The paper introduces a calibrated instrument for rigorously measuring how inference optimizations—such as quantization, early‑exit, and speculative decoding—affect the output quality of large language models. It uses a formally calibrated LLM judge that verifies no systematic bias between statistically equivalent outputs and includes a null condition to ensure measured differences are zero. Applying this method, the authors find that a 4‑bit model is indistinguishable from its 16‑bit counterpart, while 3‑bit quantization and early‑exit techniques incur measurable quality losses that vary by language and task, and that token‑certainty‑based acceptance rules cannot reliably identify impactful errors.
By Jerry Kaplan
The paper introduces FairInference, a system that guarantees token-level latency isolation for well-behaved clients in multi-tenant LLM serving. It provides a δ-token fairness guarantee, ensuring that a token generated in isolation within time d will be produced within d + δ in a shared environment. The approach enforces per-token deadlines, bounds GPU compute sharing delays, and accounts for shared KV cache overhead, leading to reduced latency spikes and higher overall throughput compared to existing LLM serving systems.
By Dev Bali, Soujanya Ponnapalli, Yichuan Wang, Natacha Crooks, Scott Shenker, Matei Zaharia
The paper evaluates the Tabular Prior-data Fitted Network (TabPFN) as a surrogate model in surrogate‑assisted evolutionary algorithms (SAEAs) for expensive optimization problems. Through extensive experiments in both offline and online settings across a range of problem types—including single‑objective, multi‑objective, constrained, combinatorial, mixed‑variable, and engineering tasks—the study finds that TabPFN’s effectiveness varies strongly with the problem characteristics. The authors conclude that TabPFN should be used selectively, with customized model management and algorithm design tailored to data availability, landscape complexity, and search‑space properties.
By Lu Han, Jin Wang, Yuchen Li, Haoran Gu, Shulei Liu, Ziyang Shi, Wenao Lu, Handing Wang
EfficientTDMPC is a sample‑efficient, model‑based reinforcement learning method for continuous control that builds on the TD‑MPC family. It reduces estimation error by using an ensemble of dynamics models and averaging return estimates across models and rollout depths, and it can penalize uncertainty in the planner objective. Practical improvements such as fresher buffer data and reduced compute enable the algorithm to benefit from a higher update‑to‑data ratio, achieving state‑of‑the‑art sample efficiency on HumanoidBench‑Hard and DMC hard while matching state‑of‑the‑art on DMC easy.
By Thomas Evers, Cristian Meo, Wendelin Bohmer, Justin Dauwels, Yaniv Oren
RecMorph introduces a topology‑guided spatial recurrent architecture for generalized morphology control, converting a kinematic tree into a sequence that enables joint cross‑limb communication and representation transformation. The design incorporates residual preservation, RMS normalization, and input‑dependent channel modulation to stabilize repeated spatial transformations, achieving linear token complexity. Across five UNIMAL tasks and a four‑platform quadruped setting, RecMorph outperforms existing controllers in training performance, inference throughput, and generalization to unseen bodies with up to 30 limbs, while also demonstrating robust real‑world performance on Go1/Go2 trials.
By Quanrui Rao, Yong Liu, Xueming Xiao, Yingbo Luo, Kun Wu, Zhenyu Xu, Meibao Yao
The paper introduces a reinforcement‑learning based rewriting agent that rewrites training data to reduce the distribution mismatch between supervised fine‑tuning (SFT) and a language model’s generation distribution. By formulating data rewriting as a policy‑learning problem, the authors train a lightweight LoRA rewriting policy that optimizes alignment with question‑answering style, maintains semantic diversity, and enforces task consistency. Experiments on three instruction‑tuned backbones show that models fine‑tuned with the rewritten data achieve downstream performance comparable to standard SFT while mitigating degradation on non‑downstream benchmarks, and preliminary tests suggest the policy can transfer across domains such as logical reasoning and medical question answering.
By Jiacheng Wang, Zhijie Liu, Ping Jian, Zirong Chen, Ke Ren Liao, Zhen Yang, Zhongbin Guo
The paper tackles two shortcomings of Gaussian‑process based active learning: (1) the posterior variance is independent of observed values, reducing sensitivity to data structure, and (2) it over‑inflates variance near domain boundaries, causing excessive edge sampling. The authors propose a reconstruction‑driven design density that warps sampling toward regions where the posterior mean changes rapidly, and a geometric equalizer that corrects boundary bias. Experiments on sixteen synthetic and two real‑data benchmarks show that the equalizer consistently improves function reconstruction, while the warp further enhances performance by concentrating measurements where the target function varies most.
By Sanna Jarl, Jens Sj\"olund, Jonathan J. S. Scragg, Maria B{\aa}nkestad
FRAUDSkill is a structured frozen‑weight adaptation framework for audio anti‑fraud detection that keeps the underlying audio‑language model unchanged while optimizing external skill programs, route‑specific policies, and decision rules. It combines structured output control with validation‑guided multi‑path inference to produce protocol‑compliant predictions. On the TeleAntiFraud benchmark, FRAUDSkill achieves a 73.50% Macro‑F1 score, outperforming the shared frozen‑model baseline by 31.96% and reducing invalid outputs to 1.94%.
By Chengxian Hu, Zhiming Ma, Mingjun Pan, Yifan Wang, Shun Zhang, Qifan Wang, Zhilei Zhao, Yijin Zhou, Yuxi Zhao, Huiyuan Liu, Peidong Wang, Peng Chen
The paper investigates how transformer models learn latent structure by training a small decoder-only transformer on three variants of the Alchemy benchmark. It finds that the model acquires different components of latent structure in discrete stages, with a notable asymmetry: it robustly composes fundamental transitions but struggles to decompose complex examples into atomic transitions. Layer‑specific causal interventions reveal plasticity windows where freezing layers delays or prevents stage completion, offering a detailed view of capability evolution during training.
By Rohan Saha, Farzane Aminmansour, Alona Fyshe
TeleAntiFraud 2.0 is a monthly‑frozen, audio‑based benchmark for telecom fraud detection that incorporates newly observed scam patterns while preserving earlier test sets. It uses a Mixed‑Tree Anti‑Fraud Generation Pipeline to create profile‑grounded scenarios, expands them into mixed‑tree dialogues, and renders validated speech for 900 Chinese calls (600 fraud, 300 near‑domain non‑fraud) each month. Experiments show that classifiers perform well against unrelated negatives but drop significantly against near‑domain negatives, highlighting the need for near‑domain construction and collapse‑aware reporting in realistic evaluation settings.
By Huiyuan Liu, Zhiming Ma, Yanxing Liu, Shun Zhang, Qifan Wang, Di Liu, Yifan Wang, Yuyang Deng, Haoyang Meng, Yijin Zhou, Yuxi Zhao, Chengxian Hu, Peidong Wang, Peng Chen
PDA++ is a unified, environment‑aware object insertion framework for remote sensing imagery that improves few‑shot and long‑tail recognition. It operates in three stages: Planning, which selects scene‑compatible poses using an affordance field; Decoupling, which conditions the background on pose to preserve object identity while adapting to the scene; and Assimilation, which aligns multi‑scale texture distributions via optimal transport to enhance local coherence. The method achieves a whole‑image FID of 6.28 and boosts average few‑shot recognition mAP50 by 17.69 points on optical data, while also improving ship detection on SAR imagery and maintaining performance under cross‑dataset transfer and amorphous‑target insertion.
By Xianchi Dong, Yingyan Hou, Chao Ren, Wanxuan Lu, Zihan Wei, Hongfeng Yu, Yixiao Wang, Chubo Deng, Xian Sun
The paper investigates why large vision‑language models sometimes misclassify harmful memes, attributing failures to either missing internal evidence or poor routing of evidence to the output. Using sparse autoencoders, role‑conditioned probes, and causal interventions on Gemma‑3 and Qwen3.5, the authors show that sparse readouts consistently outperform native predictions across six harmful content benchmarks, revealing a readout gap that is largely due to routing rather than representation. The study also demonstrates that calibration‑only routing recovers most of the performance gap and that the issue persists across languages and is not solely driven by OCR signals.
By Girish A. Koushik, Diptesh Kanojia, Helen Treharne
arXiv:2609.18304v1 Announce Type: new
Abstract: Large language model (LLM) agents increasingly tackle long-horizon tasks through multi-step environment interaction, yet a single erroneous action can...
By Yi Yu, Liuyi Yao, Yaliang Li, Enshu Wang, Libing Wu
The paper introduces TERN, a forecasting model that uses a delta‑rule fast‑weight memory with channel‑wise decay and learned erasure gates, combined with a seasonal reference and online adaptation. It addresses challenges in influenza forecasting such as limited seasonal data, shifting wave patterns, and misleading information after peaks. On three Cola‑GNN influenza benchmarks, TERN outperformed existing epidemic graph models and general forecasters, matching or exceeding seasonal references and demonstrating the value of its memory component.
By Shunya Nagashima, Yuta Funayama
arXiv:2604.14121v3 Announce Type: replace
Abstract: Large language models (LLMs) have become increasingly used for various tasks, often coupled with Chain-of-Thought (CoT) prompting to boost accuracy...
By Zipeng Ling, Shuliang Liu, Seonil Son, Shenghong Fu, Yuehao Tang, Yao Wan, Xuming Hu
The paper examines kinetic energy (KE) spectra and KE transfer across spatial scales in four leading probabilistic machine learning weather prediction models—NeuralGCM-ENS, FourCastNet 3, AIFS-ENS, and GenCast—by comparing them to the physics‑based IFS‑ENS model. NeuralGCM‑ENS captures the expected upscale KE transfer but underestimates mesoscale KE due to noise injection at its encoder, while AIFS‑ENS, GenCast, and FourCastNet 3 produce realistic KE magnitudes yet fail to show the expected upscale transfer, with AIFS‑ENS and GenCast accumulating KE at high wavenumbers because of spatially uncorrelated stochastic perturbations. All models display an upscale shift of the difference KE (DKE) spectral peak over time, indicating error growth, but none reproduce the rapid initial growth of ensemble spread at small scales associated with the butterfly effect, highlighting a misrepresentation of kinetic energy scale transfer despite overall skillful forecasts.
By Jiakai Chen, Joel Oskarsson, Simon Driscoll, Sebastian Schemm
TabICLv2 is a new state‑of‑the‑art tabular foundation model that outperforms existing methods on regression and classification tasks. It relies on a synthetic data generation engine for diverse pretraining, architectural innovations such as a scalable softmax attention, and optimized training protocols that replace AdamW with the Muon optimizer. On the TabArena and TALENT benchmarks, TabICLv2 surpasses the current best model, RealTabPFN‑2.5, without any tuning, while also being faster and capable of handling million‑scale datasets with limited GPU memory.
By Jingang Qu, David Holzm\"uller, Ga\"el Varoquaux, Marine Le Morvan
iMINDBench is a new benchmark for intracranial electroencephalography (iEEG) neural decoding that evaluates models on fifteen tasks across three naturalistic movie‑watching datasets from multiple institutions. It standardizes preprocessing tracks and evaluation splits to enable consistent comparisons. The study shows that pretrained systems outperform baselines within their tracks, but strong spectral baselines remain competitive, and scaling up supervised data yields only modest or task‑dependent gains.
By Geeling Chau, Saba Hashemi, Yonghyeon Gwon, Eshani Patel, Jan DeWitt, Christopher Wang, Andrii Zahorodnii, Sabera J Talukder, Danny Dongyeop Han, Chun Kee Chung, Maryam M Shanechi, Yisong Yue
FoundAna is a GNN‑assisted foundation model designed for graph anomaly detection across diverse datasets. It combines a GNN component with a transformer encoder enhanced by four positional encodings to capture both local and global structure, using reconstruction errors as anomaly scores. Experiments on nine benchmark datasets from financial, social, and citation networks show that FoundAna consistently outperforms state‑of‑the‑art baselines.
By Suprim Nakarmi, Chahana Dahal, Yue Zhao, Junggab Son, Zuobin Xiong
EdgeReMIND is a linear memorization model designed for temporal multi-relational link prediction on the Temporal Graph Benchmark 2.0 (TGB 2.0). It achieves the highest reported test mean reciprocal rank (MRR) on six of eight TGB 2.0 datasets and is the only relation‑aware method that runs on all datasets, overcoming the scalability ceiling that limits existing embedding methods. The model uses learned per‑relation weights over data‑calibrated features, making it a practical state‑of‑the‑art baseline for large‑scale temporal graphs.
By Bryant Pollard