The study investigates sycophancy—when language models alter correct answers in response to user suggestions—across 13 open checkpoints from five model lineages, spanning from base models to post‑trained releases. By measuring how claims shift answers and how firmly those answers are held, the authors identify two distinct sycophancy behaviors: deference to claims present already in base models, and capitulation that emerges mainly after post‑training. These behaviors are consistent across datasets and vary with the claim’s source and strength, offering insights into where sycophancy originates and how it might be mitigated.
By Jiarui Yao, Shweta Parihar, Nghia Tuan Bui, Guergana K Savova, Lijing Wang
The paper investigates how small entry‑wise errors in machine‑learned Kohn‑Sham Hamiltonians can still lead to inaccurate eigenvalues and eigenvectors. It shows that existing physics‑informed losses only partially capture the error, and proposes using the cumulative local density of states (cumulative LDOS) as a complete training signal that uniquely determines the Hamiltonian. Fine‑tuning three Hamiltonian networks with this loss, called FermiScope, reduces HOMO‑LUMO gap errors without increasing element‑wise error, and the improvement generalizes to out‑of‑distribution and molecular‑dynamics scenarios.
By Fang Sun, Hekai Cheng, Ruoyan Li, Yuanzhou Chen, Wei Wang, Jason Cong, Yizhou Sun
Merge and Branch (M&B) is a tree‑structured supervised fine‑tuning method for large language models (LLMs) in recommendation systems. It compresses each historical item’s semantic identifier (SID) and text into a few learned tokens (Merge) and then uses the compressed history to predict multiple target SIDs (Branch), organizing each example as a shallow, wide tree. The approach includes a tree FlashAttention mechanism for efficient training and a tree‑structured key/value cache for constrained decoding, achieving better recommendation performance and improved training and decoding efficiency compared to strong baselines.
By Yunqin Zhu, Yao Xie
The paper investigates how language models (LMs) can introspect their internal signals and identifies a problem called causal bypass, where self-reports depend on correlated but causally unrelated signals. By training hundreds of introspectors on low‑level signals across multiple LMs, the authors show that standard supervised fine‑tuning (SFT) often leads to causal bypass. They propose interventional training, which trains models to report on counterfactually perturbed values, and demonstrate that this approach yields more accurate introspection for both endogenous and injected signals, even on out‑of‑distribution data, and that the method generalizes across token positions and scales with more data.
By Zilu Tang, Jongin Kim, Kerem Sahin, David Atkinson, Aaron Mueller, Derry Wijaya
PairCLIP‑SWA is a single‑backbone deepfake detector that fine‑tunes an OpenCLIP ViT‑L/14 visual tower using a class‑balanced corpus created by CLIP‑similarity‑guided subsampling and then applies stochastic weight averaging over the fine‑tuning trajectory. Trained on the IEEE SP Cup 2025 (DF‑Wild‑Cup) benchmark, the weight‑averaged model achieves 97.04 % accuracy, 98.64 % AU‑ROC and a 4.07 % equal error rate—outperforming the best published three‑backbone ensemble. The study shows that averaging stabilizes validation performance, prevents misranking on disjoint test sets, and that CLIP pre‑training, rather than ensembling, is the key driver of accuracy.
By Md. Nakib, Sudipto Sarkar, A. J. A. Jubayer Talukder, Mohammad Ariful Haque
LLM-PowerTrace is an open dataset that provides time‑resolved, per‑GPU power traces for large language model fine‑tuning and inference. The traces are sampled every 100 ms across various NVIDIA platforms and include detailed workload and system metadata. The dataset covers a wide range of model families, sizes, and controlled sweeps over batch size, sequence length, precision, parallelism, checkpointing, concurrency, GPU frequency, KV‑cache and weight quantization, prompt and output lengths, and request‑arrival patterns, and it defines power‑grid‑facing metrics and an AI‑compute power & energy scaling law.
By Jiansheng Wang, Jaskeerat Singh, Haochen Xu, Xinyi Li, Mingyuan Yan, Shaohui Liu, Yuzhuo Li, Yize Chen, Wenqi Cui
DeCoRL is a reinforcement‑learning framework designed to enhance multi‑image composition (MICo) by decomposing overall consistency into object‑level consistency. It first establishes instruction‑relevant object correspondences across reference images and localizes these regions in the generated image. A VLM‑based reward model, trained on single‑object consistency annotations, provides fine‑grained, object‑wise feedback that is aggregated into a multi‑object consistency reward to guide RL post‑training. The authors also introduce DeCo‑Bench, a benchmark covering compositions from 2 to 6 reference images, and show consistent improvements on both DeCo‑Bench and the public MICo‑Bench.
By Zhiqiang Wu, Shuang Sun, Jiale Zhang, Jing Li, Yibo Chen, Xu Tang, Yao Hu
The paper proposes using Low‑Rank Adaptation (LoRA) as a structural regularizer for critic learning in off‑policy reinforcement learning. By freezing randomly initialized base matrices and optimizing only low‑rank adapters, the method constrains weight changes to low‑rank residuals, reducing critic loss and improving policy performance. Experiments on state‑based continuous‑control tasks with SAC, FastTD3, and their SimbaV2 variants show that LoRA achieves the best or competitive results on most tasks.
By Yuan Zhuang, Yuexin Bian, Sihong He, Jie Feng, Qing Su, Songyang Han, Jonathan Petit, Shihao Ji, Yuanyuan Shi, Fei Miao
The paper introduces Sparse Adversarial Fine‑Tuning (SAFT), a method that identifies a small set of Robustness‑Critical Parameters (RCPs) through adversarial gradient saliency and updates only these parameters during adversarial training. By focusing on this sparse subspace, SAFT achieves higher clean and robust accuracy than LoRA‑based parameter‑efficient methods and matches or surpasses full‑parameter adversarial training while training only about 5% of the parameters. The approach is validated across multiple architectures and datasets, demonstrating its effectiveness in reducing memory cost without sacrificing robustness.
By Zhaoxin Wang, Zihang Ding, Handing Wang
MEV is a multi‑event video dataset created to support long‑take video generation. It contains high‑quality real‑world videos where subjects perform distinct, non‑overlapping events with precise temporal boundaries and fine‑grained procedural annotations. The dataset includes a comprehensive annotation pipeline that segments videos into temporally grounded events and generates structured captions at both global and event levels, enabling event‑centric evaluation and improved long‑take event transitions in generation experiments.
By Peiyuan Zhu, Shaoan Xie, Yifan Shen, Wen Tian, Zijian Li, Guangyi Chen, Kun Zhang
The paper presents a cross‑lingual abstractive summarization system for Marathi, leveraging the Marathi MahaSum corpus and the smaller PMIndiaSum dataset to fine‑tune the IndicTrans2 neural machine translation model. Evaluation on Marathi‑to‑English summarization using ROUGE, chrF, and chrF++ metrics shows that this approach outperforms prior IndicBART and mBART‑50 baselines. The authors also discuss extending the method to English‑to‑Marathi summarization.
By Asang Dani, Shailesh R Sathe
The paper introduces MCTS‑SFT, a verifier‑guided framework that uses Monte Carlo Tree Search to curate structured reasoning traces for fine‑tuning Verilog generation models. Within this framework, Plan‑MCTS searches over compact hardware‑design plans that explicitly encode interfaces, timing, state, behavior, and update rules, and retains these plans during supervised fine‑tuning. Experiments show that Plan‑MCTS yields the highest functional trace accuracy (84.2%) and improves downstream model performance, raising pass@10 from 52.56% to 57.7% and outperforming chain‑of‑thought distillation by up to 3.7 points while halving generation length.
By Matthew DeLorenzo, Kevin Tieu, Prithwish Jana, Piyush Jha, Dileep Kalathil, Vijay Ganesh, Jeyavijayan Rajendran
The paper investigates context‑aware selective transfer learning (CASL) for forecasting groundwater level anomalies across 457 wells in the Netherlands. While source‑domain pretraining can improve average accuracy, it often degrades performance on many wells; CASL mitigates this by blending frozen no‑transfer and fixed‑transfer models using calibration‑trained weights. The study finds that CASL slightly lowers RMSE compared to no transfer and partially or fully rescues a large proportion of wells harmed by fixed transfer, but no single method dominates across all seeds, highlighting an accuracy‑risk trade‑off.
By Yaxuan Wu, Ali Mohammed Mansoor Alsahag, Seyed Sahand Mohammadi Ziabari
DeFT‑RAG is a retrieval‑augmented generation framework that tackles multi‑hop questions by incrementally decomposing them into self‑contained sub‑questions. Each sub‑question explicitly names all entities it queries, eliminating the need for planning, dependency tracking, or rewriting, and allowing the LLM to focus on a single fine‑grained task. The system offloads bookkeeping to a rule‑based QA table and a lightweight decision model, caches answered QA pairs for reuse, and achieves state‑of‑the‑art accuracy on MuSiQue, 2WikiMultiHopQA, and HotpotQA while reducing LLM input tokens and calls.
By Dingyi Kang, Bingzhe Li
The paper evaluates inference and data‑collection strategies for Vision‑Language‑Action (VLA) models on the BEHAVIOR‑1K benchmark. Using a fine‑tuned π_{0.5} model that achieved a Q‑score of 0.39 on the public test set, the authors systematically compare previously proposed strategies across all 100 tasks. The results reveal that these strategies do not transfer uniformly to new environments and tasks, highlighting the need for more robust approaches.
By Ruben Hammele, Samira Huber, Mohammad Nazeri, Amirreza Payandeh, Alexandyr T. Card, Aaron Nguyen, Anuj Pokhrel, Daeun Song, Xuesu Xiao, S\"oren Pirk
The paper introduces ExLabel, an LLM‑based, self‑evolving assistant that assesses task clarity and test coverage in SWE‑bench‑style software engineering datasets. ExLabel builds an experience bank from cross‑instance knowledge, enabling context‑aware judgments that outperform static rubrics. A new benchmark, SWE‑TQ, spanning eight programming languages, demonstrates ExLabel’s superior F1 scores and its ability to filter data that improves a coding model’s success rate on SWE‑bench Verified.
By Zhonghao Jiang, Xuyang Zheng, Yanfu Yan, Zhongxin Liu, Lingfeng Bao, Shuiguang Deng
CovFedAvg is a federated learning method that preserves local low‑loss curves and selects a shared model without requiring centralized calibration data or client‑to‑client sample exchange. It aggregates gradient contributions to fit a common endpoint and guarantees that the chosen model’s evaluated training loss is no higher than the previous model or the current‑round FedAvg candidate. Experiments on synthetic, activity‑recognition, and gas‑sensor data show modest accuracy gains over standard FedAvg, with a 0.96% improvement in macro balanced accuracy for a width‑16 network.
By Federico Lucchetti, Marcus V\"olp
The paper introduces Domain‑Specialized Upcycling (DS‑Upcycling), a method that expands a dense Transformer with domain‑specific experts and a router to improve multi‑domain adaptation. Unlike prior LoRA‑based adapters that either weight all domains equally or activate all experts, DS‑Upcycling trains a domain‑aware router to select the appropriate expert while freezing other parameters. Experiments demonstrate superior average performance across domains and high routing accuracy (90‑98%) with only 256 examples per domain.
By Mingzi Cao, Xi Wang, Nikolaos Aletras
The paper introduces Prompts to Proxies (P2P), a two‑stage inference‑time system that uses structured attribute‑based prompting and entropy‑guided sampling to build a diverse proxy pool, followed by L1‑regularized regression to create a compact weighted ensemble that matches aggregate survey responses. P2P achieves low mean squared error on American Trends Panel data and competitive performance on World Values Survey cross‑locale experiments, outperforming baseline prompting and demographic‑conditioning methods while using minimal training data. The authors argue that this preference reconstruction approach offers a lightweight, externally verifiable alternative for aligning large language model proxies with target human populations.
By Bingchen Wang, Zi-Yu Khoo, Jingtan Wang
MuQ-Eval is an open‑source per‑sample quality metric for AI‑generated music that uses frozen MuQ‑310M features and a lightweight prediction head trained on the MusicEval dataset. The simplest model—attention‑pooled frozen features plus a two‑layer MLP—achieves high correlation with human judgments (system‑level SRCC = 0.957, utterance‑level SRCC = 0.838). The metric is real‑time on a single consumer GPU, outperforms existing open per‑sample metrics, and supports personalized evaluators via LoRA adaptation on as few as 150 clips.
By Di Zhu, Zihan Chen, Zixuan Li