LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
arXiv:2507.20881v3 Announce Type: replace Abstract: Endoscopic depth estimation is a critical technology for improving the safety and precision of minimally invasive surgery. It has attracted conside...
The paper introduces X‑SG$^2$S, a feed‑forward framework that embeds 1D to 3D watermarks into 3D Gaussian Splatting (3DGS) scenes without altering the original rendering pipeline. It splits watermark messages into patches, uses a self‑adaptive gate to choose injection points, and an XD injection head to embed multi‑modal messages into sorted 3DGS points. A learnable gate and XD‑extraction heads recover the hidden messages, achieving robust watermarking with minimal interference to scene fidelity.
The paper introduces the Prospective Intention Store (PIS), a method that places lifecycle logic in code and confines language tasks to a typed action space, enabling small models to perform prospective memory tasks more effectively. Using PIS, a small model (DeepSeek-Chat) achieves 82.9% Set‑F1 on PM‑Bench, surpassing the previous best of 65.1%. On Gemma‑E2B, PIS boosts Set‑F1 from 4.2% (without a store) to 66.2%, and reaches 70.1% Set‑F1, outperforming retrospective memory approaches that max out at 54.4%.
The paper proposes the interlingua hypothesis, suggesting that large language models translate by encoding a source sentence into a latent, task‑agnostic feature space and then decoding a target sentence from that space. Three lines of evidence support this: (1) BLEU variance across language pairs is largely explained by language‑specific competences without pair‑specific interactions; (2) many model components influence both monolingual and translation tasks; and (3) fine‑tuning on monolingual data recovers most translation gains seen with aligned documents. These findings converge to support the hypothesis and point toward new ways to understand and improve LLM translation.
The paper introduces a two‑stage training framework that combines Supervised Fine‑Tuning (SFT) and Direct Preference Optimization (DPO) to improve multimodal disaster severity assessment. It creates two datasets—ReasoningSet for validated rationales and PreferenceSet for paired rationales—using a single Human‑in‑the‑Loop workflow. Experiments on InternVL‑3‑8B and LLaVA‑1.5‑7B show that SFT boosts classification accuracy and Macro‑F1, while DPO further enhances interpretability and alignment with human judgment.
The paper investigates why large language models lose safety alignment after benign fine‑tuning. It argues that safety alignment relies on a low‑rank, output‑routing geometry that becomes flatter during fine‑tuning, and that after only 100 benign examples this routing is sharpened in output‑side MLPs, leading to fragile safety while general performance remains relatively intact. Techniques like LoRA and ASAM can delay this collapse by reducing output‑side sharpness, but their effectiveness diminishes with larger fine‑tuning scales.
The paper introduces a new consistency criterion for auditing decision systems that combines ensemble margin with local prediction variability to address predictive multiplicity, or the Rashomon effect. It shows that finite ensembles converge to the expected model’s consistency score as ensemble size and sample count grow, and demonstrates that ensembling models from the Rashomon set reduces unchecked incorrect predictions while keeping diversions moderate. Experiments on transformer and fine‑tuned language models for NLP and tabular classification confirm the method’s effectiveness and stronger alignment with existing multiplicity metrics.
The paper introduces AVA, a framework that tests whether general NLP embeddings can differentiate logic-sensitive relational semantics in ontologies and knowledge graphs. AVA uses 171,007 contrastive triplets from 163 ontologies, each containing an ontology statement, a paraphrase, and a hard negative with contradictory meaning. Evaluation of over 25 embedding models shows significant limitations, with the best model achieving only 0.739 triplet accuracy and 0.135 for hard negatives; fine‑tuning helps but does not transfer well to downstream Semantic Web tasks.
The paper investigates how different training strategies affect the prompt sensitivity of large language models. It reproduces and compares methods such as refined data construction and robustness objectives, finding that while robustness fine‑tuning improves over standard fine‑tuning and in‑context learning, the prompt gap remains large (40–57%). Notably, newer techniques like CoIN and PPCL often underperform a simple data‑construction approach that uses one template per batch, and diagnostics suggest that mixed‑template batches force the optimizer to reconcile conflicting updates rather than learn a prompt‑agnostic representation.
The paper introduces WHALE, a method that alternates between updating a language model’s weights and searching for a better harness (the code that manages context and control flow). By iteratively fine‑tuning the model under the current harness and then optimizing the harness under the updated model, WHALE improves performance across search QA, math reasoning, and chess puzzles, outperforming weight‑only, harness‑only, and Fast‑Slow Training by 4.15–24.38 percentage points in mean@8 accuracy. The approach uses either fixed phase lengths or an adaptive patience rule to decide when to switch phases, and the authors provide code on GitHub.
arXiv:2609.01573v1 Announce Type: cross Abstract: How to divide a fixed annotation budget between supervised fine-tuning (SFT) and reinforcement learning (RL) during LLM post-training remains an open...
The paper demonstrates that an external observer can identify the type of workload running on an NVIDIA H200 GPU by analyzing its power draw, distinguishing training, inference, and non‑AI tasks with high accuracy. Using 930 recorded traces, the authors achieve 97% accuracy and a macro‑averaged F1 score of 0.955 on unseen model families. They also test four evasion strategies to disguise training as inference, showing that a hardened detector can catch most attacks, though one strategy (LoRA) remains partially detectable.
The study investigates whether post‑training methods—GRPO, SFT, and DPO—improve language models’ ability to follow prompt evidence that conflicts with memorized knowledge. By comparing nine training variants across different scales and families, the authors find that grounding gains are modest for GRPO, moderate for Conflict‑SFT, and near‑ceiling for DPO, but all largely rely on the same causal attention‑head set present in the starting checkpoint. Removing the starting‑model grounding direction suppresses these gains, while adding it back recovers a significant portion of DPO’s improvement, indicating that existing model machinery drives most of the observed gains.
The paper introduces CoLT-Drive, a 3,536-sample counterfactual long‑tail benchmark for evaluating decision‑level driving affordance prediction, which tests whether models can infer how rare objects affect an ego vehicle’s high‑level actions. It also proposes KPA, a knowledge‑preserving adaptation framework that combines structured prompting, expert merging, and a regime‑aware LoRA mixture‑of‑experts module to improve small VLMs on driving tasks. Experiments show KPA achieves 60.8% pair accuracy on CoLT‑Drive, outperforming the Qwen3‑VL‑2B baseline and LoRA SFT while keeping competitive in‑domain performance.
The paper presents a lightweight retraining strategy for a parameterized Reduced Order Model (ROM) that achieves full‑model accuracy using only a fraction of the computational effort and sparse observations. The ROM architecture combines a Variational Autoencoder for dimensionality reduction with a transformer network that evolves latent states while accounting for the Reynolds number as an external control variable. By leveraging the probabilistic VAE, the method generates trajectory ensembles and uncertainty estimates, and adapts to out‑of‑sample parameters through sparse data assimilation with an ensemble Kalman filter, focusing retraining on the autoencoder to correct latent manifold distortions.
The study evaluates large language models (LLMs) as rerankers in conversational movie recommendation, comparing proprietary, open-weight, and fine-tuned LLMs against collaborative-filtering and sequential baselines on the ReDial benchmark. Results show that the best proprietary LLM achieves an NDCG@10 of 0.1497 with a shared semantic candidate pool, outperforming non-LLM baselines, while open-weight LLMs do not surpass a tuned shallow autoencoder under the same protocol. The analysis also highlights that reranker performance is highly sensitive to candidate generation, pool size, scoring policy, and decoding temperature, suggesting these factors should be reported as standard evaluation fields.
The paper introduces Iterative MBR Distillation for Error Span Detection (ESD) in machine translation, a self‑evolution framework that replaces human annotations with pseudo‑labels generated by a large language model. By iteratively applying Minimum Bayes Risk decoding, the method produces high‑quality error spans without costly human effort. Experiments on WMT Metrics Shared Task datasets show that models trained solely on these pseudo‑labels outperform both unadapted baselines and supervised models trained on human data at system and span levels, while keeping sentence‑level performance competitive.
The paper introduces DroneCATS-Agent, a modular framework that places a multimodal large language model (MLLM) at the core of a drone’s control loop, allowing the model to decide actions solely from prompts. It presents the DroneCATS benchmark, evaluating MLLMs on four tasks—approaching, tracking, searching, and multi‑drone commanding—without fine‑tuning or function‑calling. Results show that while small open models can navigate reliably, they often fail by mismanaging protocol termination, highlighting a gap between perception and action planning in current MLLMs.
arXiv:2609.01564v1 Announce Type: cross Abstract: Large language models (LLMs) struggle to classify text into taxonomies with many semantically similar labels, as the distinctions are domain-specific...
The paper introduces a scalable framework that uses finetuned large language models (LLMs) to translate unstructured textual resources into structured event logs for process mining. By creating a new text-to-log dataset and finetuning LLMs on it, the authors demonstrate that the resulting models produce high‑fidelity event logs, outperforming few‑shot or zero‑shot prompting methods. This approach enables previously unused organizational data, such as incident tickets and manuals, to be incorporated into process mining workflows.