Multimodal large language models (MLLMs) have achieved strong performance on perception-oriented tasks, yet their ability to perform mathematical spatial reasoning, defined as the capacity to parse and manipulate two- and three-dimensional relations, remains unclear. While humans easily master textbook-style spatial problems, current MLLMs exhibit substantially higher error rates, exposing a fundamental weakness in spatial comprehension. To investigate this gap, we present MathSpatial, the first large-scale and systematic dataset resource dedicated to mathematical spatial reasoning in MLLMs. MathSpatial provides two complementary subsets: MathSpatial-Bench (2,000 problems) for evaluating spatial reasoning isolated from perceptual noise, and MathSpatial-Corpus (8,000 problems) equipped with structured reasoning traces for training. Benchmarking 20+ MLLMs reveals a severe deficit: non-reasoning models largely languish below 60% accuracy, and even the strongest reasoning models fall ~16 percentage points behind human performance (96.3%). Crucially, we demonstrate that fine-tuning on MathSpatial-Corpus yields consistent performance gains across model families. MathSpatial provides a vital resource for diagnosing and advancing spatial reasoning in MLLMs. The dataset and related resources are available at https://github.com/shuolucs/MathSpatial.

Extracting skills from past successes is critical for the efficient evolution of Large Language Model (LLM) agents. Prevailing agent self-evolution paradigms typically rely on a core assumption: equipping LLMs with skill memories derived from successful trajectories will monotonically improve their problem-solving capabilities. However, probe analyses reveal that extracting skills solely from successful trajectories traps the model in a Skill Imitation Trap. For tasks that resemble past successes but require different tools, retrieving more skills paradoxically increases the model’s confidence in wrong tool calls—procedure skills raise the wrong-tool margin by 47% over a memory-free baseline. To overcome this limitation, we propose Boundary-Aware Skill Memory (BASM), which augments each skill with explicit boundary fields—applicability conditions, risk cues, avoidance rules, and recovery notes. These fields transform each retrieved skill from an unconditional action template into state-conditioned guidance: the agent applies the skill when its conditions hold, suppresses inapplicable tool calls when they do not, and issues targeted repairs when execution fails. Across three agent benchmarks and four model scales, BASM consistently outperforms success-distilled skill-memory baselines: it improves task success rate by up to 23.8% on AppWorld, accuracy by up to 5.0% on BFCL, and reduces attack success rate by 4.6% on AgentDojo, while simultaneously reducing average AppWorld steps by up to 6.6% relative to the memory-free baseline.

Test-time reinforcement learning (TTRL) always adapts models at inference time via pseudo-labeling, leaving it vulnerable to spurious optimization signals from label noise.Through an empirical study, we observe that responses with medium consistency form an ambiguity region and constitute the primary source of reward noise.Crucially, we find that such spurious signals can be even amplified through group-relative advantage estimation.Motivated by these findings, we propose a unified framework, Debiased and Denoised test-time Reinforcement Learning (DDRL), to mitigate spurious signals.Concretely, DDRL first applies a frequency-based sampling strategy to exclude ambiguous samples while maintaining a balanced set of positive and negative examples.It then adopts a debiased advantage estimation with fixed advantages, removing the bias introduced by group-relative policy optimization.Finally, DDRL incorporates a consensus-based off-policy refinement stage, which leverages the rejection-sampled dataset to enable efficient and stable model updates.Experiments on three large language models across multiple mathematical reasoning benchmarks demonstrate that DDRL consistently outperforms existing TTRL baselines.The code is available at https://github.com/yuyongcan/DDRL.

As LLM-based agents operate over sequential multi-step reasoning, hallucinations arising at intermediate steps risk propagating along the trajectory, thus degrading overall reliability. Unlike hallucination detection in single-turn responses, diagnosing hallucinations in multi-step workflows requires identifying which step causes the initial divergence. To fill this gap, we propose a new research task, automated hallucination attribution of LLM-based agents, aiming to identify the step responsible for the hallucination and explain why. To support this task, we introduce AgentHallu, a comprehensive benchmark with: (1) 693 high-quality trajectories spanning 7 agent frameworks and 5 domains, (2) a hallucination taxonomy organized into 5 categories (Planning, Retrieval, Reasoning, Human-Interaction, and Tool-Use) and 14 sub-categories, and (3) multi-level annotations curated by humans, covering binary labels, hallucination-responsible steps, and causal explanations. We evaluate 13 leading models, and results show the task is challenging even for top-tier models (like GPT-5, Gemini-2.5-Pro). The best-performing model achieves only 41.1% step localization accuracy, where tool-use hallucinations are the most challenging at just 11.6%. We believe AgentHallu will catalyze future research into developing robust, transparent, and reliable agentic systems.

Large reasoning models have recently demonstrated remarkable capabilities in solving complex tasks, where supervised fine-tuning (SFT) on long chain-of-thought data serves as a crucial foundation for eliciting and enhancing their reasoning abilities. Despite rapid progress in both improving and analyzing reasoning-oriented SFT, the field still lacks a systematic survey that consolidates its fast-growing literature. To fill this gap, we present a comprehensive review of recent advancements in reasoning SFT, examining the literature through the dual lenses of methodological design and analytical investigation. First, we review methodological improvements across the SFT pipeline and categorize them into data-centric approaches and algorithm-centric innovations. Second, we reorganize analytical studies along three dimensions: data characteristics, optimization dynamics, and mechanistic insights. Finally, by synthesizing current research focuses and remaining bottlenecks, we outline promising future directions for reasoning SFT. We hope this survey deepens the understanding of reasoning SFT and paves the way for advanced reasoning models.

Test-Time Reinforcement Learning (TTRL) enables Large Language Models (LLMs) to enhance reasoning capabilities on unlabeled test streams by deriving pseudo-rewards from majority voting consensus.However, existing TTRL methods rely exclusively on positive pseudo-labeling strategies.Such reliance becomes vulnerable under challenging scenarios where answer distributions are highly dispersed, resulting in weak consensus that inadvertently reinforces incorrect trajectories as supervision signals.In this paper, we propose SCRL (Selective-Complementary Reinforcement Learning), a robust test-time reinforcement learning framework that effectively mitigates label noise amplification.SCRL develops Selective Positive Pseudo-Labeling, which enforces strict consensus criteria to filter unreliable majorities. Complementarily, SCRL introduces Entropy-Gated Negative Pseudo-Labeling, the first negative supervision mechanism in TTRL, to reliably prune incorrect trajectories based on generation uncertainty. Extensive experiments on mathematical reasoning benchmarks demonstrate that SCRL achieves substantial improvements over baselines, while maintaining robust generalization and training stability under constrained rollout budgets.

Pre-training on large, high-quality datasets is essential for improving the reasoning abilities of Large Language Models (LLMs), particularly in specialized fields like mathematics. However, the field of Multimodal LLMs (MLLMs) lacks a comprehensive, open-source dataset for mathematical reasoning. To fill this gap, we present InfiMM-WebMath-40B, a high-quality dataset of interleaved image-text documents. It consists of 24 million web pages, 85 million image URLs, and 40 billion text tokens, all carefully extracted and filtered from CommonCrawl. We outline our data collection and processing pipeline in detail. Models trained on InfiMM-WebMath-40B demonstrate strong performance in both text-only and multimodal settings, setting a new state-of-the-art on multimodal math benchmarks such as MathVerse and We-Math.

Recently, scaling test-time compute on Large Language Models (LLM) has garnered wide attention. However, there has been limited investigation of how various reasoning prompting strategies perform as scaling. In this paper, we focus on a standard and realistic scaling setting: majority voting. We systematically conduct experiments on 6 LLMs × 8 prompting strategies × 6 benchmarks. Experiment results consistently show that as the sampling time and computational overhead increase, complicated prompting strategies with superior initial performance gradually fall behind simple Chain-of-Thought.We analyze this phenomenon and provide theoretical proofs. Additionally, we propose a probabilistic method to efficiently predict scaling performance and identify the best prompting strategy under large sampling times, eliminating the need for resource-intensive inference processes in practical applications.Furthermore, we introduce two ways derived from our theoretical analysis to significantly improve the scaling performance. We hope that our research can promote to re-examine the role of complicated prompting, unleash the potential of simple prompting strategies, and provide new insights for enhancing test-time scaling performance. Code is available at https://github.com/MraDonkey/rethinking_prompting.

In this work, we present InfiMM, an advanced Multimodal Large Language Model that adapts to intricate vision-language tasks. InfiMM, inspired by the Flamingo architecture, distinguishes itself through the utilization of large-scale training data, comprehensive training strategies, and diverse large language models. This approach ensures the preservation of Flamingo’s foundational strengths while simultaneously introducing augmented capabilities. Empirical evaluations across a variety of benchmarks underscore InfiMM’s remarkable capability in multimodal understanding. The code can be found at: https://anonymous.4open.science/r/infimm-zephyr-F60C/.