The paper compares latent representations in Selective State Space Models (SSMs) like Mamba and Transformers such as Pythia using Sparse Autoencoders. Across a 10‑million token corpus, 99.98% of Mamba features align closely with Pythia’s, supporting the Universality Hypothesis that core semantic representations are similar across architectures. A tiny 0.02% of features diverge, with Mamba’s recurrent bottleneck causing it to compress syntactic anomalies into polysemantic neurons, whereas Pythia’s attention can isolate distinct formatting edge‑cases.
By Rithin Nagaraj, Rupa Laalasa Oruganti, Prerna Subhashchandra Kunder, Ashwini M Joshi
The paper introduces OrbitTrace, a benchmark of 50 physics‑grounded compute‑availability traces from satellite orbits, and investigates whether specialized interruption‑resilient optimizers are needed when training is interrupted by predictable compute gaps. Experiments on CIFAR‑10/ResNet‑18 and GPT‑2/AdamW show that a strong checkpoint‑and‑resume baseline that preserves full optimizer state and indexes learning‑rate schedules in effective time matches uninterrupted training, rendering most availability‑aware methods unnecessary. Only in a narrow regime—large models with non‑persistable optimizer state and frequent short pauses—does reactive adaptation recover a modest portion of the state‑loss penalty, and even this benefit disappears for eclipse‑scale gaps.
By Subhadip Mitra
The paper introduces EmoDiPyraTrans, a development-regularized differential graph Transformer designed to decode emotions from EEG signals across unseen individuals and populations while maintaining neural interpretability. Evaluations on five datasets (SEED, FACED, MAHNOB-HCI, DEAP, DREAMER) show high cross‑subject accuracies, with the model outperforming others in accuracy and positive‑class F1. Additional experiments demonstrate its ability to distinguish healthy from depressed participants and reveal spatial‑spectral neural signatures across frontal, temporal, central, and parietal regions with an alpha‑centered frequency preference.
By Dongyi He, Bin Jiang, Xiangkai Wang, Yun Zhao, Hongjie Yan, Wai Ting Siok, Nizhuan Wang
The study audits demographic leakage in German-language resumes generated by large language models. Using ChatGPT, Gemini, and Qwen 3 variants, the authors generate resumes from anonymized profiles, varying only gender- and ethnicity-associated names while keeping qualifications constant. Even after anonymization and gender-neutralization, classifiers can reliably distinguish male- from female-generated resumes, driven by subtle differences in gender-neutral terminology rather than overtly gendered wording; ethnicity-related leakage remains weak.
By Charlotte Leininger, Helena Veit, Matthias A{\ss}enmacher, Andreas Bender
The paper introduces the Situated Identity Test (SIT), a framework that assesses whether a language model’s behavior can be traced to a specific developmental lineage rather than merely imitating a persona. SIT requires agents to possess accurate knowledge of their recorded experiences while appropriately ignoring ungrounded information, and it demonstrates that policies based only on compressed profiles are limited in distinguishing between colliding life histories. The authors present SITBench, an evaluation suite with 25 profile‑collision pairs and 10,000 probes across nine model architectures, and provide open‑source tools and pilot results on state‑of‑the‑art foundation models.
By Jun He, Deying Yu
arXiv:2609. 22880v1 Announce Type: cross Abstract: LLM rerankers add of the order of \$0.
By Andre Bacellar
The paper demonstrates that quantum transformer blocks can be intrinsically interpretable by tracking quantum mutual information, entanglement entropy, and state fidelity across layers. Experiments on four synthetic tasks show that learned mutual information aligns with task structure, entanglement is essential for accuracy, and mutual information predicts prediction correctness. These findings are validated on IBM Quantum hardware, illustrating that quantum computation’s physics can provide observable interpretability signals.
By Diego Iacopetta, Andrea Gasparini
The paper introduces UniK, a universal knowledge perception platform designed to serve both digital AI—such as chatbots and agent workflows—and physical AI, which controls robots and autonomous systems. UniK handles the entire knowledge lifecycle—ingestion, enrichment, indexing, retrieval, and continuous evaluation—across diverse modalities including text, video, molecular data, and sensor telemetry, without task‑specific fine‑tuning. In five digital AI domains, UniK paired with a 70‑billion‑parameter model matches or surpasses larger proprietary LLMs, achieving high retrieval‑augmented generation accuracy on government data, medical QA, and chemistry tasks, and it also addresses similar data challenges in physical AI world‑model training.
By Nirmit Desai, Kunal Sawarkar, Aditya Mahakali, Dongkon Lee, Kevin Park, Eric Song
The paper demonstrates that large language models (LLMs) used for forecasting real‑world events can be manipulated by simply publishing new articles, even without direct access to the model or its retriever. By injecting a small number of targeted news pieces into a common crawl corpus, an adversary can flip over half of the forecast probabilities and significantly degrade forecast accuracy. The study also shows that common defense strategies can be cheaply bypassed, highlighting the vulnerability of probabilistic LLM judgments to information‑supply‑chain attacks.
By Yuan Lu, Yukuan Zhang
Euston is an 8‑B parameter mathematical claim‑verification model that resists producing false derivations when presented with corrupted theorems. It was trained on 3,026 matched true/corrupted statement pairs generated by GraphSynth, a probabilistic factor‑graph generator, and fine‑tuned from DeepSeek‑R1‑8B using GRPO. On a balanced held‑out split, Euston’s balanced accuracy rose from 29.50 % to 63.75 %, and its discrimination gap improved from –0.5 % to +27.5 %, while maintaining comparable general mathematical ability and reducing response length and truncation rates.
By Zehua Cheng, Wei Dai, Jiahao Sun
Re:CAP is a reference‑free audit loop for retrieval‑augmented generation (RAG) pipelines that probes for missing documents instead of enumerating all relevant ones. It identifies covered topics, generates probing questions, retrieves candidate documents, and uses an LLM judge to keep only those that add new information. On several benchmarks, Re:CAP recovers a significant portion of gold documents that flat BM25 or hybrid retrieval misses, and human evaluation shows most of these documents add new information.
By Aviral Joshi, Hanoz Bhathena, Max Nelson, Saket Sharma
The paper presents a smartphone‑compatible, retrieval‑augmented language model tailored to Bangladeshi statutory law. By distilling a 9‑billion‑parameter Gemma‑2 teacher into a 2‑billion‑parameter student using supervised fine‑tuning and QLoRA, the authors achieve significant gains in ROUGE‑L and BERTScore on an English benchmark while keeping the model lightweight (1.6 GB) and operable offline on a Pixel 6. The system retrieves from 36,029 statutory passages using a hybrid dense/BM25 approach, and cross‑lingual evaluation shows effective Bangla query handling against an English‑only corpus, with a practicing lawyer rating the responses highly in a pilot study.
By MD. Nafis Kamal, Mahadi Hasan Fahim, Talha Ridwan, Nadifa Zaman, Fariha Roushon Florin, Farig Yousuf Sadeque, Saadat Rafid Ahmed
This thesis explores how to select and adapt NLP models for global health literature when annotated data and computational resources are scarce. It compares skip‑gram word2vec models trained on increasingly large specialized corpora with BioWordVec for semantic tag discovery, finding that larger coverage does not always yield more useful domain associations. The study also evaluates convolutional spaCy models versus a RoBERTa transformer for named entity recognition, noting a trade‑off between higher F1 scores and longer inference time, and investigates MiniLM few‑shot versus BART‑MNLI zero‑shot classification for multi‑label topic classification, highlighting practical constraints of inference cost.
"whyItMatters":"The work provides empirical guidance on balancing model accuracy and resource demands for building knowledge systems in low‑resource global health settings."
By Genis Skura, Antoine Geissb\"uhler, Jean-Luc Falcone
The BD-LSC dataset introduces a bi‑directional lexical semantic change benchmark that tracks sense gain, loss, and stability across three time periods, while the ST‑WSD dataset offers fine‑grained, instance‑level sense annotations for words that blend slang and standard usage. These resources enable systematic evaluation of diverse models—including unsupervised clustering, supervised learning, transformer‑based approaches, and large language models—on tasks such as exact sense matching and multi‑label accuracy. The evaluation shows that few‑shot GPT‑4o performs best overall, yet all systems struggle with rare slang senses, highlighting a key open challenge in the field.
By Afnan Aloraini, Riza Batista-Navarro
The paper introduces a free, label‑free visual evidence signal that improves fine‑grained vision‑language reasoning. By selecting image crops that maximize the model’s answer distribution peak, the method locates answer‑bearing regions without training or annotations, boosting accuracy from 70 % to 85 %. The evidence gap also complements model confidence, enabling better correctness prediction and error flagging.
By Santi Ram Tiwari, Nihal Naik, Devbrat Pandey, Nishant Sinha
DeCo introduces an efficient decouple-to-couple learning framework for multi-task visual grounding, addressing conflicts between localization and segmentation tasks. It first applies Task-aware Semantic Decoupling (TSD) to separate shared visual cues into task-specific features guided by salient words, then uses Hybrid Prior Coupling (HPC) to merge sentence-level semantic priors with mask-derived spatial priors for improved grounding. Experiments across multiple natural and remote sensing datasets show that DeCo achieves state‑of‑the‑art performance while requiring only lightweight trainable parameters on a frozen multimodal encoder.
By Xiaoqiang Lu, Licheng Jiao, Long Sun, Yuting Yang, Xu Liu, Lingling Li, Wenping Ma, Fang Liu
The paper proposes a new sequence-to-sequence formulation for multi-view stereo (MVS) that jointly predicts 3D geometry for all input views using a global transformer architecture. It introduces ray‑map embeddings to inject camera parameters into image tokens and a unified global cost volume to capture 3D structure across all views. Experiments on public benchmarks demonstrate state‑of‑the‑art performance, outperforming both traditional MVS and feed‑forward reconstruction baselines.
By Aoxiang Fan, Corentin Dumery, Nicolas Talabot, Pascal Fua
SPHQuant introduces a rotation‑free spherical weight‑only quantization framework for Vision‑Language Models, decomposing 8‑dimensional weight vectors into sign, radius, and a positive unit direction. By isolating outlier magnitudes in the radius and allocating extra precision there, it mitigates accuracy loss at extreme low bit‑widths. The method also employs a compact positive‑direction codebook with angular fine‑tuning and a hardware‑friendly GEMV kernel, achieving state‑of‑the‑art performance while boosting decode throughput by 30.3% on RTX A6000 compared to QTIP.
By Kewei Zhang, Zheng Chen, Haotong Qin, Yulun Zhang
The paper introduces the directional linear separability measure (D‑LSM) to quantify how much linear separability is preserved or improved by injective affine maps in neural networks. It characterizes the geometry supporting D‑LSM, proves its invariance under injective affine embeddings, and derives conditions for gated activations (ReLU, GELU, SiLU) to preserve and recover samples. Experiments validate the theoretical bounds, demonstrate affine‑tube constructions that achieve guaranteed recovery, and apply the method to Vision Transformer representations to obtain early post‑activation separability certificates.
By Yi Wei, Xuan Qi, Suorong Yang, Furao Shen
The paper introduces a four-step pipeline that mines decision rules in the latent space of an FT-Transformer and then translates those rules back into measurable clinical features. By treating embedding dimensions that separate patient groups as latent biomarkers, small decision trees are used to extract rules, which are then mapped to raw features using gradient-input saliency and CLS attention attribution. Across six public clinical datasets, the translated rules generally outperformed raw-feature rules, achieving significant AUROC gains, though some high-performing latent rules could not be fully captured by simple raw-feature conditions.
By Majid Lotfian Delouee, Hamed Ayoobi, Sjors G. J. G. In 't Veld, Martijn C. Schut