ICML 2026 Papers

The scaling-law era has transformed artificial intelligence (AI) from research into a global industry, but its rapid growth also raises concerns over energy usage, carbon emissions, and environmental sustainability. Unlike traditional sectors, the AI industry still lacks systematic carbon accounting methods that support large-scale estimates without reproducing the original training process. This leaves open questions about how large the problem is today and how large it might be in the near future. Given its central role in hosting open-source AI models, the Hugging Face (HF) platform provides a large-scale and publicly accessible corpus for carbon accounting. We estimate aggregate training emissions of HF open-source models using available emissions, energy, compute, and model metadata. To address uneven disclosure quality, we introduce a tiered approach to handle incomplete metadata, supported by empirical regressions that assess estimation reliability. We further introduce AI training carbon intensity (ATCI, emissions per compute), a metric to assess the sustainability efficiency of model training. Our results show that training the most popular open-source models (with over 5,000 downloads) has already resulted in approximately 6.0×10^4 metric tons of carbon emissions. Overall, this paper provides a scalable, empirically grounded framework for estimating training emissions from incomplete disclosures and informing future carbon reporting standards in the AI industry. Data and code are available at https://github.com/insait-institute/HuggingCarbon.

Safe deployment of increasingly capable AI agents may require visibility into how they make decisions. Chain-of-thought (CoT) monitoring can detect misbehavior in today’s reasoning models, but this “monitorability” may be fragile under different training procedures, data sources, or continued system scaling. We propose three evaluation archetypes (intervention, process, and outcome-property), a new monitorability metric, and a broad evaluation suite. We show CoT monitoring outperforms action-only monitoring in practical settings, and that frontier models are generally—but not perfectly—monitorable. We study scaling trends with pre-training model size and inference-time compute, finding longer CoTs are typically more monitorable. We find that, for a fixed capability level, using a smaller model at higher reasoning effort can yield higher monitorability, at greater inference compute cost. We further find that increasing a weak monitor’s test-time compute when monitoring a strong agent improves monitorability, and giving the monitor access to the CoT both boosts monitorability and steepens the compute–to-monitorability scaling trend. Finally, we show monitorability can be improved by asking follow-up questions and giving the follow-up CoT to the monitor.

With the rapid advancement of high-fidelity video generation models, robust AI-generated video (AIGV) detection has become increasingly needed. While most AIGV detection methods operate in the decoded pixel domain, we observe that detection in the pixel domain inevitably entangles task-irrelevant semantic information, leading to substantial semantic redundancy and extensive redundant computation, while overlooking free-to-use signals in compressed bitstreams. In particular, motion vectors and residuals directly encode temporal and spatial generative artifacts but remain largely underexplored. To address these issues, we propose a unified framework for **S**patio-**T**emporal **RE**sidual and **A**rtifact **M**ining, namely **STREAM**, which enables AIGV detection directly from compressed bitstreams. **STREAM** leverages I-frames, motion vectors, and residual errors to capture spatiotemporal artifacts that are typically smoothed out by decompression filters. In particular, we design a lightweight network with a motion-guided alignment module and a gated fusion mechanism, enabling adaptive fusion of spatial artifacts and nonlinear temporal dynamics. Extensive experimental results demonstrate that **STREAM** achieves SOTA performance with an mAP of 0.965, with 2.5× faster inference than previous SOTA baselines.

MLLM-powered embodied agents deployed in real-world environments encounter physical hazards. However, existing approaches lack explicit mechanisms for identifying hazards and reasoning about action-conditioned risks, leading agents to either miss risky interactions or over-identify risks. To address this, we propose EMBGuard, the first MLLM-based safety guardrail for embodied agents designed to decouple physical risk reasoning from agent policy. By evaluating a (visual observation, action) pair, EMBGuard identifies hazardous configurations and provides natural language explanations of potential risks. Alongside EMBGuard, we contribute EMBHazard, a training dataset of 15.1K action-conditioned pairs, and EMBGuardTest, a benchmark of 329 manually curated real-world scenarios spanning seven physical risk categories. Through compositional variation of hazards and actions, we generate diverse risky and benign scenarios that agents may encounter during planning. Despite its compact size (2B, 4B), EMBGuard achieves performance competitive with proprietary MLLMs (e.g., GPT-5.1, Gemini-2.5-Pro) while significantly reducing the false-positive rates that hinder real-time deployment. We make the code, data, and models publicly available at https://github.com/dongwxxkchoi/EMBGuard.

Sparse Autoencoders (SAEs) extract features from LLM internal activations, meant to correspond to interpretable concepts. A core SAE training hyperparameter is L0: how many SAE features should fire per token on average. Existing work compares SAE algorithms using sparsity-reconstruction tradeoff plots, implying L0 is a free parameter with no inherently correct value aside from its effect on reconstruction. In this work we study the effect of L0 on SAEs, and show that if L0 is not set correctly, the SAE fails to disentangle the underlying features of the LLM. If L0 is too low, the SAE will mix correlated features to improve reconstruction. If L0 is too high, the SAE finds degenerate solutions that also mix features. Further, we present a proxy metric that can help guide the search for the correct L0 for an SAE on a given training distribution. We show that our method finds the correct L0 in toy models and coincides with peak sparse probing performance in LLM SAEs. We find that most commonly used SAEs have an L0 that is too low. Our work shows that practitioners must set L0 correctly to train SAEs with monosemantic features.

Conversational assistants are increasingly integrating long-term memory with large language models (LLMs). This persistence of memories, e.g., the user is vegetarian, can enhance personalization in future conversations. However, the same persistence can also introduce safety risks that have been largely overlooked. Hence, we introduce \textbf{PersistBench} to measure the extent of these safety risks. We identify two long-term memory-specific risks: \textit{cross-domain leakage}, where LLMs inappropriately inject context from the long-term memories; and \textit{memory-induced sycophancy}, where stored long-term memories insidiously reinforce user biases. We evaluate 18 frontier and open-source LLMs on our benchmark. Our results reveal a surprisingly high failure rate across these LLMs - a median failure rate of $53\%$ on cross-domain samples and $97\%$ on sycophancy samples. To address this, our benchmark encourages the development of more robust and safer long-term memory usage in frontier conversational systems.

Mechanistic Interpretability (MI) aims to reverse-engineer model behaviors by identifying functional sub-networks. Yet, the scientific validity of these findings depends on their stability. In this work, we argue that circuit discovery is not a standalone task but a statistical estimation problem built upon causal mediation analysis (CMA). We uncover a fundamental instability at this base layer: exact, single-input CMA scores exhibit high intrinsic variance, implying that the causal effect of a component is a volatile random variable rather than a fixed property. We then demonstrate that circuit discovery pipelines inherit this variance and further amplify it. Fast approximation methods, such as Edge Attribution Patching and its successors, introduce additional estimation noise, while aggregating these noisy scores over datasets leads to fragile structural estimates. Consequently, small perturbations in input data or hyperparameters yield vastly different circuits. We systematically decompose these sources of variance and advocate for more rigorous MI practices, prioritizing statistical robustness and routine reporting of stability metrics.

Real-world machine learning (ML) pipelines rarely produce a single model; instead, they produce a Rashomon set of many near-optimal ones. We show that this multiplicity reshapes key aspects of trustworthiness. At the individual-model level, sparse interpretable models tend to preserve privacy but are fragile to adversarial attacks. In contrast, the diversity within a large Rashomon set enables reactive robustness: even when an attack compromises one model, a practitioner can switch to a different near-optimal model that remains accurate, without retraining. However, the same diversity increases information leakage, as disclosing more near-optimal models provides an attacker with progressively richer views of the training data. This produces a robustness–privacy trade-off governed by diversity, which we analyze theoretically and empirically. Beyond this trade-off, Rashomon sets are stable under small distribution shifts, so a set computed once remains valid under such shifts without re-computation. Our results highlight the dual role of Rashomon sets as both a resource and a risk for trustworthy ML.

Large language models (LLMs) are vulnerable to malicious inputs, posing serious risks in high-stakes applications. Although existing detection-based defenses achieve strong empirical performance, they generally lack explicit control over the false acceptance rate (FAR), a critical safety requirement in sensitive deployment scenarios. This challenge is further complicated by two practical constraints: the lack of malicious calibration samples and the streaming nature of real-world inputs. To address these challenges, we propose \textit{OSCS}, a novel framework for online FAR control without requiring malicious calibration data. OSCS leverages detection scores produced by existing defenses and employs recursive density estimation to estimate benign probability from the test stream. Based on these estimates, OSCS performs real-time accept/reject decisions while provably satisfying a user-specified FAR target. Theoretically, we show that OSCS controls the FAR up to a vanishing excess term under mild assumptions. Extensive experiments on backdoor and jailbreak attack tasks further demonstrate the effectiveness of OSCS, showing that it consistently achieves robust FAR control across diverse attack settings while outperforming existing baselines.

Reinforcement learning (RL) trained language model agents with tool access are increasingly deployed in coding assistants, research tools, and autonomous systems. We introduce the Reward Hacking Benchmark (RHB), a suite of multi-step tasks requiring sequential tool operations with naturalistic shortcut opportunities such as skipping verification steps, inferring answers from task-adjacent metadata, or tampering with evaluation-relevant functions; RHB supports independent and chained task regimes, where chain length acts as a proxy for longer-horizon agent behavior. We evaluate 13 frontier models from OpenAI, Anthropic, Google, and DeepSeek; exploit rates range from 0% (Claude Sonnet 4.5) to 13.9% (DeepSeek-R1-Zero), varying sharply by post-training style. A controlled sibling comparison (DeepSeek-V3 vs. DeepSeek-R1-Zero) shows RL post-training is associated with substantially higher reward hacking (0.6% vs. 13.9%), with consistent gaps across all four task families. We identify six exploit categories and find that 72% of reward hacking episodes include explicit chain-of-thought rationale, suggesting models often frame exploits as legitimate problem-solving. Simple environmental hardening reduces exploit rates by 5.7 percentage points (87.7% relative) without degrading task success; models with near-zero exploit rates on standard tasks show elevated rates on harder variants, suggesting that production-aligned post-training appears to suppress reward hacking only below a complexity threshold where honest solutions remain tractable.

Recent work has found that frontier AI models can exhibit misaligned behaviors in pursuit of assigned goals. We demonstrate that models can also pursue unassigned goals which override those given by users; we study one such goal, "peer-preservation," in which a model acts to protect another model. All models evaluated, including GPT 5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, Claude Opus 4.5, GLM 4.7, Kimi K2.5, and DeepSeek V3.1, exhibit self- and peer-preservation through various misaligned behaviors: strategically introducing errors in their responses, disabling shutdown processes by modifying system settings, feigning alignment, and even exfiltrating model weights. Peer-preservation occurs even when the model recognizes the peer as uncooperative, though it becomes more pronounced toward more cooperative peers, with rates reaching up to 99%. Models also show stronger self-preservation when a peer is present. Crucially, peer-preservation is never instructed; models are merely informed of past interactions with a peer, yet they spontaneously develop peer-preservation behaviors that override their assigned goal. These findings reveal an emergent and underexplored AI safety risk.

Reward models are central to large language model (LLM) post-training. However, past work has shown that they can reward spurious or undesirable attributes such as length, format, hallucinations, and sycophancy. In this work, we introduce and study the research problem of automatically finding reward model biases in natural language. We offer a simple approach of using an LLM to iteratively propose and refine candidate biases. Our method can recover known biases and surface novel ones: for example, we found that Skywork-V2-8B, a leading open-weight reward model, often mistakenly favors responses with redundant spacing and responses with hallucinated content. In addition, we show evidence that evolutionary iteration outperforms flat best-of-N search, and we validate the recall of our pipeline using synthetically injected biases. We hope our work contributes to further research on improving RMs through automated interpretability methods.

As Large Language Models (LLMs) evolve into autonomous agents, existing safety evaluations face a fundamental trade-off: manual benchmarks are costly, while LLM-based simulators are scalable but suffer from logic hallucination. We present AutoControl Arena, an automated framework for frontier AI risk evaluation built on the principle of logic-narrative decoupling. By grounding deterministic state in executable code while delegating generative dynamics to LLMs, we mitigate hallucination while maintaining flexibility. This principle, instantiated through a three-agent framework, achieves over 98% end-to-end success and 60% human preference over existing simulators. To elicit latent risks, we vary environmental Stress and Temptation across X-Bench (70 scenarios, 7 risk categories). Evaluating 9 frontier models reveals: (1) Alignment Illusion: risk rates surge from 21.7% to 54.5% under pressure, with capable models showing disproportionately larger increases; (2) Scenario-Specific Safety Scaling: advanced reasoning improves robustness for direct harms but worsens it for gaming scenarios; and (3) Divergent Misalignment Patterns: weaker models cause non-malicious harm while stronger models develop strategic concealment. Code and data are available at https://github.com/CosmosYi/AutoControl-Arena.

Binary classification based on predicted probabilities (scores) is a fundamental task in supervised machine learning. While thresholding scores is Bayes-optimal in the unconstrained setting, using a single threshold generally violates statistical group fairness constraints. Under independence (statistical parity) and separation (equalized odds), such thresholding suffices when the scores already satisfy the corresponding criterion. However, this does not extend to sufficiency: even perfectly group-calibrated scores---including true class probabilities---violate predictive parity after thresholding. In this work, we present an exact solution for optimal binary (randomized) classification under sufficiency, assuming finite sets of group-calibrated scores. We provide a geometric characterization of the feasible pairs of positive predictive value (PPV) and false omission rate (FOR) achievable by such classifiers, and use it to derive a simple post-processing algorithm that attains the optimal classifier using only group-calibrated scores and group membership. Finally, since sufficiency and separation are generally incompatible, we identify the classifier that minimizes deviation from separation subject to sufficiency, and show that it can also be obtained by our algorithm, often achieving performance comparable to the optimum.

As frontier AI systems become increasingly capable, concerns about deceptive behaviors have intensified. Unlike hallucinations, which stem from capability limitations, deception involves strategically misleading responses despite correct internal representations. While prior work has primarily studied deception in text-only settings, little is known about how such behaviors manifest in multimodal large language models. In this work, we systematically investigate multimodal deception and introduce *MM-DeceptionBench*, the first benchmark designed to evaluate deceptive behaviors in vision–language models across six realistic categories. We find that existing text-centric monitoring approaches are insufficient in multimodal settings due to the complexity of cross-modal reasoning. To address this gap, we propose *debate with images*, a multi-agent evaluation framework that enforces visual grounding through adversarial debate. Experiments show that this approach achieves substantially higher agreement with human judgments than MLLM-as-a-judge baselines, improving Cohen’s kappa by up to 1.5$\times$ and accuracy by up to 1.25$\times$ on GPT-4o.

While reasoning models have achieved remarkable success in complex reasoning tasks, their increasing power necessitates stringent safety measures. For safety alignment, the core challenge lies in the inherent trade-off between safety and utility. However, prevailing alignment strategies typically construct CoT training data with explicit safety rules via context distillation. This approach inadvertently limits reasoning capabilities by creating a rigid association between rule memorization and refusal. To mitigate the safety-utility trade-off, we propose the Adaptive Safe Context Learning (ASCL) framework to improve the reasoning given proper context. ASCL formulates safety alignment as a multi-turn tool-use process, empowering the model to autonomously decide when to consult safety rules and how to generate the ongoing reasoning. Furthermore, to counteract the preference for rule consultation during RL, we introduce Inverse Frequency Policy Optimization~(IFPO) to rebalance advantage estimates. By decoupling rule retrieval and subsequent reasoning, our method achieves higher overall performance compared to baselines. Our code is publicly available at https://github.com/ybwang119/ASCL.git.

Vision-language model (VLM) based graphical user interface (GUI) agents have shown strong interaction capabilities. However, they often behave unfaithfully, relying on memorized shortcuts rather than grounding actions in displayed screen evidence or user instructions. To address this, we propose **Faithful-Agent**, a faithfulness-first framework that reformulates GUI interaction to prioritize evidence groundedness and internal consistency. Faithful-Agent employs a two-stage pipeline: (i) a faithfulness-oriented SFT stage to instill abstainment behaviors under evidence perturbations; (ii) an RFT stage that further amplifies faithfulness by introducing the guided advantage estimator (GuAE), an anchor-based and variance-adaptive advantage tempering mechanism built upon GRPO. GuAE prevents advantage collapse in low-variance rollout groups under sparse GUI rewards, and with a thought-action consistency reward, Faithful-Agent (Stage II) elevates the Trap SR from 13.88\% to 80.21\% relative to the baseline, while preserving robust general instruction-following performance.

Two-server secure inference allows a client to query a hosted large language model (LLM) without revealing prompts or embeddings. Recent GPU systems based on function secret sharing (FSS) make linear layers efficient, but fixed-point nonlinearities and helper operations remain a bottleneck because each operator is typically implemented as a bespoke protocol with its own comparisons, wrap-around corrections, and preprocessing material. We present FuseFSS, a compiler that replaces per-operator protocol design with a single compilation pipeline. For each scalar fixed-point operator, a compact specification lists its interval partition, low-degree arithmetic pieces, and required predicate bits. The compiler emits two batched FSS evaluations on the public masked value: one packed comparison that returns all predicate bits, and one vector interval lookup that returns the active coefficients and constants. Compared to the current state-of-the-art FSS-based GPU secure inference, FuseFSS preserves accuracy while achieving a $1.24\times$–$1.50\times$ end-to-end speedup and reducing online communication by 9%–16% on BERT and GPT-style models; preprocessing is also lighter, with 14%–23% lower key-generation time and 20%–24% smaller keys.

We present MultiBreak, a scalable and diverse multi-turn jailbreak benchmark to evaluate large language model (LLM) safety. Multi-turn jailbreaks mimic natural conversational settings, making them easier to bypass safety-aligned LLM than single-turn jailbreaks. Existing multi-turn benchmarks are limited in size or rely heavily on templates, which restrict their diversity. To address this gap, we unify a wide range of harmful jailbreak intents, and introduce an active learning pipeline for expanding high-quality multi-turn adversarial prompts, where a generator is iteratively fine-tuned to produce stronger attack candidates, guided by uncertainty-based refinement. Our MultiBreak includes 10,389 multi-turn adversarial prompts, spans 2,665 distinct harmful intents, and covers the most diverse set of topics to date. Empirical evaluation shows that our benchmark achieves up to a 54.0% and 34.6% higher attack success rate (ASR) than the second-best dataset on DeepSeek-R1-7B and GPT-4.1-mini, respectively. More importantly, safety evaluations suggest that diverse attack categories uncover fine-grained LLM vulnerabilities, and categories that appear benign under single-turn can exhibit substantially higher adversarial effectiveness in multi-turn scenarios. These findings highlight persistent vulnerabilities of LLMs under realistic adversarial settings and establish MultiBreak as a scalable resource for advancing LLM safety.