MemGuard-Alpha evaluates whether membership inference attacks (MIA) can detect memorization in large language models (LLMs) used for financial alpha signals. The study combines five MIA methods with a temporal proximity feature and a cross-model disagreement metric, then audits them across seven LLMs, 50 S&P 100 stocks, and 299,600 prompt-model pairs. Findings show that temporal proximity alone perfectly predicts in-sample status, MIA discriminative power largely stems from model scale differences, and filtering based on contamination scores does not improve risk-adjusted performance once transaction costs are considered.
By Anisha Roy, Dip Roy
The paper introduces JoLT, a training‑free compressor that jointly allocates rank and precision for key‑value (KV) cache compression in long‑context language models. JoLT treats grouped prefill caches as fourth‑order tensors, applies partial Tucker decomposition along token and feature modes, and uses a rotated low‑bit quantizer for residuals, all governed by a single Lagrangian dual under a global byte constraint. Across five models from four architecture families, JoLT achieves 2–3× compression with less than 0.2% perplexity loss, and near‑lossless retrieval accuracy on LLaMA‑3.1‑8B at 64K context up to 3× compression.
By Rahul Krishnan, Volker Schulz
The paper introduces a lightweight single‑layer sampler that allows masked diffusion language models to approximate joint sampling of multiple tokens in a single full‑model forward pass. By training the sampler to mimic exact joint sampling from a frozen diffusion model, the authors enable parallel unmasking of tokens while maintaining a close match to the true joint distribution. Experiments on Dream‑7B and Llada‑7B models show that unmasking four tokens per denoising step yields a MAUVE score of 0.87, a substantial improvement over the marginal baseline of 0.31.
By Parikshit Bansal, Sujay Sanghavi
The study examines how large language models (LLMs) like GPT‑5.2, Gemini 3 Flash, and Perplexity sonar‑pro recommend brands across five industries. Using 50 brands and 250 queries repeated five times, the authors measured brand inclusion, recommendation share, competitive vacuum, and co‑mention asymmetry, finding that most queries mention at least one brand and that vacuum prevalence remained stable between February and September 2026. The analysis shows strong cross‑date consistency in recommendation patterns and no emergent clustering of brand mentions, though co‑mention structures deviate from null expectations.
By Dmitrij \.Zatuchin
The paper presents Laxmi, a retrained version of the AIFS weather model that uses satellite-based precipitation observations instead of ERA5 reanalysis data. Laxmi achieves a 19% improvement in global probabilistic accuracy, reduces drizzle overprediction by 33%, and boosts the 95th percentile Brier skill score by 57%. In a case study of 10 Indian tropical storms, Laxmi accurately forecasted 150 mm event-total precipitation in 7 events, outperforming both the original AIFS and the leading physical model IFS.
By Julian F. Schmitt, Bertrand Delorme, Robert C. King, Yashica Patodia, Tapio Schneider, Aditi Sheshadri, Ravi Jain
The study introduces a two‑dimensional framework to audit French news headlines, distinguishing salience framing—captured by four wording devices—from selection framing—captured by outlet‑level story form and high‑charge distributions. Using a 10,000‑headline supervision set annotated by LLMs and human arbitration, the authors classify 902,111 headlines from 25 outlets (2022‑2025) and find that salience and selection diverge yet correlate, that default thresholds inflate salience estimates, and that headlines mentioning Jews, the Far‑right, and Muslims exhibit the highest salience rates. The authors release their dataset, lexicons, and code, claiming it is the largest French headline framing audit to date.
By Amr Sobhy
PTC-Bias is a two-stage framework that improves contextual biasing in speech large language models by using phoneme-level temporal competition. In the first stage, PTC Retrieval performs frame-synchronous phoneme decoding to generate a compact shortlist of bias words and their speech intervals. The second stage, PTC Correction, applies a local competition between retrieved candidates and mismatched transcript spans within those intervals, reducing near-homophone and word-segmentation errors without extra SpeechLLM passes. Experiments on LibriSpeech demonstrate consistent gains across two SpeechLLMs, with PTC-Bias reducing B-WER by up to 23.9% relative to CTC-Filter while keeping U-WER nearly unchanged.
By Zhiqi Ai, Han Cheng, Shiyi Mu, Yongjin Zhou, Shugong Xu
EAGER is a reinforcement learning framework for generative event extraction that uses fine‑grained verifiable rewards and Schema‑Contrastive Advantage Estimation to address advantage collapse in sparse binary reward settings. The reward design focuses on structural validity, extraction accuracy, groundedness, coverage, over‑generation, and span precision. Across seven benchmark datasets, EAGER outperforms prompting, supervised fine‑tuning, and previous reinforcement learning baselines, achieving a substantial improvement over the strongest prior method.
By Omar Adjali, Siting Liang, Omair Shahzad Bhatti, Daniel Sonntag
The study investigates how part‑of‑speech (PoS) categories are represented in the latent space of Sparse AutoEncoders (SAEs) applied to language models. Results show that PoS distinctions can be reliably recovered from SAE activations, but the mapping is not one‑to‑one; instead, PoS categories are supported by compact, distributed groups of sparse latents that vary across tags and remain stable on held‑out data. The findings suggest that SAEs encode morpho‑syntactic information in a distributed, category‑dependent manner rather than through isolated grammatical features.
By Alessandro Bondielli, Lucia Passaro, Serena Auriemma, Alessandro Lenci
The study re‑examines a reported advantage of a routed ternary (1.58‑bit) language model over a full‑precision transformer at 60K parameters. By running controlled experiments with multiple seeds and a fixed training recipe, the authors find that the apparent benefit largely stems from the choice of baseline model shape rather than the ternary architecture itself. While the routed model does outperform other shapes at a larger 130M‑byte budget, its advantage diminishes when a plain gated diagonal‑SSM block is used, and the ternary penalty varies with architecture and quantization details.
By Gautam Veldanda
The paper investigates how the order of generating explanations—whether a rationale is produced before or after the answer—affects vision‑language reasoning. By conducting controlled experiments on knowledge‑intensive QA, visual entailment, and compositional grounding tasks, the authors show that larger models are required for reliable rationale‑first generation, while answer‑first generation is less susceptible to format errors. The study concludes that explanation ordering, model scale, pre‑training knowledge, fine‑tuning, and task structure jointly influence prediction accuracy and reasoning faithfulness.
By Siting Liang, Luca Rippe, Omar Adjali, Daniel Sonntag
The study evaluates Arabic–Russian machine translation by comparing seven fine‑tuned neural machine translation (NMT) models with four few‑shot large language models (LLMs) on a new 15.47 million‑pair corpus split into 20k/5k/5k. Fine‑tuned NLLB‑1.3B achieves the best performance (BLEU 16.3, COMET 0.738), while the best few‑shot LLM, Aya‑Expanse 8B, scores only BLEU 1.7 on 500 sentences. Error analysis shows that low lexical overlap between Arabic and Russian is the main source of failures, and statistical tests confirm significant performance gaps between most models.
By Mullosharaf K. Arabov
MILO is a compression framework that reduces the key-value cache memory used in many-shot in-context learning by applying block-wise low-rank compression. It dynamically allocates rank budgets to blocks based on information entropy, preserving important information while aggressively compressing redundant parts. Experiments on Qwen2.5 models show up to a 50% reduction in KV cache memory and a 1.8× throughput improvement with negligible performance loss on classification and reasoning tasks.
By Youpeng Zhao, Tian Tan, Liqian Peng, Jun Wang, Alec Go
The article introduces the Artificial Societies Benchmark, a validation framework designed to evaluate synthetic populations used in research. It comprises eleven tests covering internal, construct, and external validity, drawing on twenty human data sources and comparing nine language models. The benchmark links specific research uses to the evidence required and assesses how results vary with different respondent information, revealing that strong performance in one domain does not guarantee fidelity in others.
By Edoardo Chidichimo, Min Jun Jung, Felix P. S. Wallis, James K. He
The paper introduces Reward‑Tilted On‑Policy Distillation (RT‑OPD), a method that enhances acoustic grounding in audio‑language models by using a frozen teacher to generate a reward based on the contrast between token predictions with and without audio. This reward reshapes the teacher distribution for reverse‑KL distillation, encouraging students to rely more on acoustic evidence. Experiments on two compact students across three benchmarks show that RT‑OPD consistently outperforms vanilla OPD, and a 3B model trained with RT‑OPD achieves 72.72% accuracy on the MMAU benchmark, surpassing other 3B models and rivaling larger 7B and 8B models.
By Kaiyang Li, Shaobo Han, Yue Tian, Shihao Ji
The paper introduces CodeGraph, an open‑taxonomy knowledge graph that semantically annotates source code by extracting entities such as algorithms, paradigms, design patterns, and application domains from millions of files. Using a specialized large language model and a three‑stage Wikidata linking process, the authors ground these entities in Wikidata and construct a graph with about 158 million nodes and 1 billion typed edges across 14 programming languages. A quality‑assurance protocol combining human evaluation and an LLM‑as‑a‑judge filter quantifies annotation precision.
By Federico Pennino, Andrea Gurioli, Stefano Zacchiroli, Maurizio Gabbrielli, Paolo Ferragina
The paper introduces a perception interface that separates vision from language in vision‑language models. A frozen perception stack detects objects, a deterministic semantic serializer converts the perceived state into text, and a standard text‑only large language model (LLM) answers questions. Experiments on a campus‑robot benchmark show that this serialized interface outperforms a zero‑shot VLM of the same language‑model size, especially as the language model shrinks, and that the advantage persists under paraphrase and different supervision regimes.
By Cong Xu, Ravi Sankar
The paper introduces ModularSQL, a lightweight runtime guardrail designed to detect and correct multiplicity errors—such as missing DISTINCT clauses, inflated aggregates, and Cartesian join explosions—in Text-to-SQL systems. It highlights the Multiplicity Blind Spot (MBS), where standard set-based accuracy metrics fail to capture these errors, and proposes Multiset-EX as a more comprehensive evaluation criterion. Experiments on several models show that ModularSQL can improve multiplicity-aware accuracy while adding minimal computational overhead.
By Tianxin Zhou, Ruixi Lin
LabFactory is a framework that transforms a scientific brief into an executable AI lab, integrating models, knowledge resources, tools, and a controller behind a fixed interface. The builder packages the lab in a metered workspace, and a separate host evaluates the delivered artifact on held‑out inputs, ensuring the system itself is the evaluation target. Across 28 constructions in seven scientific domains, the delivered labs surpassed reference values on all 33 subtests, demonstrating that an AI agent can fully realize a scientific brief into a working, inspectable lab.
By Jinge Wu, Hongjian Zhou, Mingde Zeng, Jiayuan Zhu, Junde Wu, Jiazhen Pan, Lei Clifton, Andrew Liu, David A. Clifton
The paper investigates cache replacement strategies for large language model (LLM) prefix reuse, analyzing production traces from two companies and testing 14 eviction algorithms in both high-bandwidth memory (HBM) and large memory-pool environments. It finds that sophisticated policies designed for traditional caches offer little advantage over simple LRU, because prefix reuse is largely driven by the regular pacing of active sessions, making recency a strong predictor. The study also highlights new challenges such as heavy-tailed session footprints and variable miss costs, and proposes a compute-savings ratio along with two offline oracles to better quantify these effects, suggesting that effective prefix-cache management should combine recency with selective quick demotion, compute-aware partial eviction, and capacity-dependent granularity.
By Yiyu Liu, Minlan Yu, Juncheng Yang