Generalist robot policies benefit from large-scale visual pretraining but often struggle with contact-rich dexterous manipulation, where force, compliance, and incipient…Generalist robot policies benefit from large-scale visual pretraining but often struggle with contact-rich dexterous manipulation, where force, compliance, and incipient slip are hard to observe from images.
The classical capacity of a noisy optical channel is the maximum rate at which information can be transmitted reliably, given a constraint on the mean input energy.The classical capacity of a noisy optical channel is the maximum rate at which information can be transmitted reliably, given a constraint on the mean input energy.
The Model Context Protocol (MCP) standardizes how AI agents invoke external tools. MCP servers expose tools through a declared contract consisting of a name, description…The Model Context Protocol (MCP) standardizes how AI agents invoke external tools. MCP servers expose tools through a declared contract consisting of a name, description, and argument schema.
We study the problem of approximately counting subgraphs under local node differential privacy (LNDP), a recently introduced model that enforces strong privacy guarantees…We study the problem of approximately counting subgraphs under local node differential privacy (LNDP), a recently introduced model that enforces strong privacy guarantees in analyses of distributed social networks.
Given a joint distribution PXY on a finite alphabet, the information bottleneck method seeks a representation T of X that is as informative as possible about Y while…Given a joint distribution PXY on a finite alphabet, the information bottleneck method seeks a representation T of X that is as informative as possible about Y while compressing X, where T and Y are conditionally independent given…
Tree edit distance (TED), the analog of string edit distance (ED) for hierarchical data, is the minimum number of node insertions, deletions, and substitutions…Tree edit distance (TED), the analog of string edit distance (ED) for hierarchical data, is the minimum number of node insertions, deletions, and substitutions transforming one labeled rooted ordered tree into another.
Step-level guards decide, before each tool call, whether an agent's next action is safe given the user's request and the trajectory so far.Step-level guards decide, before each tool call, whether an agent's next action is safe given the user's request and the trajectory so far.
Computer use has reached a turning point: the focus is expanding from GUI-only interaction to systems that complete tasks on users' computers through visual and…Computer use has reached a turning point: the focus is expanding from GUI-only interaction to systems that complete tasks on users' computers through visual and programmatic interfaces.
Current humanoid dexterous manipulation remains largely confined to end-effectors and simple grasps, falling short of humans' ability to engage the entire body by bracing…Current humanoid dexterous manipulation remains largely confined to end-effectors and simple grasps, falling short of humans' ability to engage the entire body by bracing and sliding bulky objects against the arms and torso.
LLM systems increasingly rely on groups of agents that discuss, critique, and check one another, assuming that peers will catch individual errors.LLM systems increasingly rely on groups of agents that discuss, critique, and check one another, assuming that peers will catch individual errors. We test this assumption against a minimal adversary.
In a seminal paper, Kratsch and Spinrad [SODA 2003 and SIAM Journal on Computing 2006] studied several important problems in structural graph theory that anticipated two…In a seminal paper, Kratsch and Spinrad [SODA 2003 and SIAM Journal on Computing 2006] studied several important problems in structural graph theory that anticipated two trends in fine-grained complexity, both from the upper- and…
Supervised fine-tuning datasets remain overwhelmingly English-centric and commercially restricted, as their content is frequently generated by proprietary models whose…Supervised fine-tuning datasets remain overwhelmingly English-centric and commercially restricted, as their content is frequently generated by proprietary models whose terms of service forbid training competing systems.
Multimodal large language models (MLLMs) can act unsafely when individually benign images and texts are interpreted jointly.Multimodal large language models (MLLMs) can act unsafely when individually benign images and texts are interpreted jointly.
Long-context language models support ever longer context windows, and evaluations typically treat context length as the main indicator of difficulty.Long-context language models support ever longer context windows, and evaluations typically treat context length as the main indicator of difficulty.
Discrete image generators learn the distribution of tokens produced by an image tokenizer. These token representations should be easy for generators to model, not merely…Discrete image generators learn the distribution of tokens produced by an image tokenizer. These token representations should be easy for generators to model, not merely enable faithful image reconstruction.
Language priors in MLLMs can override visual evidence, impairing visual perception. Existing benchmarks assume shared priors across models and rely solely on accuracy.Language priors in MLLMs can override visual evidence, impairing visual perception. Existing benchmarks assume shared priors across models and rely solely on accuracy.
Weight quantization below four bits shrinks memory and bandwidth at a perplexity that looks acceptable, but often at the cost of factual knowledge and task accuracy.Weight quantization below four bits shrinks memory and bandwidth at a perplexity that looks acceptable, but often at the cost of factual knowledge and task accuracy.
Existing safety benchmarks rely largely on synthetic prompts with loosely grounded taxonomies and offer little focused analysis of severe crime categories.Existing safety benchmarks rely largely on synthetic prompts with loosely grounded taxonomies and offer little focused analysis of severe crime categories.
Data agents use code to process heterogeneous files and derive analytical insights. Scaling their training faces costly environment curation, privacy and…Data agents use code to process heterogeneous files and derive analytical insights. Scaling their training faces costly environment curation, privacy and intellectual-property restrictions, and difficulty verifying open-ended…
With the rise of large language models, agentic systems are becoming increasingly autonomous, as they write code, call tools, revise prompts, and propose modifications to…With the rise of large language models, agentic systems are becoming increasingly autonomous, as they write code, call tools, revise prompts, and propose modifications to their own components.
Kolmogorov-Arnold networks (KANs) have a dimension-independent approximation exponent when the target admits a sufficiently smooth representation by univariate edge…Kolmogorov-Arnold networks (KANs) have a dimension-independent approximation exponent when the target admits a sufficiently smooth representation by univariate edge functions.
Existing deepfake detection models are built upon manipu- lated images that are tampered with either exclusively fully swapped or fully synthesised faces, which leads us…Existing deepfake detection models are built upon manipu- lated images that are tampered with either exclusively fully swapped or fully synthesised faces, which leads us to an open question: Can off-the-shelf deepfake models…
Large language model (LLM) agents now fix software, operate computers and serve customers, but their harnesses, the programs that run them, call a large LLM at every…Large language model (LLM) agents now fix software, operate computers and serve customers, but their harnesses, the programs that run them, call a large LLM at every step, which makes long tasks slow and costly.
Detailed captions transfer visual information into vision-language model supervision, but quality is easily conflated with length: long single-pass outputs still miss…Detailed captions transfer visual information into vision-language model supervision, but quality is easily conflated with length: long single-pass outputs still miss small objects, text, and spatial relations or promote uncertain…
Multimodal document retrieval-augmented generation (RAG) has focused largely on retrieval, yet its answers depend on how well a vision-language model reads the retrieved…Multimodal document retrieval-augmented generation (RAG) has focused largely on retrieval, yet its answers depend on how well a vision-language model reads the retrieved pages.
Large language models (LLMs) often make discrete decisions, such as flagging phishing or escalating to a human.Large language models (LLMs) often make discrete decisions, such as flagging phishing or escalating to a human. System-1 models, such as Jev and the OpenAI Decisions API, return probabilities over options in a single forward pass…
Latent chain-of-thought replaces written reasoning steps with continuous hidden states, giving models additional steps to compute before answering.Latent chain-of-thought replaces written reasoning steps with continuous hidden states, giving models additional steps to compute before answering. It remains unclear what computation these steps actually perform.
Dual-system vision-language-action (VLA) models pair a high-frequency fast policy with a slow, expensive module, and published trigger mechanisms for invoking it lack a…Dual-system vision-language-action (VLA) models pair a high-frequency fast policy with a slow, expensive module, and published trigger mechanisms for invoking it lack a shared objective, measured costs, and any standard separating…
On-screen text can describe what a video shows, or it can frame what the video is supposed to mean. Current benchmarks treat both alike, testing whether a model can read…On-screen text can describe what a video shows, or it can frame what the video is supposed to mean. Current benchmarks treat both alike, testing whether a model can read the text rather than whether it should trust it.
LLM-agent evaluation reports behavioral scores but leaves the identity of the deployed runtime stack unreported.LLM-agent evaluation reports behavioral scores but leaves the identity of the deployed runtime stack unreported. A tampered deployment that preserves evaluated outputs passes every output-only evaluator.
Label-free evaluation methods, such as self-consistency filtering, majority-vote aggregation, and agreement-based reliability scoring, rely on an implicit assumption that…Label-free evaluation methods, such as self-consistency filtering, majority-vote aggregation, and agreement-based reliability scoring, rely on an implicit assumption that consistency can serve as a proxy for correctness when…
We study a protocol-level reaction leakage in CKKS deployments where public coefficient-to-slot (C2S) evaluation keys are combined with a client's 1-bit validity…We study a protocol-level reaction leakage in CKKS deployments where public coefficient-to-slot (C2S) evaluation keys are combined with a client's 1-bit validity response.
AI agents now pause for approval before they send money, delete files or email outsiders, and that approval is usually treated as a second factor.AI agents now pause for approval before they send money, delete files or email outsiders, and that approval is usually treated as a second factor.
The twin goals of privacy-preserving federated learning (FL) – harnessing distributed data while protecting the parties who contribute it – demands three properties…The twin goals of privacy-preserving federated learning (FL) – harnessing distributed data while protecting the parties who contribute it – demands three properties: model privacy, data privacy, and integrity.
Recursive self-improvement, in which AI systems build stronger AI systems, is one promising route to superintelligence.Recursive self-improvement, in which AI systems build stronger AI systems, is one promising route to superintelligence. We study a weaker form that can be measured today: frontier agents post-training a small raw base model with…
Online reinforcement learning (RL) with verifiable rewards has driven rapid progress in mathematics, coding, and tool use, but scaling this success to GUI agents remains…Online reinforcement learning (RL) with verifiable rewards has driven rapid progress in mathematics, coding, and tool use, but scaling this success to GUI agents remains challenging due to the high cost of interactive environments…
Reinforcement learning with self-generated labels, such as majority vote over a model's own samples, often recovers most of the gain of training with verifiable rewards.Reinforcement learning with self-generated labels, such as majority vote over a model's own samples, often recovers most of the gain of training with verifiable rewards.
While foundation models and vision transformers (ViTs) have redefined visual representations, current active learning (AL) strategies for these backbones largely operate…While foundation models and vision transformers (ViTs) have redefined visual representations, current active learning (AL) strategies for these backbones largely operate at the patch or region level, where annotation requirements…
Speculative decoding pairs a cheap drafter with parallel verification against the target. At ultra-long context it stops paying, because a drafter that attends over the…Speculative decoding pairs a cheap drafter with parallel verification against the target. At ultra-long context it stops paying, because a drafter that attends over the full prefix inherits the target's context-dependent cost, so…
Skill-policy co-evolution, in which an agent's policy model and its skill bank improve each other during training, is a promising paradigm for self-improving agents.Skill-policy co-evolution, in which an agent's policy model and its skill bank improve each other during training, is a promising paradigm for self-improving agents.
A self-improving agent has to decide which new skills to keep, and in practice it keeps whatever earns a high reward, convinces a judge model, or compresses its library…A self-improving agent has to decide which new skills to keep, and in practice it keeps whatever earns a high reward, convinces a judge model, or compresses its library, even when the skill is wrong.
Archiwaranguprok C, Kirkland K, Poonsiriwong R et al.
The present work is concerned with how to measure, at scale, the multi-dimensional impact of generative AI models on the individuals who use them.The present work is concerned with how to measure, at scale, the multi-dimensional impact of generative AI models on the individuals who use them.
Residual context is key to unlocking the speed-quality Pareto frontier of diffusion language models (DLMs) by carrying forward predictive distributions across denoising…Residual context is key to unlocking the speed-quality Pareto frontier of diffusion language models (DLMs) by carrying forward predictive distributions across denoising steps.
Coding agents are increasingly used to optimize code, but existing benchmarks fix one accelerator, programming model or harness, so they cannot isolate what each…Coding agents are increasingly used to optimize code, but existing benchmarks fix one accelerator, programming model or harness, so they cannot isolate what each component contributes or show whether added skills and tools repay…
Schwarzschild A, Thakkar O, Ghalebikesabi S et al.
Canaries, or synthetic training records containing known secrets, are widely used to assess whether language models reveal information from their training data.Canaries, or synthetic training records containing known secrets, are widely used to assess whether language models reveal information from their training data. But how much do these measurements depend on the canaries themselves?
Speculative decoding's effectiveness is governed by two quantities: the acceptance rate of drafted tokens and the cost of producing them.Speculative decoding's effectiveness is governed by two quantities: the acceptance rate of drafted tokens and the cost of producing them. We show the two objectives can be co-optimized with algorithm-system co-design, and present…
Agents are increasingly interacting with other agents that act for different principals (people or organizations on whose behalf they work), such as a buyer's personal…Agents are increasingly interacting with other agents that act for different principals (people or organizations on whose behalf they work), such as a buyer's personal agent negotiating with an enterprise's seller agent.
Earth observation (EO) models are still built by hand. For each new task, experts design a training recipe and tune it against a benchmark.Earth observation (EO) models are still built by hand. For each new task, experts design a training recipe and tune it against a benchmark. We show that this loop can be handed to an agent.
We propose a unified spatial-density signal to score token importance for both inference phases: bipartite merge for prefill and KV eviction for decode, yielding…We propose a unified spatial-density signal to score token importance for both inference phases: bipartite merge for prefill and KV eviction for decode, yielding measurable gains in each stage.