A new scheduling algorithm, Smallest Volume First (SVF), reduces LLM inference latency by optimizing key-value cache management. Theoretical analysis shows a worst-case competitive ratio reduced from 48 to 5, with 1-bit SVF achieving strong performance using minimal information. Evaluations on Llama-3.1 models confirm improvements in both average and tail latency, with the approach integrated into vLLM.
HDSO enables safe, auditable skill updates for LLM agents without training, using falsifiable hypotheses and validation. On ALFWorld, it improves Qwen3-8B by +6.9 Avg. SR points and maintains a +7.1-point gain under noisy feedback, with validated skills transferable across runs and models when diagnostic alignment is achieved.
FAPS is the first function-space posterior sampling framework that unifies stochastic-process regression and PDE inverse problems. It uses pretrained flow-matching priors and Langevin correction with low-rank covariance preconditioning to enable efficient, accurate posterior inference from sparse, noisy data with coherent uncertainty quantification.
HRLLI introduces a hierarchical reinforcement learning framework that adapts natural-language instructions dynamically during decision-making. It decomposes instructions into stage-specific guidance elements and uses a select-to-act paradigm to enable real-time selection of relevant instruction pieces, improving sample efficiency and performance in complex environments.
SAFER is a training-free framework that enhances robustness of test-time adaptation by using reliability-guided augmentation. It generates stochastic augmentations, pools predictions via correlation-weighted aggregation with outlier detection, and includes adaptive mixing to preserve clean performance under adversarial attacks. Evaluations on PACS, VLCS, and OfficeHome show improved resilience without sacrificing clean accuracy.
Parsimoniously activated dictionary learning (PADL) establishes a structured generative model with auxiliary latent variables, enabling maximum a posteriori estimation. This framework provides generalization guarantees and an analytical characterization of the tradeoff between sparsity, storage cost, and reconstruction accuracy, allowing data-driven hyperparameter estimation. The resulting algorithm achieves better reconstruction performance and accelerates inference in vision-language models.
LIHA reveals a small set of first-token broadcaster heads in GPT-2 that persistently attend to the initial prompt token, driving language switches. Instruction tuning reorganizes these circuits, concentrating language identity at early layers, as seen in Qwen2.5-1.5B-Instruct and confirmed in Chinese and Russian language handling at layer 0.
ARIA addresses contextual tunneling in LLMs by conditioning knowledge use on mechanistic completeness. It uses a three-tier cascade for causal reasoning, physics-informed transfer, and parametric fallback, and improves materials discovery through auditable, physically grounded reasoning.
HyperAdapter introduces a hypergraph-based adapter that performs structured, group-aware adaptation in vision transformers by operating in hyperedge space rather than token space. It uses prototype-based assignments to build a soft hypergraph, aggregates token features into hyperedge representations, applies lightweight adaptation, and diffuses updates back via hypergraph structure, enabling explicit structural inductive bias while maintaining efficiency. Experiments show consistent performance gains over baseline PEFT methods, especially on tasks requiring structured reasoning.
MetaPS is a simulation-guided framework that enables market agents to adaptively select among programmatic strategies based on market states. It uses simulated markets to generate supervised training data, then selects strategies during inference to produce executable actions. Experiments show MetaPS outperforms fixed strategies and LLM-based agents, with compact models exceeding stronger API models in performance.
P4IR, a two-stage framework, uses supervised fine-tuning and Group Relative Policy Optimization to enhance large language model-based automated code compliance systems. It reduces tree edit and token-level Levenshtein distances by up to 23.8% and 38.6% respectively, outperforming leading LLMs like Claude Opus, GPT-5.2, and GLM-4.7 in zero-shot settings with few-shot prompting, and reduces false positives by a small but statistically significant margin.
Gold Points Sniper (GPS) enables lightweight vision-language models to perform self-guided multimodal reasoning for fine-grained human action understanding. By integrating a Gold Points Extractor, Selective Socratic Questioner, and Semantic Entailment Evaluator, GPS achieves performance comparable to GPT-4o while maintaining superior factual accuracy on CAP benchmark-based instruction-tuning data.
NVIDIA's NeMo AutoModel enables faster fine-tuning of transformer models by automating model selection and optimization. It reduces development time and improves efficiency in training large language models on NVIDIA hardware.
Large language models learn causal structure through a difference-making logic, akin to the experimental method. This approach identifies which word sequences influence outcomes and which do not, using vast text data during training. Architectural features like token embeddings and self-attention support this inductive process by detecting patterns of variation and indifference in language.
DreamUV introduces an end-to-end learning framework that treats UV unwrapping as a generative Flow Matching problem. It learns a mesh-conditioned transport process to generate artist-like UV layouts, with boundary-aware training and Model-in-the-Loop fine-tuning to ensure seam geometry and practical validity. Results show straighter seams, tighter axis-aligned islands, and superior alignment with professional artist preferences.
A fully differentiable emulator of the Atari 2600 VCS is presented, reproducing all 64 ALE games with bit-for-bit accuracy in RAM and screen output. The system enables gradient-based explainable AI by providing a complex, fully known ground truth, with both Julia and JAX implementations validated against a reference emulator and capable of high-throughput differentiable rollouts on GPU.
The paper proposes a self-evolving cognitive framework that uses causal world modeling to enable embodied systems to continuously refine their internal models through interaction. It integrates causal modeling, intervention-driven reasoning, and continual refinement, redefining embodied interaction as an epistemic process for causal discovery and knowledge acquisition. The framework supports a shift from predictive to epistemic intelligence, with a new benchmark for evaluating self-evolving embodied scientific intelligence.
This study compares characters in LLM-generated and human-written stories using narratological dimensions. It finds that while LLMs produce characters with similar basic traits, they lack diversity in complex character features like stylization and wholeness. The research highlights key differences in character depth and variety between human and machine-generated narratives.
PRIME introduces a framework to analyze how large language models handle conflicting instructions by generating calibrated conflicts in response length, format, and reasoning. The study finds that conflict type has a greater impact on model behavior than model size, revealing diverse failure modes across conflict categories. Results highlight the need for conflict awareness and suggest instruction following cannot be reliably assessed through isolated benchmarks alone.
FACTOR introduces an inference-time model that adapts verification criteria based on claim-level uncertainty. It improves factuality and reduces verification cost by dynamically allocating effort to high-risk claims, demonstrating effective and model-agnostic performance on the FactScore benchmark.