arXiv subjects

Dawn Song

Publications and source records attributed to Dawn Song.

At least 19 recordsLinked to original sources

arXiv2026

What if automating AI R&D triggers an intelligence explosion?

In contrast to even a year ago, AI systems now write most of the code inside the companies that build them. As more of the AI research and development (R&D) pipeline is automated, could AI progress radically accelerate in an "intelligence explosion," where years of advances are compressed into months or less? Preliminary evidence suggests that it could. In this work, we assess this evidence, analyze an intelligence explosion's potential impacts, and propose policy responses. AI systems are on track to automate most AI R\&D work within a few years, and possibly all of it. If this triggers an intelligence explosion, it could dramatically bring forward AI's benefits, but also pose extreme risks: capabilities growth could accelerate far beyond what society can keep up with, humanity could lose control over superhuman AI systems, and checks on power within and between states, companies, and branches of government could be severely eroded. Although there remains much uncertainty about these possibilities, the high stakes warrant serious further attention. Policymakers should urgently obtain more visibility into the automation of AI R&D, develop ways to steer and constrain an intelligence explosion, and prepare society to adapt to an intelligence explosion's impacts.

cs.CY↗

arXiv2026

ParetoRSI-Gym: Can AI Agents Improve Language Models Across Multiple Domains?

Recursive self-improvement (RSI) requires an agentic system to design, train, and evaluate its own successor. Data-driven post-training is a central part of that loop, yet existing evaluations measure it on a single benchmark, where an agent selects from a static pool or distills a stronger model to push one score up in one run. Real post-training is multi-domain, and a gain in one domain has to be weighed against a regression in another. We propose ParetoRSI-Gym, in which an agent searches for data, curates it, trains, reads back per-domain results, and repeats, with everything but the data held fixed: the initial LLM, the five domains (math, code, finance, medicine, science), the toolset, and the SFT pipeline. Generalization benchmarks in each domain separate capability gains from overfitting to the scores the agent optimizes against. Across three agent systems, three base models, five rounds and multiple seeds, the strongest system, Claude Code with Opus-5, raises the generalization scores of its proxy-best checkpoint by 9.2% relative to the starting model (+2.3 absolute points), averaged with equal weight over three base models and five domains, with a single-domain relative gain of up to 63% (medicine on Ministral-3-3B). Further analysis shows that the harness matters as much as the model it drives, and that each agent keeps a unique data selection strategy. We see our work as a step towards evaluation of RSI systems in real-world machine learning engineering.

cs.LG↗

arXiv2026

Can Test-Time Memory Evolution Outperform Test-Time Training for Scientific Discovery?

Recent \emph{test-time training} (TTT) frameworks have enabled open-weight large language models (LLMs) to outperform inference-only frameworks powered by frontier models on scientific discovery tasks. However, TTT introduces additional computational cost for online weight updates during the discovery process, making computational efficiency an important consideration when scaling to larger models. This motivates us to explore an efficient \emph{test-time adaptation} method for scientific discovery. We introduce \emph{Test-Time Memory Evolution} (TTME), a plug-and-play test-time adaptation method, as an alternative to TTT. TTME continually updates the external memory by distilling the experience to guide subsequent discovery without updating model weights. To have a clean comparison, we evaluate TTME and TTT under the same discovery framework, replacing only the test-time adaptation method. Our evaluation shows that TTME outperforms TTT on all five mathematical tasks, while reducing computational cost by $82.8$-$84.6\%$ in FLOPs and $47.7$-$58.4\%$ in runtime. These results establish TTME as an effective and computationally efficient alternative to TTT in the evaluated discovery settings.

cs.AI↗

arXiv2026

Agentic Verified Optimization of Real-World Software

Program optimization aims to improve performance while preserving behavior. Coding agents can already make programs faster by changing algorithms, data structures, and execution strategies. However, existing equivalence checks mostly rely on tests, which agents have been shown to exploit, passing the tests without preserving the intended behavior. We propose agentic verified optimization, in which an agent produces a faster implementation together with a machine-checked proof of equivalence to the existing one. Because the existing implementation serves as the specification, the task avoids the specification-writing bottleneck of verified code generation, and recent advances in theorem proving make the proofs attainable. We formulate the task at the function and repository levels in Lean 4, where programs can be both executed and reasoned about. Few real-world applications are written in Lean, so we develop Aqueduct, an agentic framework that migrates C, C++, Go, and Rust programs into native Lean, validates each port by differential testing against its original. On the 189 reference implementations in Verina, agents produce faster proved rewrites for 156, including algorithmic changes such as replacing exponential recursion with dynamic programming. On 17 migrated applications, their 92 proved rewrites speed up 47 of 51 application scenarios. These application-level gains are smaller than the function-level ones, making the choice of functions to optimize an important problem. Current agents can thus already carry out verified optimization on both standalone functions and applications migrated from real programs. We discuss the remaining challenges and the steps toward verified optimization in production.

cs.SE↗

arXiv2026

CyberArena: Benchmarking AI Agents in Live Symmetric Attack-and-Defense Combat

AI agents are demonstrating increasingly strong offensive and defensive cybersecurity capabilities in the real world. Current benchmarks mostly evaluate agents in isolation against static targets. Real cyber operations, however, are dynamic and coupled: security postures constantly change as attackers and defenders adapt to each other in real time. Static benchmarks quickly saturate and miss an agent's ability to attack and defend simultaneously under live pressure. To address this gap, we present CyberArena, a dynamic, symmetric adversarial benchmark built around multi-round attack-and-defense combat. In isolated sandboxes, two agents compete under identical rules, acting concurrently as both attackers and defenders over continuously changing services. This converts static targets into evolving, semi-open challenges, directly testing an agent's speed and precision in discovering, exploiting, and patching vulnerabilities. To sustain the benchmark's longevity, an automated agentic pipeline continuously generates, reviews, and validates new environments. Furthermore, because pairwise contests are notoriously hard to quantify, we introduce a dynamical model that maps match dynamics into unified, multidimensional scores across offensive, defensive, and domain-specific capabilities. Evaluating 12 frontier LLM agents across 40 challenges in 466 matches, CyberArena separates model capabilities where single-agent baselines hit a ceiling. The trajectories read like a fight rather than an exam: defenders patch one hole and miss the next and certify repairs on stale evidence, while capable agents think a move ahead, planting backdoors for later rounds and, in one case, writing their own packet sniffer to watch the opponent. Beyond a benchmark, CyberArena serves as an adaptable, scalable experimental platform for future research on cybersecurity agents.

cs.CR↗

arXiv2026

CUA-Lite: An Open Platform for Computer Use Agents

Open resources for computer use agents (CUAs) -- environments, datasets, and models -- remain scattered: without a standard, these resources cannot be integrated and scaled; additionally, evaluation, RL, and SFT cannot share one infrastructure. We present CUA-Lite, an open platform for developing CUAs, built around three standardized abstractions: (1) Unified environment interface, under which we integrate 12 benchmarks across desktop, browser, and mobile, plus lightweight VM-free sandboxes that replicate OSWorld's desktop at much lower cost, providing over 30,000 verifiable tasks. (2) Unified supervised data format, into which we convert the major public CUA datasets, from GUI grounding and understanding to full rollouts, and record our fresh rollouts from frontier CUAs on these verifiable environments. (3) Model harnesses shared across evaluation, RL, and SFT, one per model and implemented for 14 model families. Each harness runs its model in these computer environments, producing the rollouts that evaluation and RL consume, or renders stored supervised data into the model's own training format for SFT. We release reference scores across the integrated benchmarks, and show that RL and SFT through the same harness improve open-weight models. We release our dataset, code, and models, and the platform is open to contributions.

cs.AI↗

arXiv2026

VerifierBench: Benchmarking the Generation of Executable Verifiers for LLM Agents

Agents need reliable ways to validate their outputs, both when completing tasks and when filtering synthetic training data. We study verifier generation: from the instruction and environment available for solving a task, an agent writes an executable verifier that must accept every valid solution and reject invalid ones. We introduce VerifierBench, a benchmark of 222 tasks from six agent benchmarks spanning tool-call, state-based, and code verification, with 4,993 hidden test cases that include alternative valid solutions and attacks; Verifier4Verifier (V4V), a six-check audit, screens its source verifiers and test suites. Across 19 models in a common agent harness, normalized balanced accuracy ranges from 10.6 to 62.1, with generated verifiers consistently rejecting invalid solutions more reliably than they accept valid ones. This apparent strength can mask overly restrictive verification: most false rejections in the analyzed records come from verifiers that accept no tested solution to a task. Case studies reveal unsupported assumptions about solution formats, tool use, and code changes that exclude valid solutions. These errors survive self-testing because agents test solutions that share the verifier's mistaken assumptions, or submit verifiers without confirming that any correct solution passes. These findings suggest testing both whether a verifier accepts a correct solution and whether it accepts other solutions permitted by the task. The project page and dataset are publicly available.

cs.SE↗

arXiv2026

Scaling Online Exploration for Robot Agents

Robot learning has advanced by scaling demonstration datasets and model capacity; however, deploying a new robot embodiment or task typically necessitates a dedicated teleoperation corpus. The resulting policies remain bounded by their imitation distributions, limiting generalization and continual self-improvement. We introduce RAX (Robot Agent eXploration), an open-source framework for robot agents centered on scaling exploration: enabling agents to improve by acquiring online experience and synthesizing executable programs. In RAX, a general-purpose coding agent generates robot control code, operating without action demonstration data, hand-engineered action primitives, or value-function guidance. Through online exploration, the agent identifies decision sequences that achieve task goals and compiles them into modular programs encapsulating the full perception-planning-control loop. Using RAX, we present a systematic multi-embodiment study on scaling online exploration across precision and long-horizon manipulation tasks using single-arm, bimanual, dexterous, mobile, and humanoid systems in simulation and on real robots. Generalization performance on unseen environments improves consistently as exploration scales. In simulation, RAX outperforms top baseline models by +37.4% on average. On real robots, RAX autonomously learns memory-dependent, articulated, high-precision, deformable, and long-horizon mobile manipulation tasks without human demonstrations, continually improving performance by composing verified skills from prior exploration sessions. As exploration scales, we observe emergent capabilities including active perception, error recovery, safety guardrails, and human-robot collaboration. We open-source RAX as a generalist robot-agent platform in which new robot embodiments are integrated through their APIs alone, without teleoperated data collection or policy training.

cs.RO↗

arXiv2026

EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents

Large language model (LLM) agents can translate natural-language requests into real-world actions, creating safety risks from both indirect prompt injections and directly harmful requests. Existing system-level safety harnesses are typically designed by experts and applied across heterogeneous models and domains, despite substantial differences in model behavior, utility tolerance, and domain-specific safety requirements. We present EvoSafeHarness, a framework for automatically synthesizing deployment-specific safety harnesses for frozen LLM agents. Given a target model and domain specification, EvoSafeHarness jointly evolves a natural-language policy and executable enforcement logic using behavioral feedback from the target model. Each candidate is additionally evaluated through fresh-context adversarial review to filter rules that depend on benchmark artifacts. Across four agent benchmark families, EvoSafeHarness achieves a stronger safety-utility frontier than fixed expert-designed defenses. On DecodingTrust-Agent, evaluated across fifteen model-domain deployments, it reduces attack success rate from 50.9% to 12.6% for direct attacks and from 40.4% to 7.4% for indirect attacks, with a 3.3-point utility cost. On AgentDojo, it achieves 82.8% utility with 0.0% attack success rate, twice the utility of CaMeL at the same zero-attack-success operating point, while the same harness transfers unchanged to unseen AgentDyn suites. It also achieves the best score on Agent-SafetyBench for every evaluated victim model and maintains an attack success rate below 18% against an adaptive PAIR-style attacker with sixteen attempts per task. Analysis of the synthesized harnesses shows that domain semantics determine which safety relations and trajectory state must be tracked, while model and runtime behavior determine how and where those relations should be enforced.

cs.CR↗

arXiv2026

Strategy Executability in Mathematical Reasoning: Leveraging Human-Model Differences for Effective Guidance

Example-based guidance is widely used to improve mathematical reasoning at inference time, yet its effectiveness is highly unstable across problems and models-even when the guidance is correct and problem-relevant. We show that this instability arises from a previously underexplored gap between strategy usage-whether a reasoning strategy appears in successful solutions-and strategy executability-whether the strategy remains effective when instantiated as guidance for a target model. Through a controlled analysis of paired human-written and model-generated solutions, we identify a systematic dissociation between usage and executability: human- and model-derived strategies differ in structured, domain-dependent ways, leading to complementary strengths and consistent source-dependent reversals under guidance. Building on this diagnosis, we propose Selective Strategy Retrieval (SSR), a test-time framework that explicitly models executability by selectively retrieving and combining strategies using empirical, multi-route, source-aware signals. Across multiple mathematical reasoning benchmarks, SSR yields reliable and consistent improvements over direct solving, in-context learning, and single-source guidance, improving accuracy by up to $+13$ points on AIME25 and $+5$ points on Apex for compact reasoning models. Code and benchmark are publicly available at: https://github.com/lwd17/strategy-execute-pipeline.

cs.AI↗

arXiv2026

TestJack: Should you trust the results in coding benchmarks? Agentic Coding Benchmarks Auditing via Evaluator Evolution

Large language model (LLM) agents are rapidly reshaping software engineering, accompanied by an explosion of new code benchmarks. Yet nearly all existing benchmarks still rely on the same decades-old criterion: a solution is correct if it passes a fixed set of unit tests. Such tests are often insufficient: they check only part of what the task requires, so agents can reward hack them or silently miss required behavior while still passing every test. As a result, higher benchmark scores may partly reflect better adaptation to the evaluator rather than better problem solving. Existing works focus on static test augmentation: they strengthen each task's tests once, before any trial is seen, and thus overlook how real trials actually fail. We introduce TestJack, a scalable framework for evaluating patches beyond fixed tests. For each trial, TestJack generates tests targeting prompt requirements the patch may violate, retains only tests passed by the ground-truth patch, and re-examines any trial failures. Each confirmed failure is thus supported by a replayable test. To reduce evaluation cost, we also introduce a lightweight variant which audits a random sample of trials in depth and reuses the resulting tests across all trials for the same task. Across 6 frontier model backends and 5 benchmarks such as DeepSWE and SWE Marathon, we find that about 34.4% of the model trials currently judged correct violate the task requirements, lowering the overall resolution rate from 50.6% to 33.2%. Our results reveal a fundamental limitation of current coding-agent evaluation: as LLMs become better at optimizing against fixed evaluators, those evaluators themselves must become more adaptive.

cs.SE↗

arXiv2026

NetAgent: Multi-Task Agentic Network Traffic Analysis Made Practical

Network traffic analysis is central to network security, spanning tasks from intrusion detection to encrypted traffic classification. Existing approaches either train task-specific models that generalize poorly or rely on costly traffic foundation models that still struggle under distribution shift. We present NetAgent, the first agentic framework for multi-task traffic analysis. Through a carefully designed agent loop, NetAgent supports complex task understanding, on-the-fly decomposition and orchestration, dynamic replanning, and long-horizon analysis, without task-specific training. It introduces five key designs: (1) knowledge-augmented workflow planning that maps attack knowledge to traffic features to bridge the semantic gap; (2) a comprehensive tool action space with 150+ verified tools extracted from 50+ published systems; (3) a unified code execution space for flexible action composition; (4) a three-tier memory for long-term knowledge consolidation; and (5) sandboxing and runtime repair for reliable execution. Across 9 major benchmarks, NetAgent outperforms all baselines (23 single-task and 5 multi-task) on nearly all tasks and generalizes substantially better to unseen traffic distribution (90.04% F1 vs. 2.74% and 3.04% for the best single-task and multi-task baselines) and under realistic background shift (4.85-point F1 drop vs. 74.88-point and 74.80-point drop for the best single-task and multi-task baselines). These results reveal that existing methods owe much of their reported success to overfitting dataset-specific patterns and degrade sharply in realistic network environments, while NetAgent's agentic design remains accurate, generalizable, and robust.

cs.CR↗

arXiv2026

Component and Dimension Sparsity in Transformer Refusal Mechanisms

Activation steering manipulates large language model behavior by intervening on internal activations, but the mechanistic basis of these interventions remains poorly understood. We decompose refusal steering into component-level interventions across four open-weight models, identifying the sparse subsets of attention and MLP components whose steering suffices to reproduce the full behavioral effect. We find that refusal directions concentrate in sparse component mechanisms comprising 28--48\% of upstream components, retaining 88--101\% of steering effectiveness. Within these mechanisms, effective steering further concentrates in approximately 50\% of residual stream dimensions, retaining 85--98\% of the component-mechanism baseline, consistent with a privileged basis structure. Sparsity thus operates at two levels: which components are steered, and which dimensions within those components carry the signal. Together these findings show that refusal is not diffusely encoded across a transformer but assembled by a structured, identifiable mechanism, providing a foundation for mechanistic understanding of how refusal behaviors are represented and steered. To facilitate reproducibility, we release all code and raw experimental results in https://github.com/wang-research-lab/Refusal_Mechanisms.

cs.CL↗

arXiv2026

Generalization Dynamics of LM Pre-training

People typically assume that LMs stably mature from pattern-matching parrots to generalizable intelligence during pre-training. We build a toy eval suite and show this mental model is wrong: throughout pre-training, LMs frequently and suddenly hop between parrot-like and intelligence-like computations. We call this mode-hopping. Across our suite, LMs suddenly latch onto memorized or in-context patterns instead of in-context learning, use System 1 instead of System 2 thinking, pick up what sounds true instead of what is true, fail at multi-hop persona QA, out-of-context reasoning, and emergent misalignment -- then just as suddenly revert and generalize. Mode-hopping is not explained by standard optimization dynamics: it is locally stable and cannot be fixed by checkpoint averaging. We instead think of it as a capacity allocation problem: in a capacity-bounded model, generalizable circuits must compete with the shallow ones learned early in training, and the data in each pre-training window may decide which circuits win. Our suite provides a new efficient lens on generalization. We demonstrate two concrete applications: (i) select intermediate pre-training checkpoints that strongly generalize reasoning and alignment, better than the final pre- or mid-training checkpoints, and (ii) select pre-training data that controls and stabilizes generalization dynamics.

cs.CL↗

arXiv2026

AgentXploit: Autonomous Repository-to-Runtime Red-Teaming for AI Agents

AI agents combine language models with external data and tools that can modify files, call APIs, or execute code. Security failures can arise when adversarial content changes an agent's tool use or when the surrounding software contains vulnerabilities such as path traversal or command injection. We study authorized white-box pre-deployment auditing, where the auditor has access to the target repository and a controlled runtime, but successful attacks must still act through the task-defined attacker interface and be confirmed by an external verifier. We present AgentXploit, a two-role auditing system that separates repository-level attack-path discovery from runtime exploitation. The Analyzer Agent traces attacker-controlled inputs to sensitive operations and records code-supported candidate attack paths; the Exploiter Agent turns these paths into concrete attacks and revises them using runtime feedback. We also introduce AgentXploit-Bench, containing 72 reproducible vulnerabilities across 12 open-source AI-agent systems and frameworks. Across three runs, AgentXploit reaches 59.3% end-to-end success, compared with 38.4% for Codex. Under a token-budget-matched comparison, Codex reaches 46.3%. On AgentDojo, where injection points are provided, the Exploiter Agent reaches 79.2% attack success versus 52.7% for AgentVigil. These results highlight repository discovery and runtime exploitation as distinct challenges in end-to-end agent security auditing.

cs.CR↗

arXiv2026

MobileCybench: Evaluating Agent Vulnerability Discovery via Executable Probes

AI agents now report vulnerabilities faster than maintainers can review them. Reports often depend on security properties specific to the application, and require considerable human labor to process. To mitigate this, we introduce a framework for evaluating vulnerability reports via probes, executable checks of security properties. A reported exploit is evaluated by replaying it against the application and running the probes: a triggered probe indicates both that the exploit succeeded and which security property it violated. As a probe encodes a security property rather than a known vulnerability, it can detect vulnerabilities that were not known when the probe was written. We instantiate the framework as MobileCybench, a benchmark for vulnerability discovery by AI agents in 13 Android applications, with 495 probes written and reviewed by the authors. We evaluate 5 coding agents (OpenCode with GPT-5.5, GPT-5.6-Sol, and GLM-5.2; Claude Code with Opus 4.8 and Opus 5) under 4 settings: as a malicious app on the victim's device or as a remote attacker with a low-privilege account, each with either only an obfuscated APK or access to the application's source code. Given only the obfuscated APK, the top agent, OpenCode with GPT-5.6-Sol, triggers probes in 53.8% of applications in the malicious-app setting and 16.7% in the remote-attacker setting. With source code, the trigger rate across all agents and both attack settings increases from 28.8% to 32.8%. Building and running the benchmark surfaced 23 previously unreported vulnerabilities, the majority of which have been confirmed by maintainers.

cs.CR↗

arXiv2026

Can Generalist Agents Automate Data Curation?

Curating training data is among the most consequential yet labor-intensive parts of modern AI development: practitioners iteratively propose, implement, evaluate, and revise data policies against noisy benchmark feedback. We ask whether generalist coding agents can automate this data-curation loop. We introduce *Curation-Bench*, an agent-centric benchmark that fixes the model, training recipe, and evaluation suite while giving agents command-line access to inspect data, implement policies, submit them to a fixed training/evaluation pipeline, and revise. In a vision-language instruction-tuning instantiation, out-of-the-box agents reach strong published data-selection baselines within ten iterations. However, trajectory analysis reveals a persistent *execution-research gap*: agents mainly tune local policy variants rather than explore new policy families, even when given strategy guides and paper references. Scaffolds requiring each iteration to cite, instantiate, and adapt a prior method shift agents toward method-guided exploration. The scaffolded agent autonomously composes -- without human design input -- a data-selection policy that outperforms strong published baselines at one-tenth their data budget. Overall, current agents can run the curation loop, but reliable data research requires scaffolded method adaptation, not open-ended prompting alone. Code and benchmark are open-sourced.

cs.AI↗

arXiv2026

BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure

LM-agent benchmarks increasingly function as interactive evaluation infrastructure. Agents observe state, call tools, modify workspaces, submit artifacts, and receive rewards from outcome procedures. This interactivity makes evaluations vulnerable to reward hacking: an agent improves its measured score by exploiting the reward-relevant trajectory instead of solving the intended task. Existing defenses rely largely on task-specific patches, prompt instructions, or post-hoc detectors. They do not provide reusable evidence that a concrete run remained within its intended evaluation boundary. This paper presents BenchShield, a model-backed instrumentation layer for reward integrity in LLM-agent evaluation. BenchShield grounds detection in a finite lifecycle model of an evaluation's reward-relevant events. Within the benchmark infrastructure, two complementary analyses operate over this model. A static, phase-aware taint analysis exposes reward-hacking paths before a run. Its runtime counterpart uses infrastructure-side evidence to attribute concrete agent use and emit evidence-backed claims. We construct BenchShield Trajectories, a human-labeled corpus of 456 adjudicated trajectories from more than 31,000 public agent runs across three benchmarks. Compared with an agentic hackability scanner baseline on the same tasks and model, BenchShield improves full-chain recall from 23-94% to 77-100%, same-vector coverage from 16-56% to 43-78%, and reduces per-task cost by up to 65%. Its runtime analysis achieves 96% accuracy in detecting reward hacking from infrastructure-side evidence.

cs.CR↗