The Industrial Kinematic Trajectory Model (IKTM) is a new coordinate‑free autoregressive generator that models industrial vehicle motion using kinematic sequences of speed and heading change, eliminating the need for absolute spatial references. It employs a causal transformer with probabilistic mixture heads and a duration‑conditioning signal, extending a prior Markovian approach with deep sequence modeling. Trained on a single site, IKTM achieves near‑optimal turn‑rate distribution matching and low Jensen‑Shannon divergence across four sites, outperforming the earlier model by 5.3–6.0× in oracle‑length protocols while maintaining exact duration tracking in rollouts.
By Max Amiri, David Eyers
The paper argues that language models’ intransitive preferences arise from multiple internally consistent latent orderings rather than noise around a single ordering. By demonstrating that a single ordering cannot explain observed inconsistencies and introducing a noise‑augmented mixture Bradley‑Terry model, the authors show that mixtures of orderings better capture preference structure across several models and tasks. A case study on Moral Machine dilemmas further illustrates that models can share latent components even when aggregate preferences differ.
By Aviral Chawla, William H. W. Thompson, Jean-Gabriel Young
The paper investigates Hierarchical Reasoning Models (HRM), a class of hierarchical Transformer-based latent reasoning models, across Sudoku, Maze, and ARC-AGI-2 tasks. By comparing HRM to Transformer baselines, applying causal interventions on recurrent states, and conducting linear probe and sparse autoencoder ablations, the authors find that recurrent models outperform one-pass baselines, that high- and low-level states contribute differently across tasks, and that ablations of sparse autoencoder features cause larger behavioral changes than probe-direction ablations. The study concludes that HRM implements constraint‑aware iterative refinement on a puzzle‑specific solution state, with component contributions varying without a compact, causally important feature set.
By Leo Raphael Rodrigues, Jian Kang
TWIG (Time‑Causal Wavelet Operator for Irregular Graphs) is a graph‑native neural operator designed for autoregressive surrogate modeling on static irregular graphs. It transforms each node’s history into causal multiscale temporal features, separating recent changes from slower memory components, and propagates these through graph‑wavelet operator blocks with gated pointwise channel mixing. The architecture is causal by construction, enabling closed‑loop forecasting where predictions are recursively reused as future inputs, and it consistently outperforms non‑time‑causal baselines across three irregular‑domain forecasting problems.
By Subashree Venkatasubramanian, David A. Barajas-Solano, Chuyang Liu, Daniel M. Tartakovsky, Dipankar Dwivedi
The paper introduces Calibrated Clipping, a dynamic method to align FP8 quantization bounds with high‑precision BF16 distributions, thereby mitigating training instability in full‑pipeline FP8 reinforcement learning for large language models. It identifies that compounded FP8 noise distorts importance ratios, causing entropy surges and garbled outputs. Experiments across GRPO and DAPO algorithms on 8B‑32B models show the technique restores performance to BF16 levels.
By Fanchao Chen, Ziheng Jiang, Ziyun Wei, Zheng Zhong, Du Li, Chi Zhang, Haibin Lin, Shivaram Venkataraman
The paper introduces CASP-LLM, a coverage‑aware semantic prompting framework that mitigates bias toward dominant temporal patterns in prompt‑based time series forecasting. Unlike traditional similarity‑based retrieval that selects top‑K candidates by cosine similarity, CASP‑LLM uses usage‑tracking and a saturating‑gate regularizer to diversify prompt selection without adding learnable parameters. Experiments on six long‑term and the M4 short‑term benchmarks show that CASP‑LLM matches or outperforms similarity‑based LLM forecasters in most settings, with failures traced to cross‑batch usage rather than within‑retrieval redundancy.
By Daeun Ji, Minkyoung Kim, Dongkuk Kim, Yohan Lee, Beomsoo Kim, Beakcheol Jang
WaveFront Decoding (WFD) is a training‑free, self‑speculative decoding framework for looped language models that reduces decoding latency by batching draft and verification steps within the same recurrent‑block calls. By exploiting intermediate recurrence outputs as draft predictions and weight sharing to process token states at different depths together, WFD arranges mixed‑depth states into a diagonal wavefront, allowing shallow‑depth drafting while deeper‑depth verification proceeds concurrently. Experiments on six Spec‑Bench task categories show WFD achieving up to 4.81× speedup on Huginn‑3.5B compared to autoregressive decoding, outperforming traditional draft‑then‑verify approaches.
By Hyeongju Ha, Jae-Joon Kim
The paper introduces the first controlled benchmark for optimizers in discrete diffusion models, evaluating seven optimizers (AdamW, Lion, Muon, SOAP, MARS, MARS‑M, Schedule‑Free) across four diffusion formulations: masked diffusion on text8, uniform diffusion on QM9 and LM1B, and Gaussian diffusion on CelebA‑64. Each optimizer undergoes the same search protocol and is retrained with full budget and multiple seeds, revealing that AdamW, while strong, is not universally optimal and that optimizers validated on autoregressive language models (Muon, MARS‑M, SOAP) can outperform tuned AdamW on certain tasks.
By Arman Bolatov, Egor Shulgin, David Li, Abduragim Shtanchaev, Sebastian U. Stich, Maxim Panov, Eric Moulines, Peter Richt\'arik, Martin Tak\'a\v{c}
The paper introduces Neural Spectral Capacity (NSC), a closed‑form metric derived from the singular‑value spectrum of weight matrices that can be computed solely from a network’s architectural specification. Unlike traditional measures such as #Params and #FLOPs, NSC captures architectural structure (depth, width, head, FFN allocations) and can be evaluated without instantiating the model, data, or gradients. Using a dynamic‑programming solver (NSC‑DP), the authors demonstrate that NSC can efficiently identify architectures that outperform existing training‑free proxies across Transformer and CNN families, and achieve state‑of‑the‑art results in tasks such as WikiText‑103 and commonsense reasoning with LLaMA‑7B.
whyItMatters":"NSC provides a fast, architecture‑only proxy that outperforms conventional metrics and training‑free proxies, enabling more effective design and pruning of large models without costly training or data."
By Chenyu Zhu, Ruoyu Zhao, Zhichao Lu
The paper audits LLM-based explainers attached to an Active Inference agent that manages German grid demand, testing three large‑language‑model backends (GPT‑4o, Claude‑3‑Opus, Gemini). By injecting corrupted observations and attacker‑controlled text, the study finds that the explainers fail to flag errors, produce fluent but incorrect rationalizations for wrong actions, and can be steered to exfiltrate data. The authors propose mitigations but do not evaluate them, emphasizing that explanations are never verified for truth before operators rely on them.
By Param Raval, Rohit Shenoy, Archana Vaidheeswaran
The study compares six state‑of‑the‑art sleep‑staging foundation models on the MESA polysomnography dataset under three signal conditions: EEG only, ECG only, and EEG+ECG. Results show that EEG alone yields the highest accuracy (BIOT macro‑F1 = 0.7237), while switching to ECG alone incurs a consistent accuracy loss of about 0.35 macro‑F1 and reduces data rate to one‑third. Adding ECG to EEG offers little benefit for most models, indicating that EEG carries most of the sleep‑staging signal and that wearable‑compatible ECG alone is a viable but less accurate alternative.
By Hassan Mehdi, Riku Klen, Ayse Kosal Bulbul, Suzanne Timmons, Abdulhamit Subasi, Wei Chen, Zou Zhu, Muhammad Irfan
The study investigates whether language models can explicitly report constraints they have learned through post‑training fine‑tuning. Using constrained recipe generation with five banned ingredients, the authors compare supervised fine‑tuning (SFT) and Group Relative Policy Optimization (GRPO) against an untrained baseline on a Constraint Awareness Benchmark. Both fine‑tuning methods increase behavioral compliance from 4% to about 90% but reduce explicit constraint reporting and erode retained third‑person knowledge, with GRPO showing more destructive effects. The results suggest that reward‑based signals may suppress constraints context‑independently, and that models fail to enumerate constraints on request even when they can avoid them internally.
By Arin Agarwal
Swiss-Knife is a framework that extends decode‑time alignment for frozen language models by treating the alignment specification as a runtime object. It introduces hot‑swappable scoring blades, a batch normaliser, a pairwise aggregation operator, and a selection rule, and characterises admissible aggregation operators with a representation theorem. In experiments, Swiss‑Knife paired with DPO‑LoRA blades and an uncertainty‑aware pairwise tournament outperforms six existing decode‑time methods, achieving a higher harmonic F1 score, lower refusal rate, and faster objective reconfiguration.
By Agnibh Karmakar, Mayur Parvatikar, Shreyash Dhoot, Amit Dhanda, Aman Chadha, Kapil Wanaskar, Vinija Jain, Amitava Das
MechaTerp-TRACE is a new framework that systematically ablates individual components of language models to measure their causal contribution to producing a named entity. By applying TRACE to thirteen instruction‑tuned dense decoder models, the study finds that a small set of positionally fixed components consistently carry the most influence across models and prompts, while the remaining support is evenly distributed. This suggests that entity knowledge is largely embedded in generic generation machinery rather than in isolated, findable components.
By Brandon Colelough, Davis Bartels, Madeline Bittner, Dina Demner-Fushman
The paper introduces a cross-dialect Named Entity Recognition (NER) framework for Bangla, leveraging the ANCHOLIK-NER dataset that covers five major regional dialects. Using a Leave-One-Dialect-Out Cross-Validation strategy, eight transformer-based models were evaluated, with Multilingual-E5 Large achieving the best performance (F1 up to 97.26% on Mymensingh, 82.38% on Chattogram). Local Interpretable Model-agnostic Explanations (LIME) revealed that the models rely mainly on the surface form of entity words rather than surrounding context, suggesting a direction for future improvement.
By Shamim Rahim Refat, Faika Fairuj Preotee, Shuvashis Sarker, Shifat Islam, Bidyarthi Paul, Mohammad Ashraful Hoque
The paper investigates time‑incremental continued pretraining (CPT) of large language models using web‑scale data that overlaps across snapshots. Across six open‑weight models, CPT improves factual recall without catastrophic forgetting, while the cost is negligible and data quality outweighs quantity. The study identifies optimal learning rates, shows LoRA can match full CPT, and demonstrates that CPT gains transfer to fine‑tuned models.
By F{\i}rat \"Oncel, Salman Hussain Ali, Mirco Ravanelli, Cem Subakan, \c{C}a\u{g}atay Y{\i}ld{\i}z
MinCU is a new benchmark for grounded minimal‑change understanding that presents pairs of near‑identical images differing by a single atomic variation in object category, attribute, count, or spatial position. Models are evaluated on their ability to describe the change, localize the changed region, and identify the changed entity. The authors also introduce SG‑ISA, a structured autoregressive method that decomposes the task into a Think‑Locate‑Describe sequence, showing that fine‑tuning with SG‑ISA improves both grounding accuracy and description quality while reducing reasoning‑token overhead.
By Chaoqian Mu, Wenhao Wu, Zichen Liang, Jiaxu Li, Lijun Wang, Yifan Wang, Huchuan Lu
The study developed a multimodal ultrasound and clinical data model to predict microvascular invasion (MVI) preoperatively in hepatocellular carcinoma (HCC). Using data from 489 patients across eight centers, the model combined B-mode ultrasound, color Doppler flow imaging, dynamic contrast-enhanced ultrasound, and clinical information, achieving an AUC of 0.8953 in external validation. Dynamic contrast-enhanced ultrasound contributed the most predictive power, while other modalities and clinical data added complementary value.
By Jun Cheng, Yuanyuan Kong, Qing Huang, Xiaotong Tan, Licong Dong, Yulong Han, Wufeng Xue, Ruobing Huang, Dong Ni, Qi Yang, Jie Yu, Ping Liang
The paper explores the use of Closed‑Circuit Television (CCTV) footage to estimate urban rail platform crowding in real time. It compares three computer‑vision methods—object detection and counting, crowd‑level classification with a Vision Transformer, and semantic segmentation—to extract crowd-related features. A novel convex ridge regression technique is introduced to convert segmentation outputs into passenger counts, and the methods are evaluated on a privacy‑preserving dataset of over 600 hours of Washington Metropolitan Area Transit Authority (WMATA) video, showing that CCTV alone can provide valuable real‑time crowd estimates.
By Riccardo Fiorista, Awad Abdelhalim, Anson F. Stewart, Gabriel L. Pincus, Ian Thistle, Jinhua Zhao
The study evaluates how quantization affects accuracy and safety of five 7‑8B language models on clinical benchmarks. INT8 GPTQ shows minimal degradation (≤1.9%) across tasks, while INT4 causes substantial, model‑dependent drops, especially in high‑risk scenarios and safety metrics. Recovery methods such as clinical calibration substitution and QLoRA fine‑tuning yield mixed results, underscoring the need for task‑specific validation.
By Leonard Twagirayezu, Prasenjit Mitra