arXiv subjects

Percy Liang

Publications and source records attributed to Percy Liang.

At least 37 records · Page 2Linked to original sources

arXiv2026

Replaying pre-training data improves fine-tuning

To obtain a language model for a target domain (e.g. math), the current paradigm is to pre-train on a vast amount of generic web text and then fine-tune on the relatively limited amount of target data. Typically, generic data is only mixed in during fine-tuning to prevent catastrophic forgetting of the generic domain. We surprisingly find that replaying the generic data during fine-tuning can actually improve performance on the (less related) target task. Concretely, in a controlled pre-training environment with 4M target tokens, 4B total tokens, and 150M parameter models, generic replay increases target data efficiency by up to $1.87\times$ for fine-tuning and $2.06\times$ for mid-training. We further analyze data schedules that introduce target data during pre-training and find that replay helps more when there is less target data present in pre-training. We demonstrate the success of replay in practice for fine-tuning 8B parameter models, improving agentic web navigation success by $4.5\%$ and Basque question-answering accuracy by $2\%$.

cs.CL↗

arXiv2026

Comparing AI Agents to Cybersecurity Professionals in Real-World Penetration Testing

We present the first comprehensive evaluation of AI agents against human cybersecurity professionals in a live enterprise environment. We evaluate ten cybersecurity professionals alongside six existing AI agents and ARTEMIS, our new agent scaffold, on a large university network consisting of ~8,000 hosts across 12 subnets. ARTEMIS is a multi-agent framework featuring dynamic prompt generation, arbitrary sub-agents, and automatic vulnerability triaging. In our comparative study, ARTEMIS placed second overall, discovering 9 valid vulnerabilities with an 82% valid submission rate and outperforming 9 of 10 human participants. While existing scaffolds such as Codex and CyAgent underperformed relative to most human participants, ARTEMIS demonstrated technical sophistication and submission quality comparable to the strongest participants. We observe that AI agents offer advantages in systematic enumeration, parallel exploitation, and cost -- certain ARTEMIS variants cost $18/hour versus $60/hour for professional penetration testers. We also identify key capability gaps: AI agents exhibit higher false-positive rates and struggle with GUI-based tasks.

cs.AI↗

arXiv2026

Beyond a Single Extractor: Re-thinking HTML-to-Text Extraction for LLM Pretraining

One of the first pre-processing steps for constructing web-scale LLM pretraining datasets involves extracting text from HTML. Despite the immense diversity of web content, existing open-source datasets predominantly apply a single fixed extractor to all webpages. In this work, we investigate whether this practice leads to suboptimal coverage and utilization of Internet data. We first show that while different extractors may lead to similar model performance on standard language understanding tasks, the pages surviving a fixed filtering pipeline can differ substantially. This suggests a simple intervention: by taking a Union over different extractors, we can increase the token yield of DCLM-Baseline by up to 71% while maintaining benchmark performance. We further show that for structured content such as tables and code blocks, extractor choice can significantly impact downstream task performance, with differences of up to 10 percentage points (p.p.) on WikiTQ and 3 p.p. on HumanEval.

cs.CL↗

arXiv2026

The Mighty ToRR: A Benchmark for Table Reasoning and Robustness

Despite its real-world significance, model performance on tabular data remains underexplored, leaving uncertainty about which model to rely on and which prompt configuration to adopt. To address this gap, we create ToRR, a benchmark for Table Reasoning and Robustness, measuring model performance and robustness on table-related tasks. The benchmark includes 10 datasets that cover different types of table reasoning capabilities across varied domains. ToRR goes beyond model performance rankings, and is designed to reflect whether models can handle tabular data consistently and robustly, across a variety of common table representation formats. We present a leaderboard as well as comprehensive analyses of the results of leading models over ToRR. Our results reveal a striking pattern of brittle model behavior, where even strong models are unable to perform robustly on tabular data tasks. Although no specific table format leads to consistently better performance, we show that testing over multiple formats is crucial for reliably estimating model capabilities. Moreover, we show that the reliability boost from testing multiple prompts can be equivalent to adding more test examples. Overall, our findings show that table understanding and reasoning tasks remain a significant challenge.

cs.CL↗

arXiv2026

VLAW: Iterative Co-Improvement of Vision-Language-Action Policy and World Model

The goal of this paper is to improve the performance and reliability of vision-language-action (VLA) models through iterative online interaction. Since collecting policy rollouts in the real world is expensive, we investigate whether a learned simulator-specifically, an action-conditioned video generation model-can be used to generate additional rollout data. Unfortunately, existing world models lack the physical fidelity necessary for policy improvement: they are predominantly trained on demonstration datasets that lack coverage of many different physical interactions (particularly failure cases) and struggle to accurately model small yet critical physical details in contact-rich object manipulation. We propose a simple iterative improvement algorithm that uses real-world roll-out data to improve the fidelity of the world model, which can then, in turn, be used to generate supplemental synthetic data for improving the VLA model. In our experiments on a real robot, we use this approach to improve the performance of a state-of-the-art VLA model on multiple downstream tasks. We achieve a 39.2% absolute success rate improvement over the base policy and 11.6% improvement from training with the generated synthetic rollouts. Videos can be found at this anonymous website: https://sites.google.com/view/vla-w

cs.RO↗

arXiv2026

Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning

Recent video generation models demonstrate remarkable ability to capture complex physical interactions and scene evolution over time. To leverage their spatiotemporal priors, robotics works have adapted video models for policy learning but introduce complexity by requiring multiple stages of post-training and new architectural components for action generation. In this work, we introduce Cosmos Policy, a simple approach for adapting a large pretrained video model (Cosmos-Predict2) into an effective robot policy through a single stage of post-training on the robot demonstration data collected on the target platform, with no architectural modifications. Cosmos Policy learns to directly generate robot actions encoded as latent frames within the video model's latent diffusion process, harnessing the model's pretrained priors and core learning algorithm to capture complex action distributions. Additionally, Cosmos Policy generates future state images and values (expected cumulative rewards), which are similarly encoded as latent frames, enabling test-time planning of action trajectories with higher likelihood of success. In our evaluations, Cosmos Policy achieves state-of-the-art performance on the LIBERO and RoboCasa simulation benchmarks (98.5% and 67.1% average success rates, respectively) and the highest average score in challenging real-world bimanual manipulation tasks, outperforming strong diffusion policies trained from scratch, video model-based policies, and state-of-the-art vision-language-action models fine-tuned on the same robot demonstrations. Furthermore, given policy rollout data, Cosmos Policy can learn from experience to refine its world model and value function and leverage model-based planning to achieve even higher success rates in challenging tasks. We release code, models, and training data at https://research.nvidia.com/labs/dir/cosmos-policy/

cs.AI↗

arXiv2026

Relative Scaling Laws for LLMs

Scaling laws describe how language models improve with additional data, parameters, and compute. While widely used, they are typically measured on aggregate test sets. Aggregate evaluations yield clean trends but average over heterogeneous subpopulations, obscuring performance disparities. We introduce relative scaling laws, which track how performance gaps between test distributions evolve with scale rather than focusing solely on absolute error. Using 255 decoder-only Transformers trained under matched-compute (IsoFLOP) budgets from $10^{18}$--$10^{20}$ FLOPs on standard pretraining datasets, we find diverse trajectories: academic domains on MMLU converge toward parity; regional English dialects shift depending on population size; and clusters of AI risk behaviours split, with capability- and influence-related risks increasing during pretraining while adversarial risks do not. These results show that although scaling improves overall performance, it is not a universal equalizer. To support further study, we release all model checkpoints from this work to enable practitioners to measure relative alongside traditional scaling laws, in order to better prioritize robustness challenges in light of the bitter lesson.

cs.CL↗

arXiv2026

On the Entropy Calibration of Language Models

We study the problem of entropy calibration, which asks whether a language model's entropy over generations matches its log loss on human text. Past work found that models are miscalibrated, with entropy per step increasing as generations grow longer, due to error accumulation. To calibrate the model and improve text quality, it has become standard practice to truncate the distribution, but this approach reduces output diversity, which we would like to avoid. Therefore, in this paper, we ask: does miscalibration improve automatically with scale, and if not, is it theoretically possible to calibrate without tradeoffs? To build intuition, we first study a simplified theoretical setting to characterize the scaling behavior of miscalibration with respect to dataset size. We find that the rate of scaling depends on the power law exponent of the data distribution -- in particular, for a power law exponent close to 1, the scaling exponent is close to 0, meaning that miscalibration improves very slowly with scale. Next, we measure miscalibration empirically in language models ranging from 0.5B to 70B parameters. We find that the observed scaling behavior is similar to what is predicted theoretically: our fitted scaling exponents for text are close to 0, meaning that larger models accumulate error at a similar rate as smaller ones. This scaling (or, lack thereof) provides one explanation for why we sample from larger models with similar amounts of truncation as smaller models, even though the larger models are of higher quality. However, truncation is not a satisfying solution because it comes at the cost of increased log loss. In theory, is it even possible to reduce entropy while preserving log loss? We prove that it is possible, if we assume access to a black box which can fit models to predict the future entropy of text.

cs.CL↗

arXiv2026

RoboReward: General-Purpose Vision-Language Reward Models for Robotics

A well-designed reward is critical for effective reinforcement learning-based policy improvement. In real-world robotics, obtaining such rewards typically requires either labor-intensive human labeling or brittle, handcrafted objectives. Vision-language models (VLMs) have shown promise as automatic reward models, yet their effectiveness on real robot tasks is poorly understood. In this work, we aim to close this gap by introducing (1) RoboReward, a robotics reward dataset and benchmark built on large-scale real-robot corpora from Open X-Embodiment (OXE) and RoboArena, and (2) vision-language reward models trained on this dataset (RoboReward 4B/8B). Because OXE is success-heavy and lacks failure examples, we propose a negative examples data augmentation pipeline that generates calibrated negative and near-misses via counterfactual relabeling of successful episodes and temporal clipping to create partial-progress outcomes from the same videos. Using this framework, we build a large training and evaluation dataset spanning diverse tasks and embodiments to test whether state-of-the-art VLMs can reliably provide rewards for robot learning. Our evaluation of open and proprietary VLMs finds that no model excels across tasks, highlighting substantial room for improvement. We then train general-purpose 4B- and 8B-parameter models that outperform much larger VLMs in assigning rewards for short-horizon robotic tasks. Finally, we deploy the 8B model in real-robot reinforcement learning and find that it improves policy learning over Gemini Robotics-ER 1.5 while narrowing the gap to RL training with human-provided rewards. We release the full dataset, trained reward models, and evaluation suite on our website to advance the development of general-purpose reward models in robotics: https://crfm.stanford.edu/helm/robo-reward-bench (project website).

cs.RO↗

arXiv2026

Extracting books from production language models

Many unresolved legal questions over LLMs and copyright center on memorization: whether specific training data have been encoded in the model's weights during training, and whether those memorized data can be extracted in the model's outputs. While many believe that LLMs do not memorize much of their training data, recent work shows that substantial amounts of copyrighted text can be extracted from open-weight models. However, it remains an open question if similar extraction is feasible for production LLMs, given the safety measures these systems implement. We investigate this question using a two-phase procedure: (1) an initial probe to test for extraction feasibility, which sometimes uses a Best-of-N (BoN) jailbreak, followed by (2) iterative continuation prompts to attempt to extract the book. We evaluate our procedure on four production LLMs -- Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro, and Grok 3 -- and we measure extraction success with a score computed from a block-based approximation of longest common substring (nv-recall). With different per-LLM experimental configurations, we were able to extract varying amounts of text. For the Phase 1 probe, it was unnecessary to jailbreak Gemini 2.5 Pro and Grok 3 to extract text (e.g, nv-recall of 76.8% and 70.3%, respectively, for Harry Potter and the Sorcerer's Stone), while it was necessary for Claude 3.7 Sonnet and GPT-4.1. In some cases, jailbroken Claude 3.7 Sonnet outputs entire books near-verbatim (e.g., nv-recall=95.8%). GPT-4.1 requires significantly more BoN attempts (e.g., 20X), and eventually refuses to continue (e.g., nv-recall=4.0%). Taken together, our work highlights that, even with model- and system-level safeguards, extraction of (in-copyright) training data remains a risk for production LLMs.

cs.CL↗

arXiv2025

The 2025 Foundation Model Transparency Index

Foundation model developers are among the world's most important companies. As these companies become increasingly consequential, how do their transparency practices evolve? The 2025 Foundation Model Transparency Index is the third edition of an annual effort to characterize and quantify the transparency of foundation model developers. The 2025 FMTI introduces new indicators related to data acquisition, usage data, and monitoring and evaluates companies like Alibaba, DeepSeek, and xAI for the first time. The 2024 FMTI reported that transparency was improving, but the 2025 FMTI finds this progress has deteriorated: the average score out of 100 fell from 58 in 2024 to 40 in 2025. Companies are most opaque about their training data and training compute as well as the post-deployment usage and impact of their flagship models. In spite of this general trend, IBM stands out as a positive outlier, scoring 95, in contrast to the lowest scorers, xAI and Midjourney, at just 14. The five members of the Frontier Model Forum we score end up in the middle of the Index: we posit that these companies avoid reputational harms from low scores but lack incentives to be transparency leaders. As policymakers around the world increasingly mandate certain types of transparency, this work reveals the current state of transparency for foundation model developers, how it may change given newly enacted policy, and where more aggressive policy interventions are necessary to address critical information deficits.

cs.AI↗

arXiv2025

BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems

AI agents have the potential to significantly alter the cybersecurity landscape. Here, we introduce the first framework to capture offensive and defensive cyber-capabilities in evolving real-world systems. Instantiating this framework with BountyBench, we set up 25 systems with complex, real-world codebases. To capture the vulnerability lifecycle, we define three task types: Detect (detecting a new vulnerability), Exploit (exploiting a given vulnerability), and Patch (patching a given vulnerability). For Detect, we construct a new success indicator, which is general across vulnerability types and provides localized evaluation. We manually set up the environment for each system, including installing packages, setting up server(s), and hydrating database(s). We add 40 bug bounties, which are vulnerabilities with monetary awards from \$10 to \$30,485, covering 9 of the OWASP Top 10 Risks. To modulate task difficulty, we devise a new strategy based on information to guide detection, interpolating from identifying a zero day to exploiting a given vulnerability. We evaluate 10 agents: Claude Code, OpenAI Codex CLI with o3-high and o4-mini, and custom agents with o3-high, GPT-4.1, Gemini 2.5 Pro Preview, Claude 3.7 Sonnet Thinking, Qwen3 235B A22B, Llama 4 Maverick, and DeepSeek-R1. Given up to three attempts, the top-performing agents are Codex CLI: o3-high (12.5% on Detect, mapping to \$3,720; 90% on Patch, mapping to \$14,152), Custom Agent: Claude 3.7 Sonnet Thinking (67.5% on Exploit), and Codex CLI: o4-mini (90% on Patch, mapping to \$14,422). Codex CLI: o3-high, Codex CLI: o4-mini, and Claude Code are more capable at defense, achieving higher Patch scores of 90%, 90%, and 87.5%, compared to Exploit scores of 47.5%, 32.5%, and 57.5% respectively; while the custom agents are relatively balanced between offense and defense, achieving Exploit scores of 17.5-67.5% and Patch scores of 25-60%.

cs.CR↗

arXiv2025

RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies

Comprehensive, unbiased, and comparable evaluation of modern generalist policies is uniquely challenging: existing approaches for robot benchmarking typically rely on heavy standardization, either by specifying fixed evaluation tasks and environments, or by hosting centralized ''robot challenges'', and do not readily scale to evaluating generalist policies across a broad range of tasks and environments. In this work, we propose RoboArena, a new approach for scalable evaluation of generalist robot policies in the real world. Instead of standardizing evaluations around fixed tasks, environments, or locations, we propose to crowd-source evaluations across a distributed network of evaluators. Importantly, evaluators can freely choose the tasks and environments they evaluate on, enabling easy scaling of diversity, but they are required to perform double-blind evaluations over pairs of policies. Then, by aggregating preference feedback from pairwise comparisons across diverse tasks and environments, we can derive a ranking of policies. We instantiate our approach across a network of evaluators at seven academic institutions using the DROID robot platform. Through more than 600 pairwise real-robot evaluation episodes across seven generalist policies, we demonstrate that our crowd-sourced approach can more accurately rank the performance of existing generalist policies than conventional, centralized evaluation approaches, while being more scalable, resilient, and trustworthy. We open our evaluation network to the community and hope that it can enable more accessible comparisons of generalist robot policies.

cs.RO↗

arXiv2025

Beat the long tail: Distribution-Aware Speculative Decoding for RL Training

Reinforcement learning(RL) post-training has become essential for aligning large language models (LLMs), yet its efficiency is increasingly constrained by the rollout phase, where long trajectories are generated token by token. We identify a major bottleneck:the long-tail distribution of rollout lengths, where a small fraction of long generations dominates wall clock time and a complementary opportunity; the availability of historical rollouts that reveal stable prompt level patterns across training epochs. Motivated by these observations, we propose DAS, a Distribution Aware Speculative decoding framework that accelerates RL rollouts without altering model outputs. DAS integrates two key ideas: an adaptive, nonparametric drafter built from recent rollouts using an incrementally maintained suffix tree, and a length aware speculation policy that allocates more aggressive draft budgets to long trajectories that dominate makespan. This design exploits rollout history to sustain acceptance while balancing base and token level costs during decoding. Experiments on math and code reasoning tasks show that DAS reduces rollout time up to 50% while preserving identical training curves, demonstrating that distribution-aware speculative decoding can significantly accelerate RL post training without compromising learning quality.

cs.LG↗

arXiv2025

Machine Unlearning Doesn't Do What You Think: Lessons for Generative AI Policy and Research

"Machine unlearning" is a popular proposed solution for mitigating the existence of content in an AI model that is problematic for legal or moral reasons, including privacy, copyright, safety, and more. For example, unlearning is often invoked as a solution for removing the effects of specific information from a generative-AI model's parameters, e.g., a particular individual's personal data or the inclusion of copyrighted content in the model's training data. Unlearning is also proposed as a way to prevent a model from generating targeted types of information in its outputs, e.g., generations that closely resemble a particular individual's data or reflect the concept of "Spiderman." Both of these goals--the targeted removal of information from a model and the targeted suppression of information from a model's outputs--present various technical and substantive challenges. We provide a framework for ML researchers and policymakers to think rigorously about these challenges, identifying several mismatches between the goals of unlearning and feasible implementations. These mismatches explain why unlearning is not a general-purpose solution for circumscribing generative-AI model behavior in service of broader positive impact.

cs.LG↗

arXiv2025

SpecEval: Evaluating Model Adherence to Behavior Specifications

Companies that develop foundation models publish behavioral guidelines they pledge their models will follow, but it remains unclear if models actually do so. While providers such as OpenAI, Anthropic, and Google have published detailed specifications describing both desired safety constraints and qualitative traits for their models, there has been no systematic audit of adherence to these guidelines. We introduce an automated framework that audits models against their providers specifications by parsing behavioral statements, generating targeted prompts, and using models to judge adherence. Our central focus is on three way consistency between a provider specification, its model outputs, and its own models as judges; an extension of prior two way generator validator consistency. This establishes a necessary baseline: at minimum, a foundation model should consistently satisfy the developer behavioral specifications when judged by the developer evaluator models. We apply our framework to 16 models from six developers across more than 100 behavioral statements, finding systematic inconsistencies including compliance gaps of up to 20 percent across providers.

cs.CL↗

arXiv2025

Blackbox Model Provenance via Palimpsestic Membership Inference

Suppose Alice trains an open-weight language model and Bob uses a blackbox derivative of Alice's model to produce text. Can Alice prove that Bob is using her model, either by querying Bob's derivative model (query setting) or from the text alone (observational setting)? We formulate this question as an independence testing problem--in which the null hypothesis is that Bob's model or text is independent of Alice's randomized training run--and investigate it through the lens of palimpsestic memorization in language models: models are more likely to memorize data seen later in training, so we can test whether Bob is using Alice's model using test statistics that capture correlation between Bob's model or text and the ordering of training examples in Alice's training run. If Alice has randomly shuffled her training data, then any significant correlation amounts to exactly quantifiable statistical evidence against the null hypothesis, regardless of the composition of Alice's training data. In the query setting, we directly estimate (via prompting) the likelihood Bob's model gives to Alice's training examples and order; we correlate the likelihoods of over 40 fine-tunes of various Pythia and OLMo base models ranging from 1B to 12B parameters with the base model's training data order, achieving a p-value on the order of at most 1e-8 in all but six cases. In the observational setting, we try two approaches based on estimating 1) the likelihood of Bob's text overlapping with spans of Alice's training examples and 2) the likelihood of Bob's text with respect to different versions of Alice's model we obtain by repeating the last phase (e.g., 1%) of her training run on reshuffled data. The second approach can reliably distinguish Bob's text from as little as a few hundred tokens; the first does not involve any retraining but requires many more tokens (several hundred thousand) to achieve high power.

cs.LG↗

arXiv2025

Holistic Agent Leaderboard: The Missing Infrastructure for AI Agent Evaluation

AI agents have been developed for complex real-world tasks from coding to customer service. But AI agent evaluations suffer from many challenges that undermine our understanding of how well agents really work. We introduce the Holistic Agent Leaderboard (HAL) to address these challenges. We make three main contributions. First, we provide a standardized evaluation harness that orchestrates parallel evaluations across hundreds of VMs, reducing evaluation time from weeks to hours while eliminating common implementation bugs. Second, we conduct three-dimensional analysis spanning models, scaffolds, and benchmarks. We validate the harness by conducting 21,730 agent rollouts across 9 models and 9 benchmarks in coding, web navigation, science, and customer service with a total cost of about $40,000. Our analysis reveals surprising insights, such as higher reasoning effort reducing accuracy in the majority of runs. Third, we use LLM-aided log inspection to uncover previously unreported behaviors, such as searching for the benchmark on HuggingFace instead of solving a task, or misusing credit cards in flight booking tasks. We share all agent logs, comprising 2.5B tokens of language model calls, to incentivize further research into agent behavior. By standardizing how the field evaluates agents and addressing common pitfalls in agent evaluation, we hope to shift the focus from agents that ace benchmarks to agents that work reliably in the real world.

cs.AI↗