The paper examines privacy risks in large language models (LLMs) through a unified threat model and systematic ablation study of four attacks: Membership Inference, Attribute Inference, Data Extraction, and Backdoor Attacks. It finds that membership inference is highly sensitive to dataset and attack strategy, attribute inference depends on model and training characteristics, data extraction grows with model size and repeated training data, and backdoor attacks consistently succeed. The authors provide practical recommendations for privacy‑aware LLM deployment, such as prioritizing membership leakage checks in retrieval‑augmented systems, reducing repeated personal data in training, monitoring attribute inference, and protecting fine‑tuning pipelines from backdoors.
By Karima Makhlouf, Lamiaa Basyoni, Syed Khaderi, Gabriel Marquez, Peter Sotomango, Mahmoud Awawdah, Sami Zhioua
HyCo is a hybrid drafting method that assigns depth to a server model’s multi‑token prediction head for execution on a mobile NPU, while alternative branches are handled by a small language model in a single pass. By returning only the hidden state, input embedding, and key‑value entries, HyCo reduces server GPU time per token by 38.6% on Spec‑Bench and 29.2% on 32K‑token summarization, and achieves faster delivery and higher savings than an SLM‑only drafter. The approach also incorporates quantization, a restricted vocabulary, a 1K‑token window, and an ε‑threshold rule to relax exact sampling distribution preservation.
By Hanbin Bae, Heejin Choi, Joohyung Lee, Byoung Jin Choi, Sangjun Park, Hoon-Young Cho
The paper introduces Lexical Reference Frames (LexFrame) as a method to interpret language model residual streams by treating each hidden state as having a native reference frame induced by subsequent layers. Unlike fixed lenses or point‑cloud approaches, LexFrame reads a state through the model’s remaining layers using a set number of attention rounds, achieving high coherence (0.97) and accurate state reconstruction (within 0.16 nats). The approach also provides faithful attribution, token‑level steering, and potential early‑exit savings, making every hidden state readable, attributable, and steerable.
By Jian Gu, Hongyu Zhang, Chunyang Chen, Aldeida Aleti
The study evaluates 13 Geospatial Foundation Models (GFMs) for estimating above‑ground biomass (AGB) from satellite imagery, comparing them to a fully supervised state‑of‑the‑art baseline trained on the AGBD dataset. Most GFMs perform worse than the baseline, but TESSERA and AlphaEarth Foundations (AEF) features combined with specific prediction heads achieve superior results, with the best model attaining an RMSE of 50.79 Mg/ha. This AEF‑based model also generalizes better to unseen regions, reduces RMSE by up to 25% in excluded areas, and matches the ESA CCI biomass product when trained on limited regional labels.
By Ghjulia Sialelli, Linus Scheibenreif, Jan Dirk Wegner, Konrad Schindler
The paper introduces Identifier-Free Code Embedding Models designed to improve function association for binary reverse engineering. It formalizes the bidirectional association problem between source code and decompiled, stripped code, and evaluates embedding models on this task. By fine‑tuning a Qwen3‑Embedding model with contrastive learning, the authors achieve significant performance gains over existing baselines and demonstrate generalization to a constant‑algorithm association task.
By Eric Wolos, Michael Doyle
PiCo is a post‑hoc framework that explains Geo‑foundation model predictions by identifying concept vectors in the model’s embedding space. It generates synthetic multispectral satellite data using a 3D radiative transfer model and a game engine to reveal physical sensitivities, then projects embeddings onto these vectors to quantify concept importance. Evaluations on three GFMs show that PiCo can recover most of the full‑embedding accuracy and uncover how different physical concepts influence predictions for tasks such as ground‑cover estimation and forest classification.
By Yihang She, Andrew Blake, Jianbo Qi, Srinivasan Keshav
The paper introduces Locked Image Multiple Instance PreTraining (LIMIT), a method that aligns slide‑level pathology reports with patch‑level visual features by freezing a vision encoder and fine‑tuning a text encoder. LIMIT uses report embeddings as cross‑attention queries over bags of patch features, optimizing slide representations via InfoNCE. Applying LIMIT, the authors create CONCH‑Z, which achieves state‑of‑the‑art zero‑shot performance on 22 pathology tasks, outperforming vision‑language encoders trained on curated caption datasets and narrowing the gap with larger slide‑level foundation models.
By Daniel Shao, Dongmin Bang, Luca L. Weishaupt, Sophia J. Wagner, Nic G. Reitsam, Ming Y. Lu, Long P. Le, Richard J. Chen, Faisal Mahmood
The paper introduces a domain‑specific debugging agent for repairing near‑miss accelerator kernels, arguing that debugging is more efficient than regenerating kernels from scratch. It tackles knowledge scarcity, integrity, and cost control through pattern retrieval, anti‑cheat detection, and convergence safeguards. Empirical results show the debug agent achieves a 66.7% Pass@1 rate and uses 92.8% fewer tokens per success compared to traditional regeneration methods.
By Yansong Sun, Shenxiu Wu, Siyuan Chen, Runlin Hou, Junhao Qiu, Junming Cao, Shudi Shao, Zhichao Lu, Qingfu Zhang
Pixels to Tokens introduces TEAL, a token‑space efficient active learning framework that operates at the pixel level for semantic segmentation using a frozen Vision Transformer backbone. TEAL selects diverse candidate pixels via a two‑level MaxHerding strategy on multi‑layer token descriptors and refines uncertainty on a finer decoder grid, avoiding interpolation artifacts. Across multiple datasets, TEAL outperforms prior baselines under extreme label scarcity, achieving up to +24.36 mIoU with a single labeled pixel per image and reaching 97% of fully supervised performance with only ten pixels.
By Utku Cicek, Mahip Singh, Ziyao Shang, Kimathi Kaai, C Thomas, Pedro Pablo Guerrero Vela, Alexander Wong, Sirisha Rambhatla
ARC‑SORT is a real‑time, training‑free multi‑object tracker designed for fast, direction‑changing athletes in sports broadcasts. It uses an Interacting Multiple Model Kalman filter to generate per‑track uncertainty‑adaptive search regions (EIoU) and a lightweight corrective identity swap mechanism for track crossings, achieving high HOTA scores on SportsMOT and DanceTrack. The tracker also drops the motion gate for long‑absent players, linking tracklets solely by appearance, and runs at ~50 fps on an RTX 3060 GPU.
By Girish Kumar Adari
AbridgeKV compresses long chain‑of‑thought outputs by encoding completed text blocks into the KV cache of specially trained pause tokens, freeing the original cache space. The method uses pause tokens positioned within the block, allocates cache budget per layer based on entry value, and compresses large blocks. Experiments on 446 competition problems show that with only a quarter of the original KV cache, AbridgeKV matches the accuracy and output length of a full‑KV model, while enabling faster inference and a 3.5× smaller cache per sequence.
By Jack Cai, Lijie Yang, Vasilis Kontonis, Dimitris Papailiopoulos, Shivam Garg
RAG‑RST is a retrieval‑augmented framework that supplies a formal discourse schema as in‑context demonstrations for Rhetorical Structure Theory parsing. By retrieving exemplar EDU pairs, it improves nuclearity F1 from 0.794 to 0.846 and relation F1 from 0.738 to 0.788 without any training. Two configurations—precomputed and dynamic—both outperform existing RST parsers, with the dynamic variant consistently strongest, and the approach shows that a formal linguistic schema can be provided to a foundation model as a demonstration rather than encoded in its parameters.
By Mohammadreza Sediqin, Shlomo Engelson Argamon
RecSoup introduces a lightweight two‑stage method for Controllable Multi‑Objective Recommendation (CMOR) that merges multiple objective‑specific models without retraining. The approach first prunes parameters to reduce cross‑objective interference, then adaptively merges them using a learnable preference‑to‑coefficient mapping, enabling dynamic trade‑off control at inference time. Experiments show RecSoup achieves superior Pareto front quality and 165–228× faster inference than parameter‑generation baselines.
By Zidong Jia, Guofu Xie, Xiao Zhang
Memorilla is a memory system that encodes an entire document collection into reusable embeddings and retrieves query‑specific information as a fixed set of memory tokens for a frozen large language model. Using only 16 memory tokens, it outperforms baselines such as Mem0 and MemoRAG on most benchmarks, achieving higher accuracy on FactKG while drastically reducing token usage. In agentic settings, Memorilla helps prevent wandering and boosts success rates in a TextWorld environment, demonstrating its utility for continual learning and long‑context interactions.
By Adibvafa Fallahpour, Parsa Idehpour, Vignesh Kothapalli, Nikita Mounier, Shirley Wu, Shayan Pardis, Jure Leskovec
The paper introduces Future‑L1, an interleaved latent visual reasoning framework that alternates between language tokens and continuous latent visual spans during autoregressive decoding for video event prediction. By training on a curated dataset (Future‑L1‑50K) and optimizing latent trajectories with a latent‑aware RL objective (LA‑DAPO), Future‑L1 achieves state‑of‑the‑art results on FutureBench and TwiFF‑Bench, outperforming prior models by significant margins. The work demonstrates that preserving intermediate visual semantics in latent space, rather than verbalizing every reasoning step, leads to more accurate and visually grounded future predictions.
By Tianxiang Jiang, Linquan Wu, Sheng Xia, Songze Li, Ziang Yan, Haoyu Yang, Yu Qiao, Yi Wang
The paper investigates three ways to model permutation generation with language models: hard-assignment prediction over shuffled candidate-index sequences, soft-assignment prediction over doubly stochastic matrices, and direct ordering-score prediction followed by sorting. Each approach imposes permutation constraints at different stages of prediction and decoding, leading to distinct training objectives and sampling procedures. Experiments on Amazon recommendation reranking show that hard- and soft-assignment variants match a constrained masked-diffusion baseline in Recall@1, while direct ordering-score prediction guarantees valid permutations without additional assignment steps, though further function evaluations do not consistently improve ranking quality.
By Zhe Xu, Mingfu Liang, Kavosh Asadi, Jean-Pascal Billaud, Tianyi Chen, Yanhui Guo, Christina Huang, Baekjin Kim, Jimmy Kim, Yufei Li, Daqi Sun, Jiachen Sun, Dian Yu, Zaiwei Zhang, Kaushik Rangadurai, Frank Shyu, Sandeep Pandey, Xi Liu, Luke Simon
EigDiff proposes using eigensamples to train diffusion models on much smaller datasets while maintaining performance. By computing latent eigenvectors via PCA in a pre‑trained CLIP space, the method selects a representative training set that includes exemplars close to these eigenvectors and diverse images. Experiments show that EigDiff can reduce training data size by up to 14× and accelerate subset selection by up to 1607× with only marginal performance loss.
By Andrei Jarca, Vlad Hondru, Florinel-Alin Croitoru, Radu Tudor Ionescu
The paper introduces Self‑Programmed Execution (SPE), where a language model’s own completion serves as the orchestrator for state transitions in an agent, eliminating fixed harness policies. SPE requires a specialized language, Spell, a Lisp dialect that allows programs to edit and re‑evaluate themselves while preventing replay of side effects. Experiments show that current models can learn Spell in‑context and perform complex agentic tasks, demonstrating that a language model can act as an autonomous agent without external orchestration.
By Luke J. O'Connor
The paper introduces FPCD, a framework that stabilizes the modality gap in vision‑language models during continual learning. It combines Fast Null‑space Projection, which projects gradients onto approximate null spaces to reduce interference, with Contrastive Distillation that aligns image‑to‑text and text‑to‑image similarity distributions using a small reference set. On the 11‑task MTIL benchmark, FPCD achieves strong performance, maintaining stable modality gaps and superior zero‑shot retrieval on COCO Captions and Flickr30k.
By Tiantian Peng, Yuyang Liu, Shuo Yang, Qiuhe Hong, YongHong Tian
The paper investigates the use of partial Rotary Positional Embedding (RoPE) in transformer models, showing that applying RoPE to only about 10% of hidden dimensions can save up to ten times the memory of full RoPE while achieving comparable final loss. The authors systematically analyze training dynamics and convergence across various architectures, model sizes, sequence lengths, and datasets, finding consistent performance trends. They also note that completely omitting positional encoding (NoPE) can cause loss spikes, which are mitigated by minimal RoPE usage or QK‑Norm application.
By Mohammad Aflah Khan, Krishna P. Gummadi, Manish Gupta, Abhilasha Ravichander