Yuda Song, Alvin Zhang, Fahim Tajwar et al.
Foundation models are typically trained in stages: pretraining, midtraining, supervised fine-tuning (SFT), and reinforcement learning (RL), with changes in both data and…Foundation models are typically trained in stages: pretraining, midtraining, supervised fine-tuning (SFT), and reinforcement learning (RL), with changes in both data and learning algorithms.
Guy Zamir, Victor Boone, Matthew Zurek et al.
In this paper, we study the regret minimization problem in episodic tabular MDPs. Our algorithm achieves the first expected regret bound scaling as √H³ S A K for…In this paper, we study the regret minimization problem in episodic tabular MDPs. Our algorithm achieves the first expected regret bound scaling as √H³ S A K for sufficiently large K, where H is the episode length, K is the number…
Lie Li, Wen Li, Junxiao Shen et al.
Unlike autoregressive large language models (AR-LLMs), diffusion LLMs (dLLMs) generate text by iteratively denoising many tokens in parallel, offering a promising path to…Unlike autoregressive large language models (AR-LLMs), diffusion LLMs (dLLMs) generate text by iteratively denoising many tokens in parallel, offering a promising path to efficient inference.
Yinuo Liu
Reinforcement learning with verifiable rewards (RLVR) makes a language model solve more problems on the first try and fewer when given many tries: it raises pass@1 while…Reinforcement learning with verifiable rewards (RLVR) makes a language model solve more problems on the first try and fewer when given many tries: it raises pass@1 while the base model overtakes it at large sampling budget k.
Noel Loo, Neehal Tumma, T. Konstantin Rusch et al.
Test-time training (TTT) layers replace the linear fast-weight memory of modern recurrent architectures with a small neural network trained online by gradient descent.Test-time training (TTT) layers replace the linear fast-weight memory of modern recurrent architectures with a small neural network trained online by gradient descent.
Qiwei Di, Quanquan Gu
We study adversarial linear bandits with strong corruption, in which an adaptive adversary chooses the reward vectors and a strong adversary corrupts each observed reward…We study adversarial linear bandits with strong corruption, in which an adaptive adversary chooses the reward vectors and a strong adversary corrupts each observed reward after seeing the learner's action, subject to a known total…
A proof showed that random search with cooling and averaging could recover a hidden vector above the same signal threshold as established global methods.A proof showed that random search with cooling and averaging could recover a hidden vector above the same signal threshold as established global methods.
Xu Zhang, Ziqing Hu, Shuo Han et al.
Multimodal large language models (MLLMs) are vulnerable to visual jailbreak attacks that bypass safety safeguards.Multimodal large language models (MLLMs) are vulnerable to visual jailbreak attacks that bypass safety safeguards. Existing methods typically optimize isolated adversarial examples, limiting attack diversity, transferability, and…
Shikhar Shiromani
Post-training moves a language model's weights by about a thousandth of their norm and changes its behavior sharply; pretraining moves them by a third and buys a slow…Post-training moves a language model's weights by about a thousandth of their norm and changes its behavior sharply; pretraining moves them by a third and buys a slow drift.
Igor Shilov, Xiaoxue Yang, Alexander Panfilov et al.
We introduce Heads of Security (HoS), a prompt injection defense that enforces instruction-data separation via a small set of learned per-head attention gates.We introduce Heads of Security (HoS), a prompt injection defense that enforces instruction-data separation via a small set of learned per-head attention gates.
Pranjal Aggarwal, Sean Welleck
AI agents have made substantial progress in domains where they have a non-trivial amount of time to deliberate between actions, such as mathematics, software engineering…AI agents have made substantial progress in domains where they have a non-trivial amount of time to deliberate between actions, such as mathematics, software engineering, and performing tasks on desktop applications.
Jiafeng Lin, Yutian Zhang, Huan Zhang et al.
Predictive maintenance is vital for commercial aviation, yet progress remains hindered by the lack of public, large-scale real-world datasets.Predictive maintenance is vital for commercial aviation, yet progress remains hindered by the lack of public, large-scale real-world datasets.
Peize Ding
Label-free reinforcement learning with verifiable rewards (RLVR) rewards a language model for agreeing with its own majority vote.Label-free reinforcement learning with verifiable rewards (RLVR) rewards a language model for agreeing with its own majority vote. It improves reasoning at first but eventually collapses: the self-assessed reward goes to one while…
Elisabeth Fink
Self-attention mechanisms have established the Transformer as the dominant backbone across natural language processing, source code intelligence, and automated theorem…Self-attention mechanisms have established the Transformer as the dominant backbone across natural language processing, source code intelligence, and automated theorem proving.
Saket Reddy, Amro Abbas, Anshuman Suri et al.
Fine-tuning Large Language Models (LLMs) on instruction data can quietly erode their safety, even when no training example looks harmful.Fine-tuning Large Language Models (LLMs) on instruction data can quietly erode their safety, even when no training example looks harmful. Because this safety-degrading data looks benign, standard content filters struggle to catch…
Hubert Baniecki, Przemyslaw Biecek, Martin Pawelczyk
Language model performance depends strongly on the mixture of domains in the pretraining data corpus. Popular data mixing methods select these domain weights by training…Language model performance depends strongly on the mixture of domains in the pretraining data corpus. Popular data mixing methods select these domain weights by training many small proxy models on random mixes and then fitting a…
Geunhyoung Kim, Sung Ju Hwang
Mixture-of-Experts (MoE) models are bottlenecked by loading routed-expert weights from memory during interactive decoding at small batch sizes.Mixture-of-Experts (MoE) models are bottlenecked by loading routed-expert weights from memory during interactive decoding at small batch sizes.
Henrique Schechter Vera, Sahil Dua, Jovan Andonov et al.
We present EmbeddingGemma 2, an open multimodal embedding model which natively represents text, images, audio, video, and combinations thereof in a unified space.We present EmbeddingGemma 2, an open multimodal embedding model which natively represents text, images, audio, video, and combinations thereof in a unified space.
Vincess Dongmo, Astride Melvin Fokam Ninyim
Multi-relational graph detectors learn from clean data how much to trust each relation, but an adversary pays a different price to manipulate each relation and each…Multi-relational graph detectors learn from clean data how much to trust each relation, but an adversary pays a different price to manipulate each relation and each action; taking such costs as given, we ask when using a relation…
Giuseppe Curci, Stefano Simonazzi, Andrea Molinari et al.
Direct Preference Optimization (DPO) typically relies on preference pairs generated from diverse model pools and ranked by an external LLM judge.Direct Preference Optimization (DPO) typically relies on preference pairs generated from diverse model pools and ranked by an external LLM judge.
Pedro Carvalho Brom
Typed decision models return a probability distribution over options supplied by the caller, and software tends to act on that number as if it were a guarantee.Typed decision models return a probability distribution over options supplied by the caller, and software tends to act on that number as if it were a guarantee.
Skander Moalla, Ahmad Rammal, Niket Patel et al.
While progress in informal mathematics has come largely from scaling the chain-of-thought and agentic reasoning of frontier language models, formal theorem proving has…While progress in informal mathematics has come largely from scaling the chain-of-thought and agentic reasoning of frontier language models, formal theorem proving has increasingly relied on human-designed workflows around these…
Cheng Gui, Xiang Zheng, Yutao Wu et al.
Science agents increasingly package domain expertise as reusable skills that plan analyses, select tools, and execute workflows.Science agents increasingly package domain expertise as reusable skills that plan analyses, select tools, and execute workflows. This modularity creates a confidentiality risk: the interactions that deliver a skill's capability…
Junu Choi, Youngjoon Hong
High-resolution probabilistic weather forecasting remains computationally demanding. Here we present Astra, a 0.25° ensemble model trained on eight H200 GPUs.High-resolution probabilistic weather forecasting remains computationally demanding. Here we present Astra, a 0.25° ensemble model trained on eight H200 GPUs.
Salim Khazem
Embedding-based verification often calibrates a false-accept threshold from all impostor pairs among m identities, although these pairs share identities.Embedding-based verification often calibrates a false-accept threshold from all impostor pairs among m identities, although these pairs share identities.
Jiamu Zhang, Hanjie Chen
Large language models (LLMs) now serve as the backbone of agents that reason over many turns, call tools, and pursue long-horizon tasks.Large language models (LLMs) now serve as the backbone of agents that reason over many turns, call tools, and pursue long-horizon tasks. To avoid recomputation, a model stores the attention keys and values of all past tokens in a…
Naveen Madiraju
Ask a model "how many settlements did the Normans control?" and it answers that the passage does not say.Ask a model "how many settlements did the Normans control?" and it answers that the passage does not say. Tell it "given that the Normans controlled 12 settlements of 400 inhabitants, what is the total?" and it computes 4,800.
Ari Blondal, Pooya Hatami, Chavdar Lalov et al.
We study the list replicability number in the realizable PAC learning framework and prove that every concept class H over a countable domain with Littlestone dimension d…We study the list replicability number in the realizable PAC learning framework and prove that every concept class H over a countable domain with Littlestone dimension d and dual VC dimension v^(*) satisfies [ LR(H)≤ 2^{O(dmin{log…
Sajad Ashkezari, Shai Ben-David
We determine the optimal regret in agnostic multiclass online learning with arbitrary, possibly infinite, label spaces.We determine the optimal regret in agnostic multiclass online learning with arbitrary, possibly infinite, label spaces. For every hypothesis class of multiclass Littlestone dimension d<∞, we prove a regret upper bound of…
Awni Altabaa, Tobias Schnabel, Kiran Tomlinson
Several formal theories have been developed to characterize the capabilities and limitations of the transformer architecture, each employing different abstractions that…Several formal theories have been developed to characterize the capabilities and limitations of the transformer architecture, each employing different abstractions that capture different features of the architecture.
Ivan Viakhirev, Kirill Borodin, Grach Mkrtchian
Reporting the best of M noisy, unbiased scores overstates the winner's value by up to c√V, where V is the total noise variance and the sharp constant c depends on how the…Reporting the best of M noisy, unbiased scores overstates the winner's value by up to c√V, where V is the total noise variance and the sharp constant c depends on how the noise is dependent.
Zong-Han Bai
Covariate-aware time series foundation models (TSFMs) accept a future path for a control variable and return a forecast, inviting what-if queries.Covariate-aware time series foundation models (TSFMs) accept a future path for a control variable and return a forecast, inviting what-if queries. The query is interventional; the models are trained on observational data.
Nicholas Roberts, Ahmed Ahmed, Suhas Kotha et al.
Synthetic rewrites, where a language model rewrites a natural document e.g. into a target style, are now a standard ingredient of pretraining corpora.Synthetic rewrites, where a language model rewrites a natural document e.g. into a target style, are now a standard ingredient of pretraining corpora.
Gautam Kamath, Mahbod Majid, Ankur Moitra et al.
We study learning a discrete distribution under person-level differential privacy in the batched setting. Given n users, each contributing m i.i.d.We study learning a discrete distribution under person-level differential privacy in the batched setting. Given n users, each contributing m i.i.d.
Andrea Giuseppe Di Francesco, Rishabh Jain, Franco Maria Nardini et al.
Recent methods for injecting graph structure into LLM prompts include graph soft prompting, in which a graph neural network (GNN) learns graph-specific tokens that…Recent methods for injecting graph structure into LLM prompts include graph soft prompting, in which a graph neural network (GNN) learns graph-specific tokens that capture relational information among entities in textual prompts.
Zhaoyi Zhou, Akshayaa Magesh, Jeongyeol Kwon et al.
As LLM agents operate over corpora of more than ten thousand tools, selecting the right tool for a complex multi-step user instruction becomes a central bottleneck.As LLM agents operate over corpora of more than ten thousand tools, selecting the right tool for a complex multi-step user instruction becomes a central bottleneck.
Jiacheng Liu, Yaxin Luo, Xiaohan Zhao et al.
Scaling embedding capacity alongside Transformer computation is an emerging trend in language models. This trend makes high compression rates increasingly useful.Scaling embedding capacity alongside Transformer computation is an emerging trend in language models. This trend makes high compression rates increasingly useful.
Khouloud Saadi, Di Wang
Chemical AI workflows routinely represent the same molecule through multiple co-referring identifiers, including trivial names, IUPAC names, and SMILES, implicitly…Chemical AI workflows routinely represent the same molecule through multiple co-referring identifiers, including trivial names, IUPAC names, and SMILES, implicitly assuming that these representations provide interchangeable access…
Zhizhuo Kou, Sizai Hou, Long Chen et al.
Reinforcement learning (RL) is credited with teaching reasoning models to write {wait} before reconsidering a step, yet large RL accuracy gains do not require the word.Reinforcement learning (RL) is credited with teaching reasoning models to write {wait} before reconsidering a step, yet large RL accuracy gains do not require the word.
Haolin Liu, Chen-Yu Wei, Hanbin Zhou
We study episodic linear Markov decision processes with unknown transitions, adversarial losses, and bandit feedback.We study episodic linear Markov decision processes with unknown transitions, adversarial losses, and bandit feedback. The state-of-the-art result for this setting (Liu et al., 2024) either achieves √K regret with a computationally…