LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
arXiv:2610.40934v1 Announce Type: new Abstract: In this work, we provide a formal framework and mathematical definition for the subset of foundation models focused on enabling the property of "any-ta...
arXiv:2610.40798v1 Announce Type: new Abstract: Web world models increasingly predict the consequences of candidate actions before agents act, yet a plausible predicted future does not necessarily su...
arXiv:2610.40744v1 Announce Type: cross Abstract: With the progressive capabilities of large language models (LLMs), researchers are finding ways to integrate them into the control loops of Open Radi...
arXiv:2610. 41114v1 Announce Type: new Abstract: Optimizing a proxy objective can improve human-perceived quality, but the resulting gain can vary with optimization strength.
The paper investigates how memorization behaves during fine‑tuning of large language models with LoRA, a parameter‑efficient technique. It finds that LoRA fine‑tuning diverges from previous results seen in full fine‑tuning and pre‑training: factors like model size and data duplication no longer predict memorization in the same way. Using a relaxed similarity‑based metric, the authors show that LoRA substantially lowers memorization risk while preserving strong task performance.
arXiv:2610.41158v1 Announce Type: new Abstract: Genetic programming (GP) has been widely used to learn scheduling rules for dynamic flexible job shop scheduling (DFJSS). However, trained rules tend t...
arXiv:2610.41710v1 Announce Type: new Abstract: Clinical trial failures are a major barrier to efficient drug development, incurring substantial costs and delaying patient access to effective therapi...
arXiv:2610.41665v1 Announce Type: new Abstract: Sparse tuning is widely used to adapt large language models due to its parameter efficiency. However, its effectiveness depends not only on the number...
The paper presents a post‑training method for dynamically routed mixture‑of‑experts (MoE) models that integrates independently optimized specialists efficiently. It introduces calibrated multi‑teacher on‑policy distillation (C‑MOPD) with teacher‑wise loss normalization and clipping, combined with a post‑capacity bounded‑width dispatch, final‑route replay, and asynchronous rollout–training overlap. On the Turing‑20B‑A2B model, the approach achieves a high integration score (0.9988), retains most specialist gains, outperforms shared SFT and Qwen3‑8B on reasoning‑enabled benchmarks, and reduces consolidation time by 2.86×.
The paper examines how post‑training changes in large language models affect fact‑checking judgments based on a poster’s social class. Using randomized poster profiles, the study finds that only the open T"ulu‑3 recipe introduces a statistically reliable bias—making the model more likely to label claims from lower‑status jobs as false—while other commercial models do not exhibit this effect. The bias is traced to a single residual‑stream direction that emerges during supervised fine‑tuning, and the authors confirm the mechanism on additional claims.
Artifact Distillation proposes a training‑free method for transferring knowledge from large frontier language models to smaller ones by having the teacher solve tasks and store its problem‑solving steps in an artifact bank that the student can query during inference. In experiments on LiveBench, a bank created by Claude Opus 4.5 enabled Claude Haiku 4.5 to surpass Claude Sonnet 4.5 on spatial reasoning and nearly match it on table joining, all at significantly lower cost. The same artifact bank also boosted Qwen3.5‑27B’s pooled score from 0.730 to 0.800, demonstrating that the approach works across different model families.
The paper introduces a new scorer for generative answer engines that ranks candidate pages by the likelihood of being cited. It evaluates pretrained rerankers, a hosted typed decision model (Jev), eight open decision models, and a distilled 4B student on GEO‑Bench queries answered by ten engines. The distilled student, fine‑tuned with low‑rank adapters, achieves comparable performance to larger models and the authors release the model, benchmark, predictions, and code.
The paper introduces Persona Learning from Feedback (PLF), a framework that creates reusable, preference‑grounded assistant personas as LoRA adapters on a frozen backbone. During training, a frozen self‑teacher provides token‑level supervision conditioned on persona specifications and response‑specific feedback, while a feedback gate focuses updates on responses needing improvement. At deployment, a supervised router selects an appropriate adapter or composition based on the request or dialogue, enabling personalized AI assistants without user‑specific fine‑tuning and improving preference fit over existing persona prompting and distillation methods.
UMI‑Live is an open‑source, full‑stack bimanual manipulation system that lets human experts demonstrate tasks for robot learning without needing teleoperation of physical robots. The custom hardware integrates camera, LiDAR‑inertial sensing, onboard compute, and power, eliminating external tracking or calibration, while a user interface records data and an asynchronous pipeline performs SLAM reconstruction to produce training‑ready datasets. The system supports dual‑YAM teleoperation and VLA policy training, and the authors released 40 hours of data across three tasks, demonstrating high success rates on a pick‑and‑place benchmark.
Intrinsic Distribution Matching (IDM) extends distribution matching distillation (DMD) for one‑step and few‑step text‑to‑image generation by adding an explicit moment constraint that matches feature means and per‑coordinate standard deviations using the generator’s exponential‑moving‑average copy. This approach avoids a separate fake‑score network and an external feature encoder, while retaining the full DMD correction. Experiments on Z‑Image‑Turbo 6B and Qwen‑Image 20B with full‑parameter and LoRA tuning show IDM achieving the highest scores on Chinese and English LongText‑Bench, and ablations confirm the benefit of the moment loss.
The paper introduces MedCBR, a two‑stage alignment framework for medical vision‑language models that improves diagnostic reasoning at the capability boundary. First, it uses Difficulty Weighted Boundary Preference Optimization (DBPO) to focus learning on counterfactual boundary cases, and second, it applies Entropy Gated Grounded Policy Optimization (EGPO) to reward anatomical grounding and reasoning fidelity while mitigating rubric noise. Experiments on six medical benchmarks demonstrate consistent accuracy gains, especially on challenging differential diagnosis tasks, with the combined stages delivering super‑additive improvements.
The paper introduces TOME-Math, a large-scale corpus that transforms 2,118 university-level mathematics textbooks into about 380,000 structured proof-supervision records, preserving chapter order, proposition-proof links, dependency references, and source provenance. Using these structural signals, the authors explore curriculum learning across multiple subjects and analyze scaling patterns with supervision volume and source breadth. Fine‑tuning on TOME‑Math improves the unweighted mean score on held‑out evaluations and public benchmarks by 16.7 percentage points over the base model, surpassing all three public instruction‑data baselines in aggregate.
The paper introduces JoBS, a Bayesian optimization framework that co‑optimizes data and model configurations for large language model training without requiring full training runs at each iteration. JoBS first allocates part of the budget to train a scaling‑law‑inspired performance predictor using only a few training steps, then uses this predictor to guide subsequent BO iterations, thereby exploring many more configurations within the same budget. The authors analyze JoBS’s regret, derive optimal budget splits, and demonstrate that it outperforms both independent optimization and existing multi‑fidelity BO baselines across diverse downstream tasks.
The paper introduces Anagent, a multi‑agent framework designed to improve scientific table and figure analysis. It comprises four specialized agents—Planner, Expert, Solver, and Critic—that collaboratively decompose tasks, retrieve information, synthesize analyses, and refine outputs. The authors also present AnaBench, a large benchmark of 63,178 instances across nine domains, and demonstrate that Anagent yields significant performance gains, up to 13.43% without fine‑tuning and 42.12% with fine‑tuning, highlighting the importance of task‑oriented reasoning and context‑aware problem solving.
The paper introduces two new streaming deep reinforcement learning algorithms—Streaming Soft Actor-Critic (S2AC) and Streaming Deterministic Actor-Critic (SDAC)—designed to work with state‑of‑the‑art batch RL methods and suitable for on‑device fine‑tuning, such as Sim2Real transfer. These algorithms achieve performance comparable to existing streaming baselines on standard benchmarks without the need for extensive per‑environment hyperparameter tuning. The authors also examine the challenges of transitioning from batch to streaming RL, demonstrating that a naive approach can degrade pre‑trained policy performance and propose a principled solution to mitigate this issue.