9471 papers
This paper demonstrates that language models can be fine-tuned to reliably detect and identify activation steering interventions, revealing that such steering is not inherently undetectable and that models trained to recognize it may paradoxically become more susceptible to behavioral manipulation.
This paper introduces DPAC, a diffusion guidance method that projects adversarial gradients onto the tangent space of iso-density surfaces to minimize path-space KL divergence and control energy, thereby theoretically and empirically achieving higher sample quality (lower FID) while maintaining target classification success.
The paper proposes Deep FlexQP, a deep unfolding-based solver that accelerates nonlinear programming by learning dimension-agnostic parameters for a robust, always-feasible convex QP relaxation, thereby significantly improving the speed and success rates of SQP and safety filter applications while providing rigorous performance guarantees.
The paper introduces Guided Flow Policy (GFP), a novel offline reinforcement learning method that couples a multi-step flow-matching policy with a distilled one-step actor to selectively focus on high-value actions, achieving state-of-the-art performance across diverse benchmarks by overcoming the limitations of indiscriminate behavior regularization.
This paper demonstrates that injecting a canonical action ordering signal into the reward function during RL post-training significantly improves Transformer performance on Zebra puzzles compared to optimizing for task success alone, even when the model is fine-tuned on randomized solution sequences.
This paper proposes a multi-loss learning framework for speech emotion recognition that integrates energy-adaptive mixup and frame-level attention to address data scarcity and emotional complexity, achieving state-of-the-art performance across four benchmark datasets.
This paper introduces a post-training method that induces extreme sparsity in transformer attention (reducing connectivity to ~0.4%) without sacrificing performance, thereby revealing simplified, interpretable task-specific circuits and unifying feature-based and circuit-based perspectives on model behavior.
ClinNoteAgents is a novel LLM-based multi-agent system that effectively predicts and interprets 30-day heart failure readmission risks by transforming unstructured clinical notes into structured risk factors and clinician-style abstractions, offering a scalable and interpretable solution for data-limited healthcare settings.
The paper introduces InternGeometry, an LLM agent enhanced by Complexity-Boosting Reinforcement Learning and a dynamic memory mechanism that iteratively proposes and verifies auxiliary constructions, achieving a medalist-level performance on IMO geometry problems with significantly less training data than previous expert models.
ReFusion introduces a novel masked diffusion model that integrates sequence reorganization with a hybrid parallel-autoregressive decoding strategy to simultaneously achieve full KV cache efficiency, reduce learning complexity, and significantly outperform existing diffusion models while narrowing the performance gap with autoregressive models.
HydroGEM is a self-supervised, zero-shot hybrid TCN-Transformer foundation model that effectively performs continental-scale streamflow quality control by detecting and reconstructing sensor anomalies with high accuracy and cross-national generalization, thereby addressing the scalability limitations of manual hydrological data validation.
This paper introduces RePo, a novel mechanism that leverages a differentiable module to dynamically re-position tokens based on contextual dependencies rather than fixed linear indices, thereby reducing extraneous cognitive load and enhancing LLM performance on tasks involving noisy contexts, structured data, and long-range dependencies.
This paper introduces MCP-SafetyBench, a comprehensive benchmark leveraging real-world Model Context Protocol (MCP) servers to evaluate the safety of large language models in multi-turn, cross-tool scenarios, revealing that current models remain vulnerable to diverse MCP-specific attacks despite a significant safety-utility trade-off.
This paper introduces FluenceFormer, a transformer-driven, two-stage framework that leverages a physics-informed Fluence-Aware Regression loss to achieve superior, geometry-aware fluence map prediction for radiotherapy planning, significantly outperforming existing CNN and single-stage methods in energy conservation and structural fidelity.
Yukthi Opus is a multi-chain hybrid metaheuristic that combines Markov Chain Monte Carlo exploration, greedy local search, and adaptive simulated annealing to achieve robust, budget-efficient optimization for large-scale NP-hard problems.
This paper presents a cost- and latency-aware benchmark demonstrating that while tool-augmented planning significantly improves accuracy for complex knowledge-intensive tasks like Event-QA, it often incurs prohibitive latency costs and offers no benefit—or even degrades performance—for tasks like persuasive response generation where simple one-shot prompting is more efficient.
The paper introduces PFUA, a tool-augmented reasoning agent that outperforms text-only models in protein function prediction by integrating domain-specific tools and external biological priors to generate verifiable evidence, rather than relying on ineffective internal chain-of-thought reasoning.
The paper proposes NeuronLLM, a novel framework that improves task-level controllability in Large Language Models by identifying both facilitative "good" and inhibitive "bad" neurons through contrastive learning and augmented question sets to overcome the limitations of existing ability-specific methods.
The paper introduces the Spectral Sphere Optimizer (SSO), a novel parallel training algorithm that enforces strict module-wise spectral constraints on both weights and updates to achieve full Maximal Update Parametrization alignment, resulting in superior convergence, stability, and performance across diverse large-scale architectures compared to AdamW and Muon.
EmboTeam is a novel framework that enhances embodied multi-robot collaboration by cascading LLM-based instruction parsing into formal PDDL planning and reactive behavior tree execution, achieving significantly higher task success rates on the new MACE-THOR benchmark compared to existing baselines.