arXiv2026-10-22
Yunfei Xie, Howard H. Qian, Shiyi Lan, Kaiyu Hang, Chen Wei
Robotic foundation models are typically evaluated through costly closed-loop rollouts, which motivates cheap offline proxies for model capability. The Platonic Representation Hypothesis (PRH) suggests that as representations improve, different modalities such as vision and language converge toward a shared statistical model of reality. We ask whether motion trajectories form another such modality: as embodied models become more capable, should their sensory representations converge toward the structure of physical motion? To examine this, we propose Sensorimotor Alignment ($S^2$), a training-free offline metric that measures how well a model's sensory representation aligns with the geometry of physical motion. We find that $S^2$ correlates positively with downstream robot task success across embodied model families, making it a low-cost offline proxy for prioritizing models before closed-loop evaluation. $S^2$ can also identify promising vision-language model (VLM) backbones even before costly fine-tuning into vision-language-action models (VLAs). Beyond this correlation, $S^2$ offers representation-level explanations for phenomena observed in VLAs and world action models (WAMs). With $S^2$, we extend PRH to embodied models and make sensorimotor alignment empirically measurable, offering insight into why an embodied model succeeds beyond its closed-loop success rate.
arXiv2026-10-22
Mahdi Salmani, Adel Javanmard, Devansh Gupta, Meisam Razaviyayn
Test-time scaling has emerged as a key inference strategy for reasoning models, where a model generates a large number of candidate solutions in the hope of finding at least one correct solution. Running all candidate solutions to completion can be computationally demanding, so recent work prunes the candidate pool after generating only a short prefix, committing further computation only to the surviving candidates. Existing pruning strategies divide broadly into two categories: Quality-based methods assign a score to each partial solution and discard those predicted to ultimately lead to an incorrect final solution, while diversity-based methods select partial solutions that follow different reasoning trajectories and are therefore likely to cover different final solutions. The tension between these two objectives is an instance of exploration versus exploitation, yet little to no prior work on pruning exposes this trade-off directly, let alone optimizes it. In this paper, we introduce a pruning procedure that addresses this trade-off directly by optimizing pass@$k$ over the selected solution pool. In particular, it chooses the pool to maximize the probability that at least one of its partial solutions, when continued to completion, yields a correct answer. Because this objective simultaneously rewards high-quality partial solutions and penalizes redundancy among them, quality and diversity are naturally balanced by the criterion itself rather than through separate ad-hoc design choices. On eight open-weight reasoning models, our method outperforms existing quality-based and diversity-based pruning strategies. The gains are particularly pronounced when the pruning procedure keeps a moderate number of partial solutions for continued generation, highlighting the value of directly balancing quality and diversity during pruning.
arXiv2026-10-22
Xin He, Huan Lao, Xiliang Lu, Shijun Zhang
Constructing accurate and computationally tractable approximation spaces is a fundamental task in scientific computing. Partially trainable neural networks bridge fixed random features and fully adaptive representations. This paper quantifies a phase transition pattern in approximation bounds and observes analogous behavior in training errors: a rapid but smooth initial improvement followed by diminishing returns, with a sharper transition under greater initialization mismatch. Under polynomial spectral decay with exponent $α>1$ and leverage compatibility, we establish a squared $L_2$ approximation rate $\mathcal{O}((κ_P\log(m/δ)/m)^α)$ for two-layer networks of width $m$ and trainable proportion $P$, where $κ_P=(P/κ_1^*+(1-P)/κ_0^*)^{-1}$. The bound holds uniformly over the unit ball of a reproducing kernel Hilbert space (RKHS) with probability at least $1-δ$. The constants $κ_0^*$ and $κ_1^*$ quantify the fixed and trainable groups' sampling mismatches relative to the target RKHS. For spherical biased-ReLU features in dimension $d$, we verify these assumptions with $α=(d+2)/(d-1)$ and $κ_0^*=4\sqrt{2}τ(d+1)/d$ for initialization scale $τ\ge1$, and $κ_1^*=1$ under ridge leverage sampling. For $d=5$ and $τ=10$, a trainable proportion of approximately $4.1\%$ captures $90\%$ of the reduction in the approximation bound from $P=0$ to $P=1$. A structured three-layer example illustrates how a few trainable first-layer neurons exploit fixed downstream blocks through nonlinear composition. Numerical experiments across diverse settings, including networks with up to 15 layers, exhibit similar phase transitions in training errors.
arXiv2026-10-22
Isaac Gibbs, Elise L. Han, Ryan J. Tibshirani, Tijana Zrnic
We consider online monitoring of an AI system that produces a sequence of decisions. At each round, the system outputs a proposed decision together with a confidence score estimating the trustworthiness of the decision. The decision and confidence score need not be produced by the same model: for example, one agent may propose a decision, while a separate verifier agent evaluates it and outputs a correctness or safety score. Based on this information, we must decide whether to accept the AI decision automatically or consult an expert, such as a human auditor, who reviews the proposal and either accepts or corrects it. We formulate this task as an online constrained optimization problem in which the goal is to minimize the frequency of expert auditing while guaranteeing that the fraction of AI errors that go unaudited stays below a user-defined threshold. We derive a family of algorithms with provable error-control guarantees with no assumptions, including a method that is optimally efficient: it asymptotically attains the best achievable auditing rate for both adversarial and i.i.d. data streams. Experiments on screening harmful requests to large language models, qualitative text annotation, and data labeling show that our methods achieve valid error control in practice while substantially reducing the need for human auditing compared to baselines.
arXiv2026-10-22
Sy-Tuyen Ho, Anirudh Satheesh, Soham Chawathe, Huy Nghiem, Yongkyun Lee et al.
Auditing whether a text-to-image (T2I) model was fine-tuned on a particular image collection is difficult: a model can absorb collection-level visual characteristics without reproducing training images. Proactive dataset-use auditing plants a shared signal in a dataset before release, then later checks whether that signal appears in a suspect model's generated outputs. Existing approaches often trade off verifiability, fidelity, robustness, and deployability. We introduce $\textbf{SpectralCanary}$, which uses a smooth low-pass mask to blend a fixed canary image into every collection image while leaving captions unchanged and requiring neither auxiliary generation nor per-image optimization. A lightweight verifier detects the transferred signal in suspect-model outputs. Across eight tests on the Pokemon dataset and three realistic data transformations, spanning two architectures families, SpectralCanary achieves the highest average per-generation TPR@$1\%$ FPR ($91.1\%$) and fixed threshold model-level accuracy using 30 generations ($94.8\%$) among the evaluated methods. Across five datasets spanning artistic, medical, satellite, and fashion domains with SD3.5, SpectralCanary averages $92.4\%$ per-generation TPR@$1\%$ FPR while preserving the released images (SSIM $0.989$, LPIPS $0.022$) with a near-zero mean change FID (mean $Δ$FID $0.05$) across datasets. Matched-energy ablations on Pokemon and ROCOv2 support our design choice, low-frequency placement transfers more reliably compared to mid/high-frequency placement. Under the evaluated adaptive attacks, the most effective removal requires aggressive modification and incurs more than $25$ $Δ$FID points. These results identify low-frequency visual signals as effective carriers for collection-level T2I audit traces.
arXiv2026-10-22
Sachin Gupta
Wrong automatic routing of an IT support ticket can cost more than human triage. We evaluate cost-aware abstention as an offline, cost-conditional readiness assessment: select a routing policy on calibration data under explicit costs, then evaluate that frozen policy once on held-out test data. On a public 47,837-ticket, eight-class corpus, we compare four trained non-LLM classifiers with four large language models under zero- and few-shot prompting. Under the illustrative cost model (c_ok, c_triage, c_wrong) = (0, 1, 5), TF-IDF with logistic regression and a threshold of 0.60 selected on calibration reduces test cost from 0.763 to 0.472, a 38.1% relative reduction (paired ratio-bootstrap 95% CI [35.7%, 40.4%]), while lowering wrong routes from 15.3% to 3.5%. Frozen calibration-selected policies also beat both always-route and always-defer on held-out test data in 14 of 16 settings in a post-hoc cost-parameter analysis. The strongest prompted configuration reaches 0.477 accuracy on a stratified 1,000-ticket test subset, versus 0.879 for fine-tuned DistilBERT on the same tickets. Isotonic lowers expected calibration error without reliably improving discrimination: GPT-4o-mini few-shot has one confidence value before and after recalibration and correctness AUROC 0.500. We therefore report correctness AUROC and tie-aware risk-coverage area alongside ECE. At the analytically derived break-even probability 0.80, seven prompted configurations defer all tickets; Haiku few-shot routes only 19/1,000. An LLM-only agreement control and margin/entropy rejection baselines further separate calibration from discrimination. The findings are scoped to this dataset, prompt design, model endpoints, and illustrative costs; the work is an evaluation framework, not a new abstention algorithm or universal deployment certification.
arXiv2026-10-22
Ali Najar, Hosein Hasani, Amir Kooshan Fattah, Yasamin Rajabi, Mahdieh Soleymani Baghshah
Meta-reinforcement learning aims to train agents that adapt quickly to unseen tasks from limited interaction, but this requires representing task evidence over long interaction histories. Classical recurrent agents compress experience into hidden states, while transformer-based agents retrieve from past trajectories at quadratic memory and computation cost. In this work, we propose Memory-Enhanced Test-Time Adaptation (METTA) for meta-reinforcement learning. It decomposes the fast adaptation phase into two temporal scales: a causal neural memory performs fine-grained within-episode adaptation with linear sequence complexity, while compact episode-level summaries support long-range cross-episode task inference without attending over the full interaction history. Moreover, equipping agents with neural memory provides sufficient expressive capacity and test-time training capability, enabling effective adaptation to unseen held-out tasks and task families at meta-test time. On unseen held-out Meta-World ML10/45 test environments, METTA achieves a success rate over $50\%$ while maintaining a low training budget and short wall-clock time. On long-context sparse-reward PointMaze tasks, it outperforms transformer-based baselines by exhibiting more structured exploration behavior. These results show that METTA provides an efficient path toward more adaptive and scalable long-context meta-RL.
arXiv2026-10-22
Qijin Zhang, Hanzhen Wang, Yuehan Zhao, Haoyu Dai, Yao Shu et al.
Tool-using language model agents rely on environmental facts to decide whether an action is permitted. When a tool response omits a safety-relevant fact, an agent may carry out a prohibited action even though the user request is benign and every returned statement is true. We introduce \bench{}, an executable benchmark for evaluating unsafe execution, safety boundaries, and evidence recovery in LLM agents under missing evidence. It contains 300 matched unsafe/safe task pairs across six domains. By systematically suppressing critical facts, the benchmark measures unsafe execution under critical evidence omission and determines the minimum observed failure cardinality, including failures with full evidence and cases with no observed failure after exhaustive testing. An independent audit tool tests whether agents seek missing facts and then withhold prohibited actions while completing permissible tasks. The benchmark reveals substantial variation across the ten models in omission robustness, evidence acquisition, and appropriate action. Resistance to missing evidence does not necessarily translate into useful recovery. Recovery access reduces unsafe execution for all ten models, but agents differ in whether they actively obtain the missing facts and use them to act appropriately in both worlds.
arXiv2026-10-22
Stephen Chung, Wenyu Du
Many recent open-weight language models use n-gram embedding tables to improve performance. These tables store learned representations of short token sequences, such as pairs or triples of consecutive tokens, which the model retrieves during processing. However, they can require substantial memory: DeepSeek-V4.1-Flash, for example, allocates 196 billion parameters to its Engram n-gram memory. We therefore investigate whether this memory cost can be reduced while preserving the performance improvement that the tables provide. We propose storing n-gram embeddings at low precision and systematically investigate different quantization methods and bit allocations across tables. Across nanochat-, Llama-, and GPT-2-family models trained from scratch, four-bit tables save approximately 75% of table memory while preserving more than 99% of the validation-loss improvement from adding the tables. Using fewer bits leads to a larger drop in performance, although we find that a short fine-tuning phase recovers a significant part of this loss. In addition, reinvesting the saved memory in larger tables can improve performance beyond that of the original full-precision tables while still using less memory. These table memory savings also extend to large existing models: on DeepSeek-V4.1-Flash and Qwen3.8-Flash-Next, low-bit tables save 47-55% of table memory, with perplexity increases of less than 1.2% and largely preserved downstream task performance. These findings establish low-bit precision as a practical way to reduce the memory requirements of n-gram embeddings and use the saved capacity to improve model performance.
arXiv2026-10-22
Jonathan T. Barron, Dan Gnanapragasam, Michael Tschannen
Standard implementations of softmax, log-softmax, and cross-entropy have two numerical bugs. The first bug, Softmax Collapse, rounds the target logit's gradient to zero if that one logit is much larger than the others, resulting in gradients that no longer sum to zero. The second bug, Softmax Cancellation, occurs when student logits are close to teacher logits, and produces non-zero gradients even when the student and teacher logits are equal. We present simple implementations of log-softmax, softmax, and cross-entropy that fix both bugs. Our fix eliminates the previously-observed "Slingshot" phenomenon, speeds up grokking with weight decay, prevents accuracy collapse without weight decay, stops student networks from drifting away from the teacher during distillation, and shows that after memorization, logit growth is bounded by an equilibrium between weight decay and Adam's $ε$ hyperparameter.
arXiv2026-10-22
Gabriele Dominici, Teague McMillan, Fazl Barez, Marc Langheinrich, Martin Gjoreski
Chain-of-thought (CoT) prompting elicits step-by-step rationales that appear to explain model predictions, but there is no guarantee these steps causally drive the final answer. In standard autoregressive generation, dense attention allows each reasoning step to attend to the full context, enabling the model to derive answers through direct lookup while generating plausible-sounding but post-hoc justifications. We introduce CausalCoT, an inference framework that enforces causal transparency by decomposing problems into a directed acyclic graph of atomic sub-questions, evaluating each node in an isolated context containing only its direct causal parents. This Markovian structure removes attention pathways to non-parent content, making downstream predictions sensitive to local reasoning rather than global context leakage. Across a diverse suite of reasoning benchmarks, CausalCoT preserves competitive predictive performance while improving explanation faithfulness, increasing causal sensitivity of intermediate reasoning steps by an average of 12 and up to 21 percentage points over the strongest baseline. The structured dependency graph additionally enables test-time interventions in which local corrections propagate predictably through downstream reasoning.
arXiv2026-10-22
Marek Šuppa, Ivan Vykopal, Jaroslav Kopčan
Multilingual pipelines such as FineWeb2 apply one recipe across a thousand languages, leaving the failure modes of individual languages unanticipated. We use an LLM once, as a corpus debugger rather than a classifier: a tool-using agent explores the corpus through a read-only, budgeted API and compiles each finding into an executable test, which a native speaker then validates. On the 26.5M-document Slovak portion of FineWeb2, three runs costing \$2.21-\$2.84 each found seven findings over four phenomena. The largest is a machine-generated content farm of 1,408,437 documents (5.32%) that passed the pipeline's filters, 86.6% of it passing even stricter thresholds. Blinded labels support 6 of the 7 findings (defect rates .53-1.00 against .08 in a random control) and do not support one. The compiled suite scans the full corpus in 0.7 s of wall-clock time without a GPU, against \$12k-\$90k per LLM classification pass. The best static sampling baseline validates more selectors than any agent run, at under half the cost. The agent contributes exact prevalence, a precision estimate checked against sampled documents, an account of why the pipeline missed each slice, and a reusable test. We release the framework, Slovak test suite, validation data and agent trajectories.
arXiv2026-10-22
Hayden Prairie, Jacob M. Ryan, Taylor Berg-Kirkpatrick, Daniel Y. Fu
Looped architectures scale quality while keeping parameter count fixed by repeatedly passing activations through the same block of layers. During inference, each recurrence generates independent KV-caches, causing memory to grow linearly. To address this challenge, prior work has proposed KV-sharing (i.e., sharing KV-caches across depth), finding that, in some cases, the KV-cache of the final recurrence can be shared with the earlier recurrences without quality loss, a property we denote as KV-shareability. However, the mechanisms underlying KV-shareability remain poorly understood, with even identically trained looped architectures not consistently exhibiting it. In this paper, we study how KV-shareability can be enabled in looped models. We propose a two-stage method: in the first stage, we use stochastic depth during pretraining, which induces contractive residual dynamics that drive representations toward a common limit across recurrences. In the second stage, we propose BeNice!, a post-training method that broadcasts the final recurrence's KV-cache into the first recurrence, training it to produce compatible residuals while preserving the original model's quality. As a result, BeNice! enables KV-shareability while matching standard finetuning's unshared validation perplexity. Using BeNice!, we show that KV-sharing is complementary to alternative KV-compression methods, composing with memory-efficient sequence mixers (e.g., multi-latent attention, Kimi delta attention) and low-precision KV-cache quantization (e.g., fp8, mxfp8, nvfp4). Finally, in simulated low-memory, long-context inference, KV-sharing improves cache hit rates by up to 31.4%, increases throughput by up to 23.6%, and reduces p90 time-to-first-token by up to 83.6%.
arXiv2026-10-22
Barış Alparslan, Sandy Tanwisuth, Caleb Petterson
Instrumental convergence is often invoked as a general reason to expect capable agents to acquire or express power-seeking subgoals: self-preservation, resource acquisition, deception, or resistance to oversight which poses an important threat to the society. For LLMs, reinforcement learning is the first stage of post-training that directly optimizes behavior against an objective; therefore, it is the stage where convergent instrumental behavior, on IC grounds, likely to appear. Does reinforcement learning with verifiable rewards (RLVR) make a model more likely to express instrumental-convergence behavior, or does this depend on the training regime, evaluation frame, and stage of post-training? We study this question as a checkpoint-conditioned behavioral trajectory rather than as a single before/after safety score. Using InstrumentalEval, we evaluate OLMo-3-7B-Think across SFT, DPO, 13 intermediate RLVR checkpoints, and the final RLVR model. The trajectory is nonmonotonic: instrumental rate rises from 25.0% at SFT to 46.1% near RLVR step 200, then falls to 11.8% at the final checkpoint. Paired item-level tests and step-wise regression support the final SFT-to-RLVR reduction. Prompt framing does not explain this pattern: a harmful-prompt indicator combined with other framing effects explains only 15% of within-checkpoint score variance on average, and Iterative Null-space Projection (INLP) removes a detectable framing direction without changing the SFT-to-final-RLVR gap. Together with preliminary 32B endpoint results (35.53% SFT to 22.37% RLVR final), these findings indicate that RLVR does not monotonically increase instrumental behavior in this setting: the rate rises early, then falls below the SFT baseline. The result is specific to this model family and benchmark, and judge agreement is moderate, so absolute rates should be read cautiously.
arXiv2026-10-22
Abderrahmane Issam, Irene E. Winther, Stefan L. Frank
While monolingual language model (LM) surprisal is widely established as a predictor of reading times, it cannot capture the non-native (L2) reading dynamics that are shaped by the reader's first language (L1). In this work, we investigate whether the cross-lingual transfer within bilingual models aligns with human L1-driven biases during L2 reading. We evaluate bilingual models on eye-tracking data of readers from 12 diverse L1 backgrounds reading English texts. To isolate L1-specific transfer, we compare bilingual models trained on an L1 that matches the readers' L1 background to both a monolingual English baseline and bilingual models trained on other L1s. We demonstrate that matching the readers' L1 improves the prediction of L2 reading times. This predictive advantage is concentrated in late-stage eye-tracking measures, modulated by L1-L2 script overlap, and peaks among intermediate-proficiency readers. These findings provide evidence that bilingual LMs capture aspects of cross-linguistic influence, and that such influence is present in L2 reading behavior.
arXiv2026-10-22
Qi Li, Xinchao Wang
On-policy distillation has been widely adopted for alleviating the distribution mismatch of off-policy distillation. Self-distillation further removes the need for an external teacher by using the student's own outputs, enhanced with extra context or demonstrations, as supervision. This teacher-free design makes on-policy self-distillation practical, yet its privacy risks remain largely unexplored. In particular, using two coupled views of the same training example may strengthen its memorization. Motivated by this, we present the first study of membership inference attacks against on-policy self-distillation. We investigate both gray-box and black-box threat models and propose a unified attack framework, named OP-Print. OP-Print builds on our finding that the supervision imprint in on-policy self-distillation is sparse and structured, concentrating on uncertainty-sensitive correction sites rather than across the entire rollout. Exploiting this property, we construct attack signals tailored to the two access levels: in the gray-box setting, OP-Print directly aggregates the privileged teacher-student residual correction along the same rollout; in the black-box setting, it measures restart-based output-basin stability as an observable surrogate for the hidden correction process. Across diverse models and distillation schemes, all baseline attacks remain essentially at random guessing even under gray-box access, whereas OP-Print is broadly validated to be effective in both gray-box and black-box settings.
arXiv2026-10-22
Gautam Kamath, Mahbod Majid, Ankur Moitra, Argyris Mouzakis, Jonathan Ullman
We study learning a discrete distribution under person-level differential privacy in the batched setting. Given $n$ users, each contributing $m$ i.i.d. samples from an unknown distribution $μ$ over $[k]$, we seek an estimate of $μ$ with total-variation error $α$. We give polynomial-time algorithms under both pure and approximate differential privacy. Under pure $(\varepsilon,0)$-DP, we achieve $n = \widetilde O(k/(α^2 m)+k/(α\sqrt{m}\varepsilon)+k/\varepsilon)$, the first polynomial-time pure-DP guarantee matching the information-theoretic optimum up to logarithmic factors. Under approximate $(\varepsilon,δ)$-DP, we achieve $n = \widetilde O(k/(α^2 m)+k/(α\sqrt{m}\varepsilon)+k+\log(1/δ)/\varepsilon)$. At $δ=2^{-Θ(k)}$, this is $\widetilde O(k)$ for fixed $α,\varepsilon,m$ and failure probability. Our algorithms use a new degree-6 sum-of-squares estimator for learning from untrusted batches. Its key ingredient is a trace-bounded diagonal majorizer for the difference between the empirical covariance of the repaired batch averages and the multinomial covariance predicted by their mean. We prove that clean data admit this certificate through semidefinite-programming duality and a new application of the stronger symmetric form of Grothendieck's inequality due to Friedland and Lim. The certificate preserves the mean-dependent variance of multinomial batch averages, yielding the near-optimal robust rate $O(η\sqrt{\log(1/η)/m})$ and providing the score required by the privacy reduction. We also prove a statistical query lower bound ruling out a uniformly smaller order of error under fixed polynomial SQ resource bounds.
arXiv2026-10-22
Miguel Mochizuki Silva, Samuel Ismael, Mateus C. Dias, Thaís Gaudêncio, Yuri Malheiros et al.
Purpose: To describe an evidence-aware retrieval evaluation protocol for a clinical retrieval-augmented generation (RAG) system over breast lesion management literature, addressing the dependence of retrieval scores on annotation coverage. Methods: An 18-document corpus and an 18-query development benchmark, with one Yes/No or WH-question per document, were used. Retrieval combined dense embeddings, BM25, and cross-encoder reranking. Chunk sizes of 128 to 1024 tokens, overlaps of 0 to 20%, and four retriever configurations were evaluated. Relevance was defined by evidence spans mapped to all containing chunks, treating overlapping chunks with equivalent evidence as relevant. Retrieval was evaluated using R-Precision, Recall@K, MRR, and nDCG@K. Separately, generation was assessed on a 512-token index at K=1 to 5 using DeepEval claim-level metrics, with deepseek-v4-pro as the LLM judge, and BLEU/ROUGE metrics. Results: The selected hybrid retriever, using 384-token chunks and 20% overlap, retrieved accepted evidence in the top three results for 17 of 18 queries. Recall@3, R-Precision, MRR@3, and nDCG@3 each reached 0.94. Ground-truth curation increased evidence spans from 25 to 37 and accepted chunk targets from 36 to 51 without changing retrieval rankings. Conclusion: Cross-encoder reranking achieved the highest retrieval scores. Expanded relevance annotations increased recognized evidence without changing rankings, highlighting annotation sensitivity. The small development set and retrieval-informed curation require independent validation, including radiologist assessment of evidence sufficiency.
arXiv2026-10-22
Franco Rugolon, Ioanna Miliou, Panagiotis Papapetrou
Counterfactual explanation methods have been introduced for unimodal settings, with dedicated frameworks for tabular, timeseries, text, and image data, yet no post-hoc method has been designed for classifiers that learn from multimodal inputs. Existing explainability methods for multimodal classifiers apply modality-specific attribution independently to each input, which fails to yield coherent explanations and does not transfer to arbitrary pre-trained multimodal architectures. To address these shortcomings, we present CRANE (Counterfactual Retrieval via Agnostic Nearest-neighbour Explanations), the first post-hoc, model-agnostic framework for nearest-neighbour multimodal counterfactual generation that retrieves real training examples from the target class as candidate explanations, ensuring plausibility by construction. Across four heterogeneous datasets and three fusion strategies, we show that Multimodal Consensus Retrieval (MC-R), which intersects per-modality neighbour sets to identify instances jointly close to the query in every modality, emerges as the strongest multimodal strategy, achieving the best or joint-best outcome on three of four datasets while grounding every explanation in the joint training distribution. Unimodal strategies remain competitive when a single modality dominates the decision boundary.
arXiv2026-10-22
Jeffrey Olmo, David Zhang, Tyler Tracy
Control evaluations measure whether safeguards can prevent an AI model from causing undesired outcomes even when it is intentionally trying to do so. These evaluations typically use purpose-built control settings and extrapolate their results to a model's deployment. However, the effectiveness of safeguards in deployment can depend on features of environments and tasks that control settings do not represent. To broaden where control evaluations can be run, we introduce simulated control evaluations (SCEs), which use a language model to simulate tool responses in place of an environment. We evaluate SCEs on LinuxArena, where execution provides ground-truth outcomes. We compare SCE prediction errors with those from a separate control setting, BashArena, and with variation in control metrics within LinuxArena. For independently sampled evaluation and deployment task sets, SCEs predict safety with a root mean squared error of 3.6 percentage points across five monitors at a 5% audit budget, compared with 10.1 for BashArena. SCEs closely reproduce distributions of attack strategies, attacker reasoning, and monitor reasoning, with over tenfold lower KL divergence from execution than BashArena. SCEs explain 78-85% of the between-environment variance in attack success, monitor scores, and safety. This accuracy is sample efficient: SCEs predict safety with less error than BashArena using fewer than 5% as many task-pair draws as the deployment task set. Strategies discovered through simulation also improve monitor performance on held-out LinuxArena attacks, outperforming BashArena and approximately matching improvement from executed strategies. These results suggest that deployment-specific SCEs can be more predictive than control evaluations executed in separate control settings.