can serve as its own continuous memory encoder. We empirically show that this design improves performance on complex multimodal reasoning tasks. Building on this, we introduce a data-efficient and parameter-efficient method to fine-tune the VLM into a memory encoder, requiring only 1.2% of the model's parameters and a small corpus of 15.6K self-synthesized samples.

We introduce EvaLearn, a pioneering benchmark designed to evaluate large language models (LLMs) on their learning capability and efficiency in challenging tasks, a critical, yet underexplored aspect of model potential. EvaLearn contains 648 challenging problems across six task types, grouped into 182 sequences, each sequence dedicated to one task type. Diverging from most existing benchmarks that evaluate models in parallel, EvaLearn requires models to solve problems sequentially, allowing them to leverage the experience gained from previous solutions. EvaLearn provides five comprehensive automated metrics to evaluate models and quantify their learning capability and efficiency. We extensively benchmark nine frontier models and observe varied performance profiles: some models, such as Claude-3.7-sonnet,

Large Language Models (LLMs) exhibit impressive performance on complex reasoning tasks, yet they frequently fail on basic numerical problems, producing incorrect outputs. Inspired by Benford's Law, a statistical pattern in which lower digits occur more frequently as leading digits, we hypothesize that the skewed digit distributions in web-collected corpora may be learned by LLMs during pretraining, leading to biased numerical generation. To investigate the hypothesis, we first examine whether digits frequencies in pretraining corpus (OLMo2) follows Benford's law. We then construct an evaluation benchmark in which the ground-truth digits are uniformly distributed within each of the seven numerical reasoning tasks. Our evaluation results demonstrate that leading open-source LLMs show a consistent pattern of digit bias that resembles Benford's law. Through logit-lens tracing and neuron-level dissection, we identify that this bias arises predominantly from a small subset of highly digit-selective feed-forward network (FFN) neurons in the deeper layers. Finally, we demonstrate that pruning these neurons mitigates imbalanced overgeneration and partially corrects erroneous outputs, providing causal evidence that fine-grained pretraining digit bias can propagate into model behavior. Our findings reveal a fundamental connection between corpus-level statistics and symbolic failure modes in LLMs, offering a new lens for diagnosing and mitigating hallucinations in numerical tasks.

Reward Models (RMs) are crucial to aligning large language models (LLMs), but the degree to which an RM specialized to one task (e.g.

Self-improvement via RL often fails on complex reasoning tasks because GRPOstyle post-training methods rely on the model's initial ability to generate positive samples. Without guided exploration, these approaches merely reinforce what the model already knows (distribution-sharpening) rather than enabling the model to solve problems where it initially generates no correct solutions. To unlock reasoning ability in such settings, the model must explore new reasoning trajectories beyond its current output distribution. Such exploration requires access to sufficiently good positive samples to guide the learning. While expert demonstrations seem like a natural solution, we find that they are often ineffective in RL post-training.

Computer-Aided Design (CAD) is pivotal in industrial manufacturing, with orthographic projection reasoning foundational to its entire workflow--encompassing design, manufacturing, and simulation. However, prevailing deep-learning approaches employ standard 3D reconstruction pipelines as an alternative, which often introduce imprecise dimensions and limit the parametric editability required for CAD workflows. Recently, some researchers adopt vision-language models (VLMs), particularly supervised fine-tuning (SFT), to tackle CAD-related challenges. SFT shows promise but often devolves into pattern memorization, resulting in poor out-of-distribution (OOD) performance on complex reasoning tasks. To tackle these limitations, we introduce CReFT-CAD, a two-stage finetuning paradigm: first, a curriculum-driven reinforcement learning stage with difficulty-aware rewards to steadily build reasoning abilities; second, supervised post-tuning to refine instruction following and semantic extraction. Complementing this, we release TriView2CAD, the first large-scale, open-source benchmark for orthographic projection reasoning, comprising 200,000 synthetic and 3,000 real-world orthographic projections with precise dimensional annotations and six interoperable data modalities. Benchmarking leading VLMs on orthographic projection reasoning, we show that CReFT-CAD significantly improves reasoning accuracy and OOD generalizability in real-world scenarios, providing valuable insights to advance CAD reasoning research.

Existing methods to enhance the reasoning capability of large language models predominantly rely on supervised fine-tuning (SFT) followed by reinforcement learning (RL) on reasoning-specific data. These approaches critically depend on external supervisions-such as labeled reasoning traces, verified golden answers, or pre-trained reward models. In this work, we propose Entropy Minimized Policy Optimization (EMPO), which makes an early attempt at fully unsupervised LLM reasoning incentivization. By minimizing the semantic entropy of LLMs on unlabeled questions, EMPO achieves competitive performance compared to supervised counterparts. Specifically, without any external supervision, EMPO boosts the accuracy of Qwen2.5-Math-7BBase from 33.7% to 51.6% on math benchmarks and improves the accuracy of Qwen2.5-7BBase from 32.1% to 50.1% on MMLU-Pro. Primary analysis are also provided to interpret the effectiveness of EMPO.

Foundation models (FMs) are pre-trained on large-scale datasets and then finetuned for a specific downstream task. The most common fine-tuning method is to update pretrained weights via low-rank adaptation (LoRA). Existing initialization strategies for LoRA often rely on singular value decompositions (SVD) of gradients or weight matrices. However, they do not provably maximize the expected gradient signal, which is critical for fast adaptation. To this end, we introduce Explained Variance Adaptation (EVA), an initialization scheme that uses the directions capturing the most activation variance, provably maximizing the expected gradient signal and accelerating fine-tuning.

Visual reasoning abilities play a crucial role in understanding complex multimodal data, advancing both domain-specific applications and artificial general intelligence (AGI). Existing methods improve Vision-Language Models (VLMs) reasoning via Chain-of-Thought (CoT) supervised fine-tuning, using meticulously annotated training data to enhance visual reasoning capabilities. However, this training paradigm may lead to overfitting and cognitive rigidity, restricting the model's generalization ability to transfer visual reasoning skills under domain shift and limiting its real-world applicability. To address these limitations, we propose Reason-RFT, the first two-stage reinforcement fine-tuning framework for visual reasoning: (1) Supervised Fine-Tuning (SFT) with curated CoT data activates the reasoning potential of VLMs, followed by (2) Group Relative Policy Optimization (GRPO)-based reinforcement learning that generates multiple reasoning-response pairs, significantly enhancing the capability to address ubiquitous domain shift in visual reasoning tasks. To evaluate the visual reasoning capabilities of Reason-RFT, we reconstructed a comprehensive dataset encompassing visual counting, structural perception, and spatial transformation, serving as a benchmark for systematic assessment across three core dimensions. Experimental results demonstrate three key advantages: (1) Performance Enhancement: achieving state-of-the-art results across multiple tasks, outperforming mainstream open-source and proprietary models; (2) Generalization Superiority: consistently maintaining robust performance in addressing domain shift in typical visual reasoning tasks, outperforming alternative paradigms; (3) Data Efficiency: excelling in few-shot learning scenarios while surpassing full-dataset SFT baselines. Reason-RFT introduces a rebust training paradigm in visual reasoning, and please refer to project website: Reason-RFT.

Direct Preference Optimization (DPO) is a widely used preference optimization algorithm in large language model (LLM) alignment, which reparameterizes the reward function in reinforcement learning with human feedback (RLHF) without requiring a separate reward model. However, during the DPO training process, when a large negative gradient is applied to low-confidence samples, LLMs with a softmax output head tend to squeeze the confidence in the model's output distribution towards the highest-confidence sentence, which may lead to a decrease in the confidence of both preference and non-preference samples, while increasing the confidence of unrelated tokens. This phenomenon becomes more complex in reasoning tasks. In this work, focusing on reasoning tasks, we propose VPO, a negative gradient constraint method for human non-preference samples based on $\mathcal{V}$-usable information. By using $\mathcal{V}$-usable information to measure the similarity between preference pairs and selectively constrain the negative gradient, VPO can alleviate the squeezing effect of DPO, enhance alignment with the generation objective, and maintain the model's ability to distinguish between preference and non-preference samples. We compare VPO with DPO and its latest variants on mathematical reasoning tasks using the LLama 3.1 and Qwen 2.5 series, including both Base and Instruct models. Our results demonstrate that VPO consistently and significantly outperforms existing methods.