The paper investigates whether extra computation in recurrent Transformers should be allocated to more temporal steps or greater physical depth. Using Latent Recurrent Transformers (LRTs), the authors insert a latent thought token between vocabulary tokens, allowing each token to pass through the same $L$ layers twice while sharing parameters. Experiments on 16‑ and 20‑layer mixture‑of‑experts NanoChat backbones show that a single thought token brings a shallower model within 0.006–0.004 bits per byte of a double‑depth counterpart, recovering 67–81% of the improvement with roughly 48% fewer parameters.
By Zeyi Huang, Xuehai He, Yong Jae Lee, Yelong Shen
SpecQuant is a training‑free framework that merges speculative decoding with multi‑parent quantization to enable adaptive, efficient inference of large language models. It generates several quantized variants (INT4, FP8, FP16) from a single base model and routes queries to the appropriate variant based on predicted complexity, using lightweight models for simple tasks and full‑precision models for complex reasoning. Evaluations on Qwen2.5 models across MMLU, AlpacaEval, and GSM8K show 35‑43% speedups with less than 2% accuracy loss, facilitating practical on‑device LLM deployment without specialized infrastructure.
By Harish KB, Jagadeeswaran M, Pradheep P, Yuvanesh S, Sivakumar T
The paper argues that zero‑shot time‑series forecasting should be treated as an evidence‑access claim rather than merely a no‑parameter‑update condition. It introduces a source‑first taxonomy that distinguishes three evidence sources—frozen LLM prior reuse, parametric time‑series pretraining, and retrieval‑augmented external memory—from the architectures that implement them. The authors further outline four audit questions—task interface, forecast object and scoring, prediction‑time context, and resource budget—to make zero‑shot leaderboards transparent and comparable.
By Delun Kong, Wanyun Ling, Chenxi Liu, Ziyue Li
FairLMs is a Python library designed to streamline fairness research in language models by unifying bias measurement, mitigation, and evaluation evidence. It offers 33 intrinsic and extrinsic metrics, 14 mitigation components across four intervention categories, 14 diagnostic tools, adapters for major Transformer architectures and hosted APIs, and benchmark loaders. The library enforces explicit declarations of model capabilities and input requirements, ensuring compatibility and reproducibility across components and datasets.
By Jiale Zhang, Michael Larionov, Zichong Wang, Zhipeng Yin, Wenbin Zhang
The paper introduces Exemplar Partitioning (EP), an unsupervised technique that constructs interpretable feature dictionaries from large language model activations by clustering streamed activations into Voronoi regions defined by exemplars and their averages. EP allows comparison of dictionaries across layers, checkpoints, and architectures, and demonstrates utility in interpreting model behavior, tracking training dynamics, detecting hidden concepts, and enabling targeted interventions. Experiments on Gemma‑2‑2B and Llama‑3.1‑8B show EP can reveal how instruction tuning reorganizes harmful prompt activations, facilitate interventions that alter model responses, and achieve high concept‑detection performance while requiring far fewer construction tokens than comparable methods.
By Jessica Rumbelow
The paper presents a systematic study of combining defenses against jailbreak attacks on Large Language Models across different pipeline stages. It introduces a standardized evaluation framework that defines attack-success-rate, controls query budgets, and applies explicit fairness rules. Across 19 attacks and 15 defenses, the study finds that no single defense dominates, but carefully selected combinations can provide strong safety with minimal loss of utility, offering practical guidance for layered defense pipelines.
By Jiale Luo, Eric Han
The paper introduces OverThink, a slowdown attack that forces reasoning language models (RLMs) to produce many more reasoning tokens while still giving correct answers. By injecting decoy reasoning problems—such as Markov decision processes, language translation, or graphic comprehension—into the model’s context, attackers can dramatically increase token generation (up to 46× on SQuAD and 17× on coding agents). The study evaluates the attack on both proprietary and open-source RLMs across multiple datasets, explores multimodal and coding‑agent variants, and tests several defenses, concluding that defending against OverThink is challenging and that newer RLMs are even more vulnerable due to higher per‑token costs and increased reasoning token usage.
By Abhinav Kumar, Jaechul Roh, Ali Naseh, Marzena Karpinska, Mohit Iyyer, Amir Houmansadr, Eugene Bagdasarian
The paper investigates how restricting a language model’s view to isolated subtask contexts—via recursive language models—affects learning compared to standard chain‑of‑thought (CoT) approaches. While CoT can simulate recursive reasoning and offers similar in‑distribution performance, it tends to exploit contextual shortcuts that fail out of distribution; recursive models avoid this by isolating context, thereby preventing such shortcuts. The study shows that merely covering the correct rule is insufficient for robust reasoning beyond the training distribution.
By Chenxiao Yang, Zhiyuan Li, David McAllester, Nathan Srebro
MemeTAG introduces a dual‑objective framework for classifying harmful memes by combining keyword generation from a pretrained Vision‑Language Model with an Aggregated Tag Inference Network (ATIN) that condenses these keywords into a rich semantic embedding. The embedding is used as a target for an auxiliary reconstruction loss, encouraging deep alignment between visual and textual features. This approach, along with a three‑stage training strategy, achieves new state‑of‑the‑art results on the HarMeme, Hateful Memes Challenge, and PrideMM datasets.
By Akshit Sharma, Prashant W. Patil
The paper proposes a method for culturally aligning large language models (LLMs) using soft prompt tuning optimized via Differential Evolution (DE). Unlike traditional fine‑tuning or reinforcement learning, this approach keeps model weights frozen and requires no preference data, instead leveraging aggregated survey scores from Hofstede's Value Survey Module (VSM13). Experiments on four countries and four instruction‑tuned models show that DE‑optimized prompts reduce cultural discrepancy, improve agreement with the World Values Survey, and are preferred in blinded pairwise evaluations by LLM judges.
By Reem I. Masoud, Martin Ferianc, Philip Treleaven, Miguel Rodrigues
The paper presents an AI-based methodology that uses embedding-based similarity search and large language model classification to map Dutch job profiles to harmonised ESCO occupations. It introduces a Digital Semantic Score, which quantifies the association of job titles and skills with digital concepts relative to a non-digital reference, moving beyond keyword-based approaches. The study finds that digitalisation is unevenly distributed across occupations, with managerial, professional, and ICT roles showing the strongest digital language, while career-transition and skill analyses reveal pathway-dependent movement toward digital work and a multidimensional nature of digital capability.
By Sadegh Shahmohammadi, Xavier Pinho, Mairi Bowdler, Suhendan Adiguzel-van Zoelen, Joost van Genabeek
The paper introduces a framework to assess how linearly encoded linguistic relations are within language model embeddings. Using a constrained linear approximation on related and unrelated word pairs, the authors evaluate an extended BATS dataset across inflectional, derivational, lexicographic, and encyclopedic relations in GloVe, RoBERTa, and ModernBERT. Results show near-perfect linear encoding for inflectional and derivational relations, higher errors for lexicographic and encyclopedic relations—especially in one-to-many and many-to-many cases—and that RoBERTa and ModernBERT encode relations more linearly than GloVe.
By Vasudevan Nedumpozhimana, Fathima Thekkekara, John Kelleher
Conformal Privacy Auditing (CPA) is a distribution‑free framework that calibrates re‑identification risk for each released document against large language model (LLM)‑empowered adversaries. It outputs a conformal ambiguity set of candidate identities that is guaranteed to contain the true identity with a user‑chosen confidence level under exchangeability, along with an interpretable leakage proxy derived from the set size. CPA supports both logit‑access and sampling‑only attackers, enabling audits of both open‑source and proprietary models, and demonstrates calibrated coverage across various benchmarks and attacker configurations.
By Shuo Huang, Gholamreza Haffari, Xingliang Yuan, Ting Yu, Lizhen Qu
The paper introduces an extension of the Decoupled Vision Transformer (DC‑ViT) to handle multi‑channel imaging (MCI) data, where each channel carries a distinct semantic signal. By tokenizing each channel separately and then decoupling intra‑channel from inter‑channel updates, the model preserves channel‑specific features. The authors further enable independent per‑channel masking by solving a linear assignment between retained patches, allowing masked training without restricting visible tokens. Experiments on fluorescence microscopy, imaging mass cytometry, and satellite imaging datasets demonstrate that this approach outperforms the strongest Multi‑Channel Vision Transformer baselines on both classification and dense‑prediction tasks.
By Umar Marikkar, Sameed Husain, Muhammad Awais, Sara Atito
The paper introduces SceneLM, a vision‑language model that maintains an open‑vocabulary 3D scene map using only a structured text list of objects as persistent memory. The model updates this textual map by adding, editing, and removing objects for each input image, learning the process through supervision tasks and an automatic annotation pipeline. Evaluations on language‑grounded retrieval and localization benchmarks show competitive performance with traditional mapping systems while producing a 6‑12× more compact representation, and the model can run online on an edge device such as a quadruped robot.
By Adam Lilja, Fabio H\"ubel, Siming He, Junsheng Fu, Claire Tomlin, Lars Hammarstrand, Jitendra Malik, Jonas Frey, Marco Pavone
VeriFuse is a bounded arbitration framework that integrates vision‑language models (VLMs) into vehicle‑infrastructure cooperative 3D perception. Each agent first generates independent detections, then VeriFuse creates a unified candidate pool of geometric proposals and cross‑source hypotheses. A frozen VLM selects among three actions—SELECT, REFINE, or REJECT—to resolve ambiguity and produce final 3D detections, achieving strong AP50/AP70 scores on the DAIR‑V2X dataset while keeping vehicle‑side BEV AP50 drop minimal under delay.
By Hongyi Lin, Yiyao Liu, Qi Kang, Heye Huang, Yang Liu, Haris Koutsopoulos, Jinhua Zhao
TatBLiMP is the first benchmark of linguistic minimal pairs for the Tatar language, covering 16 morphosyntactic phenomena across 1,248 sentence pairs that differ by a single morpheme. Each pair contains one grammatical and one ungrammatical sentence, with the ungrammatical version generated by a deterministic perturbation and ratified by a native speaker. The benchmark evaluates models by comparing their assigned probabilities, allowing assessment without text generation or parsing, and tracks performance across from-scratch, cross‑lingual, and multilingual large language models.
By Ilshat Saetov, Dmitry Gaynullin
The paper investigates how Large Language Models can help detect misconfigurations in Kubernetes, the dominant platform for orchestrating containerized applications in cloud‑native environments. It presents a taxonomy of common misconfiguration types, evaluates existing detection tools, and analyzes which Kubernetes objects are most vulnerable and how severe the issues can be. The study demonstrates that advanced machine learning, particularly LLMs, can offer new insights and improve the effectiveness of misconfiguration detection methods.
By Mostafa Anouar Ghorab, Mohamed Aymen Saied
The paper shows that multi‑hop retrieval failures cluster in predictable subpopulations and formalizes this with two theoretical results: (1) confident‑failure reduction is possible only when retrieval features carry mutual information about success, and (2) no single ANN score feature dominates across all failure regimes. Building on these insights, the authors introduce RegimeAbstain, which computes a Retrieval Confidence Score (RCS) from up to nine query‑ANN structural features and uses it to calibrate an abstention policy. Across three benchmarks and two retrieval architectures, RCS achieves the best or co‑best AUC‑AC and significantly reduces the Confident‑Wrong‑Answer Rate, demonstrating its effectiveness and domain‑agnostic applicability.
By Andre Bacellar
ArenaFlow is a hierarchical credit propagation framework designed to improve reinforcement learning for open-ended agent tasks. It uses tournament-based relative ranking to generate trajectory-level rewards and structured reflective evaluation to identify pivotal success steps, reusable strategy skills, and skill usage attribution. The framework propagates advantages to high-confidence steps and maintains a global skill memory, enabling more targeted optimization and reusable skill priors for future exploration.
By Qiang Zhang, Ruixue Ding, Fanrui Zhang, Xi Chen, Boli Chen, Shihang Wang, Yinfeng Huang, Yi Zheng, Pengjun Xie, Kaipeng Zhang, Jiawei Liu, Zheng-Jun Zha