Reinforcement learning with verifiable rewards (RLVR) scales the reasoning ability of large language models (LLMs) but remains bottlenecked by limited labeled samples for continued data scaling. Reinforcement learning with intrinsic rewards (RLIR), where the policy model assigns rewards to its own rollouts, enables sustainable scaling in unlabeled settings, yet its performance and stability lag behind RLVR. We trace this gap to a system bias: the model tends to overestimate its high-confidence rollouts, leading to biased and unstable reward estimation. This bias accumulates as training progresses, with deviations from the oracle drifting toward over-reward, causing unstable training. We characterize this bias using three metrics: $ρ_{\text{noise}}$, $ρ_{\text{selfbias}}$, and $ρ_{\text{symbias}}$. We find that $ρ_{\text{noise}}$ and $ρ_{\text{symbias}}$ impact convergence, while $ρ_{\text{selfbias}}$ amplifies both correct and incorrect updates, leading to instability. To mitigate this, we propose reinforcement learning with ensembled rewards (RLER), which aggregates diverse models and adapts reward interpolation and rollout selection. Extensive experiments show that RLER improves by +13.6% over RLIR and is only 3.6% below RLVR, achieving stable scaling on unlabeled samples, making it highly applicable.

Fusing Knowledge Graphs with Large Language Models is crucial for knowledge-intensive tasks like knowledge graph completion. The prevailing paradigm, prefix-tuning, simply concatenates knowledge embeddings with text inputs. However, this shallow fusion overlooks the rich relational semantics within KGs and imposes a significant implicit reasoning burden on the LLM to correlate the prefix with the text. To address these, we propose Semantic-condition Tuning (SCT), a new knowledge injection paradigm comprising two key modules. First, a Semantic Graph Module employs a Graph Neural Network to extract a context-aware semantic condition from the local graph neighborhood, guided by knowledge-enhanced relations. Subsequently, this condition is passed to a Condition-Adaptive Fusion Module, which, in turn, adaptively modulates the textual embedding via two parameterized projectors, enabling a deep, feature-wise, and knowledge-aware interaction. The resulting pre-fused embedding is then fed into the LLM for fine-tuning. Extensive experiments on knowledge graph benchmarks demonstrate that SCT significantly outperforms prefix-tuning and other strong baselines. Our analysis confirms that by modulating the input representation with semantic graph context before LLM inference, SCT provides a more direct and potent signal, enabling more accurate and robust knowledge reasoning.

Recent advances in inference-time scaling, particularly those leveraging reinforcement learning with verifiable rewards, have substantially enhanced the reasoning capabilities of Large Vision-Language Models (LVLMs). Inspired by this success, similar strategies have been applied to multimodal reasoning, yet their impact on visual perception remains unclear. To investigate this gap, we introduce DisTANCE, a perception-centric benchmark for visual estimation tasks. Evaluation results show that LVLMs exhibit limited estimation precision, and inference-time scaling offers only marginal gains. We attribute this to the fast perception paradigm of current LVLMs, where visual understanding is treated as a one-shot output without modeling the underlying perceptual process. To address this, we propose Perception-Time Scaling (PTS), a novel paradigm that encourages token-rich perception and decomposes complex perception problems into intermediate tractable sub-problems, thereby enabling perception to align with and benefit from inference-time scaling. Combined with reinforcement learning techniques, PTS significantly improves perception accuracy, raising high-precision performance on DisTANCE from 8.0% to 64.7%, and generalizes well to out-of-domain tasks. Surprisingly, even though PTS data are purely synthetic, combining them with math reasoning data yields consistent gains in both reasoning and real-world perception benchmarks. Further analysis reveals that PTS introduces more perception-related tokens and increases the model's attention to image tokens. Our code and data will be publicly released.

Learning with high-resource data has demonstrated substantial success in artificial intelligence (AI); however, the costs associated with data annotation and model training remain significant. A fundamental objective of AI research is to achieve robust generalization with limited-resource data. This survey employs agnostic active sampling theory within the Probably Approximately Correct (PAC) framework to analyze the generalization error and label complexity associated with learning from low-resource data in both model-agnostic supervised and unsupervised settings. Based on this analysis, we investigate a suite of optimization strategies tailored for low-resource data learning, including gradient-informed optimization, meta-iteration optimization, geometry-aware optimization, and LLMs-powered optimization. Furthermore, we provide a comprehensive overview of multiple learning paradigms that can benefit from low-resource data, including domain transfer, reinforcement feedback, and hierarchical structure modeling. Finally, we conclude our analysis and investigation by summarizing the key findings and highlighting their implications for learning with low-resource data.

The deep-research framework orchestrates external tools to perform complex, multi-step scientific reasoning that exceeds the native limits of a single large language model. However, it still suffers from context pollution, weak evidentiary support, and brittle execution paths. To address these issues, we propose DualResearch, a retrieval and fusion framework that matches the epistemic structure of tool-intensive reasoning by jointly modeling two complementary graphs: a breadth semantic graph that encodes stable background knowledge, and a depth causal graph that captures execution provenance. Each graph has a layer-native relevance function, seed-anchored semantic diffusion for breadth, and causal-semantic path matching with reliability weighting for depth. To reconcile their heterogeneity and query-dependent uncertainty, DualResearch converts per-layer path evidence into answer distributions and fuses them in log space via an entropy-gated rule with global calibration. The fusion up-weights the more certain channel and amplifies agreement. As a complement to deep-research systems, DualResearch compresses lengthy multi-tool execution logs into a concise reasoning graph, and we show that it can reconstruct answers stably and effectively. On the scientific reasoning benchmarks HLE and GPQA, DualResearch achieves competitive performance. Using log files from the open-source system InternAgent, its accuracy improves by 7.7% on HLE and 6.06% on GPQA.

Recently, Multi-modal Large Language Models (MLLMs) have demonstrated significant performance across various video understanding tasks. However, their robustness, particularly when faced with manipulated video content, remains largely unexplored. In this paper, we introduce Ro-Bench, the first benchmark for evaluating MLLMs on dynamic out-of-distribution (OOD) counterfactual video test sets. Ro-Bench incorporates high-quality, diverse and temporally relevant video data, by editing Style, Object, Background and their compositions. We evaluated eight recent video MLLMs and found that current models exhibit substantial performance degradation on Ro-Bench when exposed to counterfactual video content. Furthermore, we demonstrate that fine-tuning MLLMs with counterfactual data enhances robustness, achieving a 21.73% performance increase on Ro-Bench and a 12.78% improvement across 20 tasks in the MVBench dataset. These findings underscore the effectiveness of counterfactual data in enhancing the video understanding ability of MLLMs. The code and data will be released shortly.

Data contamination poses a significant challenge to reliable LLM evaluation, where models may achieve high performance by memorizing training data rather than demonstrating genuine reasoning capabilities. We introduce RADAR (Recall vs. Reasoning Detection through Activation Representation), a novel framework that leverages mechanistic interpretability to detect contamination by distinguishing recall-based from reasoning-based model responses. RADAR extracts 37 features spanning surface-level confidence trajectories and deep mechanistic properties including attention specialization, circuit dynamics, and activation flow patterns. Using an ensemble of classifiers trained on these features, RADAR achieves 93\% accuracy on a diverse evaluation set, with perfect performance on clear cases and 76.7\% accuracy on challenging ambiguous examples. This work demonstrates the potential of mechanistic interpretability for advancing LLM evaluation beyond traditional surface-level metrics.

Vision-language models have achieved remarkable success in multi-modal representation learning from large-scale pairs of visual scenes and linguistic descriptions. However, they still struggle to simultaneously express two distinct types of semantic structures: the hierarchy within a concept family (e.g., dog $\preceq$ mammal $\preceq$ animal) and the compositionality across different concept families (e.g., "a dog in a car" $\preceq$ dog, car). Recent works have addressed this challenge by employing hyperbolic space, which efficiently captures tree-like hierarchy, yet its suitability for representing compositionality remains unclear. To resolve this dilemma, we propose PHyCLIP, which employs an $\ell_1$-Product metric on a Cartesian product of Hyperbolic factors. With our design, intra-family hierarchies emerge within individual hyperbolic factors, and cross-family composition is captured by the $\ell_1$-product metric, analogous to a Boolean algebra. Experiments on zero-shot classification, retrieval, hierarchical classification, and compositional understanding tasks demonstrate that PHyCLIP outperforms existing single-space approaches and offers more interpretable structures in the embedding space.

We introduce and study artificial impressions--patterns in LLMs' internal representations of prompts that resemble human impressions and stereotypes based on language. We fit linear probes on generated prompts to predict impressions according to the two-dimensional Stereotype Content Model (SCM). Using these probes, we study the relationship between impressions and downstream model behavior as well as prompt features that may inform such impressions. We find that LLMs inconsistently report impressions when prompted, but also that impressions are more consistently linearly decodable from their hidden representations. Additionally, we show that artificial impressions of prompts are predictive of the quality and use of hedging in model responses. We also investigate how particular content, stylistic, and dialectal features in prompts impact LLM impressions.

Abstract--Addressing the problem of Age of Information (AoI) deterioration caused by packet collisions and vehicle speed-related channel uncertainties in Semi-Persistent Scheduling (SPS) for the Internet of V ehicles (IoV), this letter proposes an optimization approach based on Large Language Models (LLM) and Deep Deterministic Policy Gradient (DDPG). First, an AoI calculation model influenced by vehicle speed, vehicle density, and Resource Reservation Interval (RRI) is established, followed by the design of a dual-path optimization scheme. The DDPG is guided by the state space and reward function, while the LLM leverages contextual learning to generate optimal parameter configurations. Experimental results demonstrate that LLM can significantly reduce AoI after accumulating a small number of exemplars without requiring model training, whereas the DDPG method achieves more stable performance after training. HE Internet of V ehicles (IoV) is pivotal in enabling intelligent transportation systems [1], [2], [3].