The paper introduces two token-error supervision methods for sparse mixture-of-experts (MoE) large language models, aligning routing affinities with token-level cross-entropy loss. The first method predicts an error score per expert and uses it to adjust affinities before top‑K selection, while the second directly aligns router affinities to the model’s objective without extra heads or inference changes. Experiments on Granite and ARC‑Challenge show accuracy gains of about 2.3–2.94 percentage points over a parameter‑matched baseline, preserving the native sparse execution budget.
By Yury Nahshan, Nati Daniel, Jacob Goldberger, Yoli Shavit
arXiv:2609. 38023v1 Announce Type: new Abstract: Physics-Informed Neural Networks (PINNs) embed governing equations into deep learning, but enforce them only through loss residuals, leaving highly oscillatory wave behavior to be discovered by optimization.
By Huiwen Zhang, Feng Ye, Chu Ma
The paper introduces agentic meta‑reasoning, a structured inference‑time framework that explicitly manages control decisions—such as selecting partial work, restarting, or stopping—during long‑horizon agentic tasks. By delegating task execution to workers and consolidating decisions through a lightweight controller that references persistent memory, the method reduces the need to replay full histories. Experiments on ProgramBench and other benchmarks show that meta‑reasoning improves performance over direct control baselines, especially as computation budgets increase, and reveals greater reuse of earlier work and higher solution coverage.
By Paras Dahal, Anton Bakhtin, Taco Cohen, Zhengxing Chen, Carole-Jean Wu, Rob Fergus, Scott Yih, Gabriel Synnaeve, Ruslan Salakhutdinov, Sanjeev Arora, Jason Weston, Anirudh Goyal
The paper introduces CALIBRA, a calibration-first multimodal temporal learning framework designed for transferable asthma‑risk forecasting across varying patient cohorts and sensor ecosystems. It processes multiple data streams—environmental, pulmonary, symptom, medication, wearable, and context—using dedicated recurrent encoders, a reliability‑conditioned gate, and gradient‑reversal training to mitigate cohort bias. CALIBRA employs a shrinkage‑based hierarchical logistic layer for probability calibration and split conformal prediction for abstention‑capable prediction sets, achieving competitive performance on a semi‑synthetic three‑cohort benchmark with controlled distribution shift.
By Taimoor Ahmad
TRACE is a deployable framework that enhances oncology language models by separating offline structure learning from online inference. It organizes oncology concepts and relations into an updatable tree‑relational structure, refines it with LM‑loss evidence, and retrieves compact prompt evidence during inference. The approach improves performance on ten classification tasks and a QA benchmark, outperforms vanilla RAG and generic GraphRAG, and provides interpretable evidence paths aligned with clinical reasoning.
By Jizheng Lai, Yingyun Li, Ying Qin, Haiyang Qian
PCaPaint is a prostate cancer inpainting method that uses latent diffusion models (LDMs) and introduces a conditioning strategy where the condition image is filled with Gaussian noise to mitigate shortcut learning. It also proposes a new training objective that focuses on errors within the lesion region and a multi‑sequence latent design that separately compresses T2w and DWI&ADC scans to preserve their distinct frequency characteristics. Experiments show that the synthetic data generated by PCaPaint improves downstream tasks such as prostate lesion segmentation, patient‑level classification, and lesion‑level detection, outperforming recent state‑of‑the‑art LDM‑based tumor inpainting methods in both performance and image quality.
By Levente Lippenszky, Hongxu Yang, Marcell D\"om\"ot\"or, Krisztian Koos, L\'aszl\'o Rusk\'o
The paper introduces the Agentic Commerce Bench (ACB), a benchmark for measuring fraud in AI agents that autonomously spend money. It presents a taxonomy of agentic commerce fraud, a dataset of twenty fraud classes derived from real production data, and an open‑source detector stack called gordonguard for auditing and replaying hostile counterparties. The study shows that current reasoning layers and security scanners perform poorly on many classes, highlighting the need for better detection mechanisms.
By Ankit Srivastava, Debjyoti Paul
UNBIND is a code unlearning framework that selectively removes memorized code from large language models at inference time while keeping the model weights unchanged. It constructs separate directional steering for hidden states that correspond to target code, enabling high forgetting rates (97.3–99.1% reduction in target code reproduction) with minimal loss in programming utility. Across multiple baselines, corpora, and evaluation metrics—including F‑BLEU, HumanEval+, and MBPP+—UNBIND consistently achieves the best joint forgetting and utility scores, and it effectively eliminates long exact code spans in repeated extraction tests.
By Zhengyang Shan, Jiayun Xin, Yanjun Lin, Xu Qian, Zhiang Liu, Minghui Xu, Yue Zhang, Qin Hu, Kun Li, Xiuzhen Cheng
The paper investigates how multimodal large language models are more susceptible to prompt injection when adversarial instructions are presented as text rather than as non-textual inputs like images. It proposes a training‑free defense that renders untrusted payloads into typographic images (or audio) before they reach the model, a method called Pictionary. Experiments on ten models and two benchmarks show that this approach significantly lowers attack success rates while maintaining normal functionality, and that fine‑tuning on image‑rendered instructions can further reduce the modality gap.
By Jie Zhang, Andrei Baroian, Jan N. van Rijn, Avital Shafran, Florian Tram\`{e}r
ThinQuant introduces efficient rotation learning for low‑bit weight and activation quantization of large language models by reducing calibration data through a geometric selection of activations and solving a lower‑dimensional optimization problem via an ADMM algorithm. The method achieves comparable or better quantization performance with dramatically fewer calibration points, completing rotation calibration for Llama‑3‑70B in under 12 minutes and for Llama‑3.1‑405B in just over 2 hours on a single GPU. ThinQuant outperforms existing gradient‑free approaches such as DartQuant and gradient‑based SpinQuant in both speed and perplexity metrics on WikiText‑2.
By Mehdi Makni, Ryan Lucas, Rahul Mazumder
The paper introduces ReviveBench, a benchmark designed to evaluate coding agents’ ability to revive non‑running software and reconstruct industrial engines from open specifications. It comprises two families of tasks—revival (ten tasks addressing dependency issues, missing modules, legacy builds, and GPU models) and reconstruction (thirteen tasks covering numerical, geometric, hardware, and transactional systems). The benchmark uses hidden verifiers calibrated against native environments, engineering tools, or reference implementations, and the authors report that the strongest evaluated model passes all revival tasks and most reconstruction tasks, while also uncovering verifier defects that highlight measurement error in executable verification.
By Tianyu Liu, Dingyuan Dai, Yufan Du, Zhen Yang
The paper introduces Encoder‑Sharing Hierarchical Multi‑Task Federated Learning (EN‑HMTFL) for vehicular ad hoc networks, combining cluster‑based hierarchical federated learning with a globally shared encoder and vehicle‑local decoders. This design allows vehicles performing different perception tasks to collaboratively learn a transferable feature representation while keeping raw data and task‑specific decoders local. Experiments on MNIST and GTSRB datasets show that EN‑HMTFL can improve accuracy by up to 24.0% and reduce communication rounds by up to 69 (28.8%) compared to a representation‑sharing benchmark.
By M. Saeid HaghighiFard, Sinem Coleri
The paper introduces MATE, a reinforcement‑learning‑based post‑training framework for unified multimodal models that lets the generation and understanding branches challenge each other instead of cooperating. In MATE, each branch proposes candidate outputs that the other must reproduce, and the solver is trained on the worst‑handled candidate, creating an evolving adversarial loop without a separate adversary. Experiments on Janus‑Pro‑1B show that MATE improves generation and understanding metrics, including GenEval (+2.4), DPG‑Bench (+1.7), and an average of nine understanding benchmarks (+0.7), while enhancing consistency across image‑text cycles.
By Wentao Zhou, Weijie Gan, Jiayun Wang
The paper investigates how chunked KV‑cache compression, which reduces memory and attention costs by grouping consecutive tokens, introduces a new positional coordinate called a token’s phase. It finds that models using this compression exhibit phase sensitivity: retrieval performance can vary by up to 40 percentage points depending on a token’s phase, creating periodic weak spots that average benchmark scores hide. By pretraining transformers with different compression designs and applying causal interventions, the authors show that attention components specialize asymmetrically for different phases, and theoretical analysis suggests gradient dynamics may favor such specialization.
By Xingyu Zhu (Luke), Pu (Luke), Yi, Ziheng Cheng, Ang Lv, Jing Liu, Lexing Ying, Yiyuan Ma, Xin Dong
The paper introduces Explainability from Training (EFT), a model‑agnostic method that tracks how deep learning models learn and organize evidence during training. EFT is applied to four leading T cell receptor‑epitope prediction models, revealing distinct learning trajectories for CNNs and transformers, conflicts between TCR alpha and beta chain evidence, and differences in feature preferences when using real versus predicted structural data. The authors also present a new benchmark, TCR‑XAI2, comprising 388 experimentally resolved TCR‑epitope structures and several predicted models to evaluate these insights.
By Jiarui Li, Zixiang Yin, Samuel Landry, Zhengming Ding, Ramgopal Mettu
The paper introduces Chameleon, a channel‑dependent state space model for multivariate time series forecasting that allows data‑dependent, fine‑grained interactions across variables while maintaining linear scaling with the number of variables. By integrating selective state space models with a Kalman filter and adapting GatedDeltaNet as the backbone, Chameleon improves generalization and achieves lower MSE and MAE on strongly dependent ODE and PEMS datasets compared to both channel‑independent and prior channel‑dependent methods. Across 28 benchmark settings, it outperforms baselines in the majority of cases and demonstrates competitive training‑time and memory efficiency on Traffic and ETT datasets.
By Yu-Cheng Wu, Fan-Keng Sun, Li-Chun Lu, Duane S. Boning
The paper introduces Online Versatile Incremental Learning (Online VIL), a scenario where both class concepts and visual domains evolve simultaneously without explicit boundaries. To address this challenge, the authors propose TopFlow, a framework that combines Domain‑agnostic Flow Matching (DFM) and Global Topology Preservation (GTP) to maintain domain‑agnostic representations and the global structure of the feature space without storing past examples. Experiments show that TopFlow outperforms existing methods in the Online VIL setting, offering a promising direction for building continual learning systems in realistic dynamic environments.
By Jae-Ho Lee, Min-Yeong Park, Jun-Yeong Moon, Jung Uk Kim, Gyeong-Moon Park
arXiv:2609.36515v1 Announce Type: cross
Abstract: A common assumption in language model development is that cognitive abilities are organized around a general, domain-free intelligence factor, like f...
By Faiz Ghifari Haznitrama, Afrizal Hasbi Azizy, Faeyza Rishad Ardi
arXiv:2609.36525v1 Announce Type: cross
Abstract: Distributed inference has become an indispensable part of deploying medical models under practical latency, memory, and throughput constraints. Altho...
By Yifei Wang, Xiaohan Zhang, Youtao Ding, Tianlin Li, Xiaoyu Zhang, Yida Yang, Li Pan
arXiv:2609.36570v1 Announce Type: cross
Abstract: Indirect prompt injection makes an LLM agent treat untrusted retrieved text as instructions. We present CounterSteer, an inference-time defense that...
By Mark Russinovich