LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
Robots operating in real-world environments must execute complex, multi-step bimanual tasks over long horizons rather than single, isolated actions. Current manipulation datasets struggle to support t...
The paper explores Retrospection-Only Fine-Tuning (ROFT), a method where a language-model agent improves its behavior by generating and training on explanations of its own experiences, without external teachers or reward signals. In software‑engineering tasks with Qwen3.5‑4B, ROFT achieves comparable or better solve rates than GRPO while requiring fewer updates and training time, and can learn from failures alone. Behavioral analysis shows ROFT indirectly assigns credit to actions and can produce shorter, more direct solutions when prompted to focus on direct solutions.
DynaTokens introduces lightweight, scene‑specific tokens that enable camera‑controlled video models to learn and generate dynamic scene motion while keeping the base model frozen. By leveraging cross‑attention, the tokens are trained from a few example trajectories, allowing the model to handle both camera‑induced motion and localized object dynamics. Experiments on VBench2 and WorldScore show that DynaTokens outperforms LoRA, block finetuning, and other trainable‑layer baselines in balancing dynamics and camera control.
The paper introduces Geometrically Modified Outputs (GMOs), a technique that reweights the singular values of a generative model’s input-output Jacobian to strengthen the negative signal used in self‑training. By amplifying mode‑seeking behavior and distortions in standard outputs, GMOs provide a more targeted negative guidance for models such as Neon and SIMS. Experiments on one‑step generative models show that GMOs consistently improve the performance of negative‑guidance self‑training methods compared to using unmodified outputs.
BiMoGen introduces a unified masked discrete diffusion framework for bidirectional motion‑text generation, addressing the limitations of autoregressive models in capturing bidirectional dependencies between language and motion. The approach employs a two‑stage training strategy—decoupled uni‑ and cross‑modal pretraining followed by supervised fine‑tuning—to establish robust cross‑modal correspondence, and incorporates Generation‑Aware Self‑Correction to mitigate error propagation during inference. Experiments on HumanML3D and KIT‑ML show competitive performance on both text‑to‑motion and motion‑to‑text tasks, demonstrating the effectiveness of the proposed training and correction mechanisms.
The paper introduces P2P‑T, a data‑efficient, object‑centric framework that learns tool manipulation directly from human video demonstrations. It uses a two‑stage approach: first pretraining an object‑centric world model to extract stable pose priors, then integrating these priors into a pose‑aware low‑level policy. By automating data processing with foundation models, P2P‑T eliminates the need for human‑robot aligned data and achieves a 73% improvement over prior state‑of‑the‑art performance on complex real‑world tool manipulation tasks.
The paper presents an expert-driven method for turning normative principles—specifically Islamic ethical, theological, and jurisprudential traditions—into alignment data for language models. Over a year, seven experts curated 2.8K supervised fine-tuning examples and 5.4K preference pairs in Arabic-English, then evaluated models trained on these datasets. Experiments show that models trained with the curated SFT data outperform a baseline in expert judgments, while adding preference data yields a smaller, non-significant improvement.
The paper proposes a two‑stage framework to improve images from low‑cost handheld ultrasound devices by mapping them to high‑quality hospital images. First, a cycle‑consistent style‑transfer model generates pixel‑aligned low‑quality/high‑quality pairs from unaligned real scans. Second, a Dual Degradation‑Guided Low‑Rank Adaptation (DDG‑LoRA) model fine‑tunes an LQ‑to‑HQ enhancement network, achieving a 16.7% FID improvement on the USenhance2023 dataset.
The study investigates how well a single consumer earbud IMU can estimate 3D body pose and whether adding foot IMUs improves accuracy. Using a multimodal capture pipeline with RGB‑D video, an AirPods head IMU, and Striv insole IMUs, the authors benchmark pose estimation across various motions and train recurrent models (IMUPoser and MobilePoser). Results show that a head IMU alone achieves 79.0 mm rigid‑MPJPE and 0.809 macro‑F1 for foot contact, while adding foot IMUs does not significantly improve pose and can even degrade performance due to insole orientation errors.
The paper investigates post‑training quantization of transformer attention blocks by optimizing a joint loss over the Q, K, V projections rather than individual weight matrices. Using this joint attention‑based objective (JAB), the authors achieve significant compression on Mistral‑7B, recovering 77‑90% of the performance gap at 3 bits, but the method fails when MLP layers are included. A role‑aware offset rule that ignores sensitivity estimates outperforms JAB on GPT‑2 and full Mistral‑7B, demonstrating that the matrix a weight belongs to is more critical than sensitivity metrics.
The paper introduces ATLAS, a method that uses retained-domain activation atlases to guide task-specific fine‑tuning of language models. By providing local reference centers and directional filters, ATLAS learns a low‑rank residual that adapts to new tasks while preserving existing behavior. Experiments on Qwen3‑8B and other backbones show lower retained‑output KL divergence, fewer rewritten answers, and more stable responses compared to seven baselines.
The paper proposes Metric-based Loss Weighting to enhance visual grounding in multimodal machine translation. By increasing loss for tokens that benefit from image context—identified via the Point-wise Cross-mutual Information (PCXMI) metric and its Congruency-based variant—the method improves translation accuracy on the CoMMuTE dataset by over 7 percentage points. Experiments fine-tune three pretrained multimodal LLMs across three language directions, showing superior performance compared to standard fine-tuning while preserving overall translation quality.
JevSoup is a training‑free framework that separates System One expert routing from System Two execution for low‑rank adaptation (LoRA) models. It selects two experts based solely on input and expert descriptions, retains the leading expert’s update, projects the second onto the orthogonal complement of the first update’s row space, and combines them with equal weights. On 14 PorTAL tasks and three Qwen3 scales, JevSoup improves task‑macro accuracy by up to 1.19 % and sample‑micro accuracy by up to 1.21 % over the strongest evaluated baselines.
arXiv:2609.31193v1 Announce Type: new Abstract: Current Audio-Visual LLMs (AVLLMs) struggle with reasoning over videos featuring multi-speaker dialogues. In such videos, resolving "who says what" is...
Auteur is a language‑driven method that generates human‑centric camera framing for generative video models. It treats shots as framings relative to an actor, encoding shot size, angle, and composition as functions of human pose and motion, and uses a domain‑specific language that converts to standard 6‑DoF camera parameters. A fine‑tuned multimodal large language model acts as a virtual director, mapping natural language descriptions and coarse human motion to sparse DSL keyframes that are interpolated into continuous camera trajectories for video generation.
arXiv:2609.31356v1 Announce Type: new Abstract: Vision-Language Models (VLMs) exhibit remarkable zero-shot generalization, yet they often encode unwanted or hazardous stylistic domains such as ideali...
FeatMark is a watermarking framework that protects images from text‑to‑image diffusion model mimicry attacks by embedding small, scene‑consistent micro‑features instead of pixel‑level perturbations. It constructs domain‑specific feature banks, selects executable features, and injects them via mask‑guided concept editing to create highly localized, natural edits. Experiments on VGGFace2, CelebA‑HQ, and WikiArt show FeatMark remains robust against ten strong watermark removal attacks and several adaptive attacks, with minimal impact on perceptual quality and extending to video mimicry scenarios.
MACBT is a clinician‑facing AI decision‑support system that models the five‑stage CBT workflow with five collaborative agents and incorporates a CBT‑specific longitudinal memory module (CD Memory) to track cognitive distortions across sessions. The system generates pre‑session pathology reports and prioritises interventions, and was trained on a Chinese CBT dialogue corpus using a Qwen3‑14B backbone with supervised fine‑tuning and preference optimisation. Evaluation by GPT‑4 judges shows MACBT outperforms several existing chatbots in professionalism and clinical authenticity, and the memory‑augmented version improves session quality by 12.6% and achieves a longitudinal mean of 2.29 on continuity, progression, and personalization.
The paper investigates why retrieval‑based open‑ended evaluation fails in medical fact verification. By creating two detailed taxonomies—one for retrieval‑stage errors across five quality dimensions and another for verifier‑reasoning errors across six steps—the authors automatically label evidence quality and reasoning errors using an LLM‑as‑Judge pipeline. Their large‑scale stress tests across multiple retrieval methods and verifier models show that increasing model size, reasoning effort, source breadth, or medical fine‑tuning does not eliminate these failure modes, indicating fundamental limits of the retrieve‑then‑verify paradigm in open‑ended medical contexts.
arXiv:2609.31461v1 Announce Type: new Abstract: Background: Large volumes of unlabeled knee MRI scans are available across repositories but remain insufficiently leveraged. We developed KneePreM, a k...