LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
Reasoning models often continue generating after their answers have settled. Settle learns when to stop from answer stability in completed traces. It trains the existing end-of-reasoning token while k...
Demographic fairness gaps in automatic speech recognition are almost always reported from a single training run. We fine-tune the Q-former projector and LoRA adapters of a speech LLM at five audio com...
arXiv:2609.33772v2 Announce Type: replace Abstract: Executable environments are critical for post-training agents on tasks that require tool use and multi-step interaction, but constructing executabl...
Beacon is a new agentic visual reasoning model that improves multimodal large language models (MLLMs) by better deciding when to use tools and how to use them. It introduces two key concepts—Mode Adaptiveness, which ensures tools are invoked only when necessary, and Tool Effect, which measures the net benefit of tool use— and trains the model with supervised fine‑tuning and reinforcement learning that rewards necessity-aware decisions and expands capability through expert hints. Across 13 benchmarks, Beacon outperforms other open‑source models, achieving the highest average score and the largest net tool‑gain on diagnostic tests.
The paper investigates whether input ablations on predictive models can reliably reveal the importance of information for explaining human sequential choice behavior. Using two synthetic bandit tasks with known generating policies, the authors compare GRUs, Transformers, a fine‑tuned LLaMA, and cognitive models under varied reward contributions. They find that while neural models can predict choices well, their responses to ablations often diverge from the true generating process, indicating that predictive accuracy alone does not guarantee faithful model ablations.
The study trains ten open‑weight large language models (LLMs) to predict their own accuracy on factual multiple‑choice questions before answering. Results show that the models’ confidence signals split into two distinct patterns: early in training, confidence aligns with output consistency (how concentrated the answer distribution is), while later, it aligns with true accuracy but only on data similar to the training set. This indicates that calibration training may not universally teach LLMs to detect their own errors.
Back2Struct is a system that converts structured images—such as diagrams, charts, and flowcharts—into editable vector graphics code (SVG/XML). By predicting semantically rich, object-level SVG code rather than low-level pixel vectorization, it allows the generated graphics to be imported into tools like PowerPoint for easy editing, restyling, and reuse. The model is trained with supervised fine‑tuning and reward‑based learning that enforces syntactic validity, concise length, and visual fidelity to the input, leading to higher accuracy, editability, and user alignment compared to baselines.
arXiv:2609.37323v1 Announce Type: new Abstract: Discrete diffusion models have emerged as a powerful paradigm for solving combinatorial optimization (CO) problems on graphs by learning to sample high...
arXiv:2609.36590v1 Announce Type: cross Abstract: Self-speculative decoding accelerates large language model (LLM) inference by drafting tokens from the target model itself, but faces a sharp tradeof...
The study investigates whether language models tailored to specific cognitive domains better align with corresponding brain systems. By prompting and fine‑tuning large language models into six domain experts—sensory, spatial, numerical, reasoning, social, and abstract—the authors find that each expert’s representations more closely match the brain region associated with its domain than other experts. This domain‑specific alignment holds across multiple base models and fMRI datasets, while overall prediction accuracy remains largely unchanged, indicating that regional alignment can be obscured when summarizing across the brain.
The paper introduces Alignment Forecasting, a method for predicting whether fine‑tuning a language model on a given dataset will increase specific alignment failures such as deception or sycophancy. It presents ALIGNMENTFORECASTBENCH, a benchmark of over 5,000 forecasting questions across many models, datasets, and failure modes, and shows that a simple forecasting scaffold using an LLM’s assessment of dataset bias can outperform baseline forecasters. The authors demonstrate that filtering out high‑risk training examples identified by the forecaster can improve alignment in multiple‑choice evaluations, though benefits in open‑ended conversations remain uncertain.
arXiv:2609.35793v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly trained with reinforcement learning from verifiable rewards (RLVR). An exact verifier can also support te...
arXiv:2609.35868v1 Announce Type: new Abstract: Is human readability necessary for effective fine-tuning of large language models? We investigate whether model-conditioned training representations ca...
THEIA is a multimodal dataset that pairs thousands of analog circuit layout images with question‑answer conversations, and it introduces a benchmark using a fine‑tuned vision‑language model to analyze GDSII files. The dataset and benchmark enable designers to interact with and query physical layouts as intuitive, meaningful entities. Experiments on five realistic tasks show the fine‑tuned model outperforms general‑purpose vision‑language models by up to 73%, revealing a significant gap between general multimodal reasoning and domain‑specific layout understanding.
TT-VidT is a video pretraining method that decouples the temporal axis by combining a per‑frame ViT-B/16 spatial encoder with a compact Temporal Transfer Layer trained via Diff Compression. The authors conduct a systematic 24‑configuration study to isolate architecture, objective, and decoder effects, showing that the full TT-VidT design yields the strongest motion‑sensitive representations. In downstream fine‑tuning, TT‑VidT outperforms state‑of‑the‑art baselines on Jester, Something‑Something V2, ARID, and Diving48 while using significantly fewer encoder FLOPs.
The paper introduces TriageRA-CCF, a method for adaptively allocating low‑rank LoRA channels in medical large language models based on source‑side signals: answer confidence, clinical coverage, and a counterfactual close‑miss proxy. By supervising a budget router that selects among 2, 4, or 8 active ranks, the approach improves average accuracy over existing LoRA variants on Qwen3‑8B and Llama3.1‑8B, though gains vary across benchmarks. Ablation studies confirm that each signal contributes to better budget decisions, though their combined effect is not uniformly superior across all backbones.
The paper introduces a dataset watermarking technique that embeds a watermark by increasing the co‑occurrence of randomly selected word pairs through meaning‑preserving local edits. The watermark can be detected solely from generated text with provable false‑positive control, and experiments on four base models and three datasets show reliable detection (p < 0.01) even when the watermarked data constitutes less than 5% of fine‑tuning tokens. Compared to existing methods, the approach better preserves benchmark utility and semantic integrity.
The paper shows that pretrained transformers often stop using their depth early, following only a few lines of context. A small rank‑8 LoRA applied to an early layer can extend this chain‑following ability, enabling models like Qwen3‑8B to achieve near‑perfect accuracy on 24‑line chains and significantly longer chains with further training. The LoRA acts as a relay, passing chain identity through middle layers and allowing frozen heads to read further up the chain, with the last useful intervention layer identified in most held‑out models.
The paper introduces Population Fidelity, an evaluation framework for assessing how well large language models (LLMs) represent human population attitudes. It focuses on three dimensions: group-level accuracy, between-group variation, and the structure of that variation. Using the framework, the authors replicate a prior study on machine bias and test cultural fine-tuning, finding that while fine-tuning improves overall alignment, it does not enhance representation of within-population differences.
arXiv:2609.37858v1 Announce Type: new Abstract: Many localized large language model (LLM) unlearning methods select a small parameter subset from a localization signal and keep it fixed during optimi...