LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
arXiv:2609.18971v1 Announce Type: new Abstract: Surgical phase recognition maps each video frame to a clinically meaningful workflow phase, supporting context-aware assistance, documentation, and pos...
arXiv:2609.18037v1 Announce Type: new Abstract: Segment Anything Models provide reusable priors, yet they require user prompts and cannot support fully automatic instrument segmentation. Automatic pr...
The paper introduces Code Consistency Preference Optimization Verification (CCPO), a method that generates computationally sound solutions with dependency graphs to improve execution-consistent preference optimization for language models. By building a scientific reasoning dataset and extracting reasoning steps, prerequisites, and derivability relationships, the authors compute execution consistency scores that are used to fine‑tune models such as Llama‑3‑8B and DeepSeekMath‑7B, achieving significant performance gains on MATH (+17.0%) and GSM8K (+15.1%). The extended Scientific Feasibility Control framework further boosts accuracy on PhyX physics reasoning to 50.1%, surpassing existing models while maintaining high scientific validity and reducing law violations.
arXiv:2609.19134v1 Announce Type: new Abstract: Scientific code repositories encode decades of human knowledge in executable models, methods, and tools. Yet fragmented toolchains, implicit domain con...
arXiv:2609.19071v1 Announce Type: new Abstract: Extracting SNP-phenotype associations from biomedical literature is vital but challenging. We benchmarked diverse NLP models, including MLMs, hybrid ar...
arXiv:2609.17848v1 Announce Type: new Abstract: Limited controlled evidence exists on how training data, adaptation method, and model scale jointly affect tool-calling performance in language-model a...
The paper compares Knowledge-Graph Based Augmentation (Graph-RAG) with Retrieval-Augmented Generation (RAG) for answering culturally specific questions. Using the LatamQA dataset, Graph-RAG, built automatically from Wikipedia via KGGen, matches RAG performance and reduces the base LLM’s error by 72% with a standard KG and 78% with a benchmark-aware variant. The approach also transfers zero‑shot to Portuguese, showing multilingual applicability.
Safety tuning pipelines judge only the final answer, which makes it difficult to distinguish robust refusal from two undesirable shortcuts: blanket refusal on benign requests and polished but unfaithf...
World models endow perceptual systems with the ability to predict how scenes evolve under interaction. They are most beneficial when trained on diverse volumes of data, to instill a rich prior into do...
Existing legal benchmarks cover diverse tasks, while continuous monetary remedies remain comparatively underexplored. We introduce ECtHR-NPD, to the best of our knowledge, the first benchmark for pred...
Current Mixture-of-Agents (MoA) paradigms generally treat query routing and agent fine-tuning as separate processes, limiting their ability to respond to evolving agent capabilities. This disconnect p...
While Large Language Models excel in natural language processing, efficiently extending their capabilities to spoken input remains a significant challenge. Existing methods for building SpeechLLMs oft...
This paper studies energy-aware manipulation as a physically grounded learning problem. We define a joint-space mechanical-work proxy from joint torque and angular displacement, and train a differenti...
HoloAegis is a minimally parametric topological inference framework that uses frozen representations to map text onto the unit sphere and makes decisions via Gibbs‑Boltzmann free‑energy differences over pre‑computed anchor centroids. On a frozen three‑benchmark protocol, it matches WildGuard‑7B on toxicity, outperforms it on harmful behaviors, but underperforms on oversafety detection, while ShieldGemma‑2B fails on indirect harms. The study demonstrates that geometric guardrails can substitute for LLM judges in some cases and must defer to them in others, with anchor banks reducing score variance and boundary displacement.
The paper introduces a deep dictionary network (DDN) foundation model designed for ultra‑low‑dose CT (ULDCT) denoising across multiple organs. By cascading convolutional sparse coding layers with iterative soft‑thresholding, the architecture offers inherent interpretability, while dynamic dictionary and threshold modules enhance representation. The model is pre‑trained on over one million normal‑dose CT images and fine‑tuned on multi‑organ ULDCT datasets, achieving state‑of‑the‑art performance that consistently outperforms existing methods.
GrowMTP is a method that trains a draft head entirely within the reinforcement learning (RL) loop, using supervision from the RL verification step and a rollout distribution that is narrower than pretraining. By detaching draft‑head updates from the policy backbone, it enables online training of the draft head from scratch. Experiments on Qwen3‑4B, MiMo‑7B‑SFT, and Qwen3.5‑4B‑Base show rollout speedups ranging from 1.36× to 2.13× and overall end‑to‑end speedups from 1.20× to 1.60×, making it a modular acceleration component for RL frameworks lacking pretrained draft heads.
arXiv:2609.16229v1 Announce Type: new Abstract: Machine unlearning aims to remove specific knowledge from a trained large language model (LLM) without retraining from scratch. Existing methods modify...
The paper introduces MIMIC, a framework that uses executable code to generate rigorous reasoning data for large language models (LLMs). By converting algorithms into verifiable reasoning trajectories through narrative fusion, code-guided test synthesis, and dynamic code instrumentation, MIMIC creates a Code-Instrumented Reward (CIR) that supplies dense, high‑fidelity supervision for reinforcement learning. Models trained with MIMIC’s synthetic dataset show significant, consistent improvements in general reasoning, complex mathematics, and fine‑grained deterministic tasks.
EMODY Flow is a lightweight flow‑matching framework that generates synchronized full‑body motion and facial expressions conditioned on speech and emotion. It attaches to a frozen Qwen‑3 Omni model, reusing its audio codecs to drive two parallel DiT generators for SMPL‑X body pose and FLAME facial expressions. An auxiliary emotion classifier at training time restores emotion sensitivity, enabling EMODY Flow to achieve state‑of‑the‑art gesture quality on BEAT2 and zero‑shot facial animation on TFHP, with significant improvements in FGD, Beat Correlation, and Diversity metrics.
arXiv:2609. 17435v1 Announce Type: new Abstract: We submit M\'eTRON-FR, a 125M GPT-2 pretrained on 92.