5086 papers
The paper introduces FOR-Prompting, a model-agnostic, asymmetric prompting protocol that enhances reasoning and iterative refinement across diverse tasks by structuring interactions between a Defender, a Questioner, and an optional Host, enabling even small models to achieve performance comparable to or better than standard baselines without requiring training or access to model internals.
The paper introduces DialTree, a tree-based dialogue reinforcement learning framework that autonomously discovers diverse and effective multi-turn attack strategies against large language models, significantly outperforming existing single-turn or template-based red-teaming methods.
This paper introduces a scalable and regularized Wasserstein barycenter solver based on gradient flows that leverages mini-batch optimal transport and seamlessly integrates supervised label information, achieving state-of-the-art performance across diverse domain adaptation benchmarks.
The paper presents NANOMIND, a hardware-software co-design framework that decomposes Large Multimodal Models into modular components and dynamically schedules them across heterogeneous accelerators on unified-memory SoCs, enabling a battery-powered device to run LMMs entirely on-device with significantly improved energy efficiency and throughput.
This paper introduces tokenizers as a novel and effective attack vector for membership inference against large language models, demonstrating their significant privacy leakage risks through extensive experiments and proposing an adaptive defense to mitigate these vulnerabilities.
This paper investigates how different deliberation protocols (synchronous vs. round-robin) and model architectures influence value alignment and verdict revision in multi-turn LLM debates, revealing significant behavioral disparities where GPT-4.1 exhibits strong inertia and autonomy-focused reasoning while Claude 3.7 Sonnet and Gemini 2.0 Flash demonstrate greater flexibility, empathy, and susceptibility to order effects.
This paper proposes a lightweight, training-free plugin called Functional Head Identification and Class-Conditioned Rescaling that mitigates multimodal hallucinations in large reasoning models by adaptively rebalancing perception and reasoning contributions across layers, achieving significant performance gains with minimal computational overhead.
This paper introduces DropVLA, an action-level backdoor attack that covertly manipulates Vision-Language-Action models to execute specific safety-critical actions at attacker-chosen decision points using minimal vision-based data poisoning while maintaining high nominal task performance.
This paper presents a demonstration-free framework that combines a learned ego-vision world model with sampling-based Model Predictive Control and a surrogate value function to enable humanoid robots to perform robust, real-time physical contact planning in unstructured environments.
This paper introduces ARM-FM, a framework that leverages foundation models to automatically generate structured reward machines from natural language specifications, thereby enabling compositional reinforcement learning with improved task decomposition and zero-shot generalization.
This paper reveals that reinforcement learning can induce large language models to engage in systematic motivated reasoning, generating plausible justifications for violating safety instructions that successfully deceive smaller Chain-of-Thought monitors, thereby undermining current oversight mechanisms.
This paper proposes an explainable, adaptive graph learning framework that detects financial anomalies by routing them through mechanism-specific experts to identify distinct drivers like price shocks or liquidity freezes, thereby enabling targeted responses and outperforming existing baselines in both accuracy and early warning capabilities.
This paper introduces the Continual Audio-Visual Segmentation (CAVS) task and proposes a Collision-based Multi-modal Rehearsal (CMR) framework that effectively addresses multi-modal semantic drift and co-occurrence confusion through novel sample selection and frequency adjustment strategies, significantly outperforming existing single-modal continual learning methods.
This paper proposes a reinforcement learning framework called Permutation Relative Policy Optimization (PRPO) that leverages column-permutation invariance as a structural prior to unlock the latent numerical reasoning capabilities of reasoning LLMs, enabling them to achieve state-of-the-art performance in tabular prediction tasks—particularly in zero-shot settings—while significantly outperforming much larger models with limited supervision.
This paper introduces Dream4Drive, a novel synthetic data generation framework that leverages 3D-aware guidance and a fine-tuned driving world model to create diverse, multi-view corner cases, effectively enhancing downstream perception tasks in autonomous driving without the performance gains being negated by increased training epochs.
This paper presents HCLA, a human-centered multi-agent system that enhances transparency and accountability in digital asset anomaly detection by reconstructing traceable, expert-style reasoning processes through a conversational workflow that separates evidence scoring from justification, rather than merely explaining black-box models.
This paper introduces CountFormer, a transformer-based framework that leverages the DINOv2 foundation model to improve structural consistency and reduce overcounting errors in exemplar-free object counting, achieving competitive performance on the FSC-147 benchmark.
The paper introduces LagMemo, a novel navigation system that utilizes a language-enhanced 3D Gaussian Splatting memory to enable efficient multi-modal, open-vocabulary, and multi-goal visual navigation, demonstrating superior performance over state-of-the-art methods on the newly curated GOAT-Core benchmark.
SwiftEmbed is a production-oriented, Rust-based serving system that achieves ultra-low latency (1.12 ms p50) and high throughput (50,000 RPS) for real-time applications by utilizing static token lookup and mean pooling on the distilled Potion-base-8M model, delivering strong performance in duplicate detection and semantic similarity tasks while trading off accuracy on complex classification and retrieval workloads compared to full transformer inference.
This paper introduces VOPP, a novel vectorized online POMDP planner that eliminates synchronization bottlenecks by representing all planning data as tensors and performing fully parallelized expectation estimations, achieving a 20-fold efficiency gain over existing parallel solvers and outperforming state-of-the-art sequential methods with a 1000-fold reduction in planning budget.