GyroNovo is a new framework for de novo peptide sequencing that improves fragment imputation by guiding the process with decoder errors observed during training. It introduces mass-aware attention using rotary embeddings to encode pairwise mass differences between spectral peaks, and creates easy and hard augmented views of spectra to train the decoder under varying corruption levels. Experiments on NovoBench demonstrate significant gains, with about 9 percentage points higher peptide-level precision and 7 percentage points higher amino-acid-level precision compared to the state-of-the-art baseline.
By Abdellah El Mekki, Laks V. S. Lakshmanan, Muhammad Abdul-Mageed
The paper discusses position‑independent KV cache reuse, a technique designed to cut latency in retrieval‑augmented generation by reusing chunk‑level KV caches across prompts. It argues that current evaluation methods overstate the accuracy of such reuse because they do not accurately capture the loss of accuracy, and that existing datasets lack the necessary reuse dynamics for thorough testing. To remedy this, the authors propose a new evaluation methodology that unambiguously measures accuracy loss and introduce Boxoffice, a tool that programmatically creates datasets with challenging KV cache reuse patterns.
By Samuel Cestola, Tianxiang Xia, Pengfei Zheng, Weiyan Zheng, Bo Wang, Yi Zhao, Diego Didona
The paper introduces TRACE, a fine‑tuning framework for Retrieval‑Augmented Generation (RAG) that addresses conflicts between retrieved knowledge and a model’s internal knowledge. TRACE uses multi‑agent debate traces to identify correct and incorrect candidates and answer‑shift patterns, providing fine‑grained supervision for reliable knowledge‑source selection. It also incorporates an answer‑completeness regularization mechanism to prevent empty, overly short, or prematurely terminated responses, thereby improving robustness against misleading retrieved content and enhancing answer quality.
By Zhengchen Huang, Yundong Sun, Minrui Song, Shuanglong Yao, Ye Liu, Ji Chen, Xing Wang
The study evaluates whether a portfolio of compact, semantically named descriptor blocks can match the performance of a 2048‑dimensional CheMeleon embedding in low‑data molecular assays. Using a fixed 11‑dimensional physicochemical base and greedily adding provenance‑screened blocks, the portfolio achieves a mean test AUC of 0.762 across nine ADME/Tox assays, comparable to CheMeleon’s 0.764 and better than Mordred’s 0.756. The results meet a predeclared pooled parity threshold but not all per‑assay thresholds, and further analysis confirms the competitiveness of the auditable representation while highlighting unresolved assay‑level differences.
By Yiqi Yao, Miquel Duran-Frigola
The paper introduces POROS, a framework that uses peer‑grounded counterfactual explanations to generate incremental behavioral steps for chronic disease management. POROS builds a directed acyclic graph of patient states where each edge represents a behavior change that peers have successfully achieved and that improves health outcomes. In two diabetes cohorts, POROS reduces the required improvement per step from over 25–30 percentage points to about 5–6, while most multi‑hop paths involve cross‑patient comparisons.
By Saman Khamesian, Hassan Ghasemzadeh
The paper studies when joint-embedding predictive architectures (JEPAs) can recover underlying causal states from high‑dimensional observations. It introduces a latent variable model where observations arise from causal states with action‑conditioned dynamics, and proposes an information‑theoretic objective that maximizes conditional likelihood while preserving state entropy. The authors prove identifiability conditions—particularly sufficient action‑induced variation—and instantiate the objective as an action‑modulated Gaussian additive‑noise model (A‑JEPA), demonstrating theoretical and empirical success in synthetic and visual benchmarks.
By Yuhang Liu, Zhuo Huang, Javen Qinfeng Shi
The paper describes a hybrid search system developed at Target that combines lexical and vector search to improve retail product search. It details data processing, embedding training, precision control, multi‑channel result fusion—specifically weighted interleaving—and performance optimizations for low latency. The system achieved measurable gains in click‑through rate, order conversion, and demand per visitor while reducing zero‑result searches.
By Darshan Sonagara, Qujiaheng Zhang, Ankit Singh, Alex Li
The paper introduces Distribution‑Conditioned Transport (DCT), a framework that learns transport maps conditioned on embeddings of source and target distributions, allowing generalization to unseen distribution pairs. DCT supports semi‑supervised learning for distributional forecasting by leveraging distributions observed at only one condition. It is agnostic to the transport mechanism and is demonstrated on synthetic benchmarks and four biological applications, including batch effect transfer in single‑cell genomics and modeling T‑cell receptor sequence evolution.
By Nic Fishman, Gokul Gowri, Paolo L. B. Fischer, Marinka Zitnik, Omar Abudayyeh, Jonathan Gootenberg
The paper introduces a low-cost method for detecting hallucinations in large language models by treating the model as a black-box dynamical system. It projects responses into a high-dimensional manifold, models the latent state-space dynamics with Koopman operator theory, and uses differential residual scores from transition operators to distinguish factual from hallucinated outputs. The approach requires only a single-sample pass and shows state-of-the-art performance across three benchmarks with reduced resource overhead.
By Dan Wilson, Mohamed Akrout
The paper introduces the Geometric Simplicial Weisfeiler–Lehman (GSWL) test, which extends the classic WL and SWL tests by incorporating vertex coordinates into color refinement for geometric simplicial complexes. It demonstrates that geometry‑aware simplicial message passing schemes are bounded by GSWL in expressivity and can match GSWL’s discriminating power on any fixed finite family of complexes. By combining GSWL with the Euler Characteristic Transform, the authors provide a complete invariant and an approximation framework, validated through experiments that reveal a clear hierarchy from combinatorial to geometry‑aware models.
By Elena Xinyi Wang, Bastian Rieck
EAServe introduces an encode-aware disaggregated serving framework for multimodal large language models (MLLMs), restructuring the traditional Prefill-Decode pipeline into a three-stage Encode-Prefill-Decode (EPD) system. By treating Encode as the control point, EAServe coordinates load‑adaptive micro‑batching, rate‑controlled offloading to prefill workers, and dynamic SM partitioning to balance GPU utilization across stages. Its Hybrid Auto Selection (HAS) layer optimizes GPU allocation, encode batch size, and offload ratio using capacity profiling and Bayesian optimization, achieving up to 4.3× higher goodput compared to NVIDIA Dynamo and 1.7× higher than vLLM on various MLLM architectures.
By Kunxiong Zhu, Zhihao Shu, Hangyu Zheng, Minghai Qin, Miao Yin, Gagan Agrawal, Wei Niu
GAUDI is a geometry-aware diffusion model designed to impute missing blocks in air‑quality time‑series data. It aligns conditional diffusion with block‑specific geometry, preserving temporal and feature processing while conditioning on visible values, masks, variable identity, and diffusion steps, but omitting absolute time‑position embeddings. On the ItalyAir dataset with 50% block missingness, GAUDI achieves an RMSE of 0.340, outperforming full‑context and local CSDI baselines.
By Xinjin Li, Yudi Xia, Calvin Chang Liu, Weiru Lin, Bojun Li, Ziwei Hong, Bolun Zhang, Jinghan Cao, Yu Ma, Tianxin Zhou
The paper reports on applying AutoResearch—a large language model that iteratively edits training scripts—to optimize embedding systems for a book recommendation pipeline at production scale. Over twelve weeks, the authors ran 220+ experiments across two representation‑learning systems, uncovering five recurring failure modes (infrastructure fragility, agent memory decay, search‑direction stagnation, iteration‑cost asymmetry, and metric fixation) that were not present in smaller settings. They propose a three‑principle scaffolding (prevent, persist, redirect) to address these failures, achieving a 1.82× lift in Recall@6, a 2.1× lift in coherence, and an autonomous text‑only fallback that expanded catalog coverage by 5.8×.
By Aparajith Chandran, Juwon Kim, Saurav Jha, Pablo Castells, Florian Hottier
BeatGraph is a self‑supervised model that represents infant ECG recordings as graphs of individual heartbeats rather than fixed‑length patches, allowing it to capture the higher heart rates and distinct waveform patterns of infants. The model uses a shared beat encoder, a Transformer for temporal ordering, and graph attention layers to produce a window embedding, which is pretrained on a large unlabeled infant ECG corpus and fine‑tuned for tasks such as sleep‑wake detection, infant‑state classification, activity‑source identification, and affect recognition. BeatGraph achieves state‑of‑the‑art performance on multiple infant‑specific benchmarks and transfers well to pediatric and adult ECG datasets, while also releasing the first public infant ECG corpus collected in diverse home and classroom settings.
By Mohammad Nur Hossain Khan, M. S. Krafczyk, Beverly G. Bolster, Nancy McElwain, Mark A. Hasegawa-Johnson, Bashima Islam
Synth-JEPA introduces a renderer‑free approach to synthesizer parameter search by learning mutually predictive audio and parameter representations from paired synthesizer data. During inference, candidate parameters are scored directly in this learned space, avoiding the need to render each candidate and shaping audio geometry through parameter correspondences. Evaluations on Surge XT and out‑of‑domain datasets show Synth‑JEPA outperforms inverse models, direct search, and learned proxy objectives, with listeners preferring its matches in 85% of pairwise tests.
By Ben Hayes, Haokun Tian, Stefan Lattner
Location encoders transform geographic coordinates into high‑dimensional embeddings for machine learning, yet it is unclear how well these embeddings capture interpretable spatial effects. This study benchmarks GeoShapley—a game‑theoretic explainer treating all location features as a single joint player—against eleven TorchSpatial encoders on a synthetic process with known coefficients, across grid, county, and global scales, with and without raw coordinates and under different training regimes. The results show that primary coefficient recovery is consistently high across encoders, while secondary coefficient recovery varies more with scale, especially at the global level, and raw‑coordinate baselines remain competitive throughout.
By Daniel Kiv, Shaowen Wang
The paper introduces a reproducible benchmark for evaluating attention mechanisms in tabular foundation models, focusing on the distinct row and column attention patterns that differ from language model attention. It compares several backends—Torch SDPA, FlashAttention variants, vLLM, and SageAttention—across realistic tabular shapes on A100, H100, and B200 GPUs, revealing that optimal backend choice varies by attention type, hardware, and model specifics. The study finds FlashAttention generally performs best, but CuDNN can outperform it for column attention on longer sequences, while SageAttention excels for large row sequences beyond 16k rows.
By Maximilian Schambach, Clemens Biehl, Sam Thelin
Tetra introduces a new Leech‑lattice based codebook that reduces the memory footprint of quantized LLMs to about 2.15 bits per weight, enabling efficient 2‑bit quantization without a massive lookup table. The method employs a 64‑state Golay trellis and a shared 16 KiB table, decoding each 24‑weight block with only six table loads and two small lookups. When applied to Qwen3 models (4B, 8B, 14B), Tetra achieves 2.70–2.73 bits per parameter, scoring 63–75 on MMLU and generating 57–114 tokens per second, while maintaining close performance to 4‑bit AWQ and outperforming llama.cpp’s IQ2_XXS on 4B.
The paper investigates whether temporal link predictors can forgo learned node representations in favor of simple statistical counts of interaction patterns. It introduces a predictor that aggregates transition and co‑occurrence counts, smooths them with destination frequencies or Kneser‑Ney continuation counts, and combines these with popularity, source history, and recency via a shared log‑linear rule. With only 9–13 learned parameters, the model achieves the best mean reciprocal rank on 7 of 16 datasets and outperforms several baselines across all evaluated datasets, demonstrating that a lightweight, count‑based approach can rival more complex neural methods.
The paper introduces PaCTS, a method that generates instance‑adaptive latent prompts—continuous embedding tokens—to provide compact contextual information for frozen time‑series foundation models (TSFMs). These prompts are constructed from instance‑specific global statistics and refined with segment‑level temporal data, enabling the model to capture both global characteristics and local temporal variations. Experiments show that PaCTS improves forecasting performance across various context lengths and model architectures, often outperforming the same backbone with double the context while reducing inference computation, and it also offers stronger improvements and better out‑of‑distribution generalization compared to weight‑space adaptation methods.