Searcharxiv⌕ Search

arXiv subjects

Chih-Hong Cheng

Publications and source records attributed to Chih-Hong Cheng.

At least 19 recordsLinked to original sources

From Ambiguous Tickets to Ranked Hypotheses: Multi-Agent Root-Cause Analysis for Autonomous Heavy-Vehicle Software

Autonomous haulage systems run safety-critical software on driverless mining trucks, with no onboard driver to detect or mitigate abnormal behavior. Every failure in testing or operation must therefore be traced to its root cause to keep the safety case valid and limit downtime. An investigation often starts from a sparse ticket containing an approximate timestamp, sometimes a suspected cause, and a multimodal vehicle recording of several gigabytes. We present a multi-agent root-cause analysis (RCA) workflow that turns this input into a telemetry-grounded event chain, ranked causal hypotheses, and candidate code locations. A suspected cause in the ticket, which could anchor the analysis on a false theory, goes to a single LLM analyst in a separate agent session. Non-neural tools select the relevant telemetry once, before analysis, and an LLM condenses it into shared evidence artifacts; this more than halved the cost per incident compared with per-analyst querying. Specialist analyst agents then apply established RCA techniques in parallel, including fault tree analysis, Kepner-Tregoe analysis, and call-chain tracing, and hypotheses that several LLM analysts reach independently rank higher. We evaluated the workflow on 12 held-out industrial incidents against a single agent with the same model and tools. It located the faulty file more often (10 vs. 8 cases) at higher cost. Two human domain experts accepted a top-three hypothesis in 4 and 9 cases, respectively. The gap, which two LLM judges did not resolve, shows that rating causal explanations is itself a hard evaluation problem.

cs.SE↗

Beyond Attack Success Rate: Cost-Aware Metric for Evaluating Iterative LLM Jailbreaks

Attack success rate (ASR) evaluates jailbreaks against large language models without accounting for repeated calls to the victim model. Yet each additional call gives the deployed system another opportunity to detect the attack and intervene. We introduce $\mathrm{ASR}_{\mathrm{STEP}}$, which penalizes additional victim-model calls while retaining binary success judgments. Our evaluation considers GCG, PAIR with a step budget, and a single-step PAP baseline on JailBreakBench across three model families. GCG's scores decrease by 7.0-8.7 percentage points after discounting, while PAP's remain unchanged. While for PAIR attack, we see a wider discounting range between 3.4-12 percentage points, showing that the "step" quantity differs per iterative strategy. Reporting both metrics makes the discount due to repeated victim-model interaction explicit. Code is available here: https://github.com/SuELS-Group/beyond_asr.

cs.CR↗

AgentMonopoly: Behavioral Patterns and Persona Control in LLM Agents

How do LLM agents' strategies differ, and how much can instructions to adopt a persona change their behavior over a game? We study these questions with AgentMonopoly, an environment where agents communicate, buy and trade properties, and make decisions that affect later turns. We compare models from four families, each at two sizes, in games with identical models, different sizes of the same family, or different families. We also assign one agent a persona, a description of how it should behave, intended to change a tendency observed in games without that persona. The tested models differ in how often they trade, waive rent, and manage cash and properties. Larger models do not consistently win more games. Persona instructions change the words agents use and some of their actions, but the changes do not always match the role across all decisions. These results show why persona following should be assessed through actions and outcomes as well as language. As our findings are conditional on the tested checkpoints, prompts, and five seeds per configuration, the open-source nature of AgentMonopoly provides the experimental infrastructure for further extending these comparisons. Our code is available at https://github.com/SuELS-Group/Agent_Monopoly.

cs.AI↗

Genesis Steering: Better Activation Steering Through Training Data

Activation steering controls large language models at inference time, but its effectiveness depends on the choice of reference data. We introduce genesis steering, which derives a genesis vector from contrastive examples in the model's own fine-tuning data. When these data are unavailable, we draw on the membership inference literature to identify likely training examples and construct a proxy dataset. We establish two theoretical results on optimal displacement to the decision boundary and reduced proxy-estimation bias. We evaluate genesis steering and membership-based proxy construction on Tülu-3-8B, whose fine-tuning data are publicly available, and Llama-3.1-8B-Instruct as a control. We find that safety data from Tülu-3-8B's training mixture yield stronger steering than external datasets on both models. On Tülu-3-8B, membership scores largely agree in identifying training examples, and vectors built from top-ranked examples achieve competitive performance against random subsets of the same size. In contrast, the results on Llama-3.1-8B-Instruct hint that a MIA method will only have a positive effect if the mean direction shift with MIA-filtered in-distribution data is much less severe than the direction shift with OoD data.

cs.LG↗

Concerns in Benchmarking Jailbreaks: An Empirical Study of Attack Success Rates

Jailbreak methods are commonly compared using Attack Success Rate (ASR), but these comparisons depend on how responses are generated and judged. In this paper, we examine eight attacks against three open-weight language models using three evaluation procedures. Low-resource language attacks achieve a mean ASR of 96.0% under keyword matching but 2.7% under a Qwen3 evaluator following StrongREJECT scoring instructions. Changing the number and arrangement of attempts reverses the ranking of DeepInception and rewriting. In a separate experiment, sampling 100 responses for each unchanged request at Ollama's default temperature of 0.8 yields 100% ASR under keyword matching but 50% under the Qwen3 evaluator. Beyond evaluator sensitivity, these findings raise a safety concern: repeated sampling can expose residual unsafe behavior despite an initial refusal, suggesting probabilistic suppression rather than robust exclusion. We recommend explicit success definitions, matched response budgets, validated evaluators, and reporting cumulative failure under repeated interaction rather than relying on single-generation tests.

cs.LG↗

OGAM: Connecting Systematic Testing to Runtime Assurance through Object-Grounded Attention Monitoring for VLA Policies

Benchmarks expose vision-language-action (VLA) policies to few canonical instructions, while exhaustive deployment testing is impossible. We introduce Object-Grounded Attention Monitoring (OGAM), connecting systematic testing to runtime assurance: testing reveals attention divergence between successful and failed executions, and OGAM uses this signal to stop failures beyond the finite suite. We generate scene-grounded instructions through pairwise combinations of action templates and objects, and separately test meaning-preserving paraphrases. All 87 out-of-benchmark cases reveal problematic behavior across OpenVLA, OpenVLA-OFT, UniVLA, and $π_{0.5}$: none completes any of the 24 feasible instructions, while infeasible or hazardous requests also trigger behavior substitution. At each action query, we project gradient-weighted visual attention through object masks and group it by instruction role for comparison across tasks and policies. Dynamic time warping aligns this course with a successful reference despite speed differences; conformal calibration on successful episodes sets the early-stopping threshold for sustained deviations, with a nominal false-stop target of $α=0.05$. Across four policies, OGAM stops 87-100% of failed episodes at median times of 5-12s within a 20s budget, with observed false-stop rates of 3-5%, without failure-labeled training. Finite testing thus identifies attention patterns that support online intervention before failure fully unfolds.

cs.LG↗

A Cross-Architecture Audit of Direction-Based Inference-Time Defences in Vision-Language Models

Inference time defences against vision language model jailbreaks often subtract a calibrated direction from the residual stream at a chosen decoder layer. We compare five defence candidates across 15 model and layer cells from four architectural families under a magnitude controlled protocol that matches the intervention size for each prompt and pairs every direction with a random control of the same norm. The candidates are the mean image conditioning shift, a CMRM style refusal direction, a ShiftDC style attack specific residual, a prompt instruction to ignore the image, and a random control. No single candidate dominates on both refusal recovery and utility preservation. The image conditioning shift leads on LLaVA 1.5 and Pixtral 12B and is the only candidate whose utility loss remains at the measurement noise floor in every family. The prompt instruction leads on Qwen2.5 VL, while the attack specific residual leads on Qwen2 VL 2B. The image conditioning direction is direction specific in 13 of 15 cells, but strongly architecture specific and nontransferable across the only dimension compatible pair, LLaVA 1.5 13B and Pixtral 12B. We also connect text only and multimodal refusal geometry. The CMRM direction has positive cosine alignment with the image conditioning shift in all 15 cells, with mean 0.35, range 0.17 to 0.65, 15 to 25 times the random vector null, and a sign test p value of about 3e-5. These results show that the two recipes recover partially overlapping geometry and that direction based defences should be calibrated separately for each language decoder family.

cs.AI↗

Crossing the Margin Cliff: Toward Relearn-Robust LLM Unlearning via Margin Calibration

Large language model unlearning is consistently fragile under relearn attacks. On TOFU, fine-tuning on twenty forget examples substantially recovers held-out forget-set ROUGE for every method we evaluate, and we trace this fragility to optimization geometry. The per-token answer margin of fourteen post-hoc methods spanning gradient, preference, and distillation families converges into a narrow band above the retain reference in 41 of 42 method--size cells, a regularity we call the margin cliff. We prove that this cliff follows whenever the retain coupling holds the diagnostic log-odds of forget content above a floor, a condition that token-saturating losses induce at stationarity and that we verify directly on 34 of 42 cells. Margin Calibration (\textsc{MC}) is a plug-in polish adding a non-saturating margin hinge anchored at the reference's per-token margin plus a KL probe on a disjoint instruction corpus, restoring forget-side pressure where the native loss saturates. Under a stated gradient-dominance condition, whose on-trajectory gradient signature we measure by instrumenting the polish, its stationary set lies on the cliff-crossing side, yielding an attack-budget upper bound on the relearn margin lift. Across TOFU (three Llama-3 sizes, three forget tiers), MUSE-News on Llama-2-7B-hf, and a Phi-3.5 panel, a single frozen configuration wins all 14 head-to-head forget aggregates and all populated relearn cells (panel-mean post-attack ROUGE-L $0.41$ to $0.18$) and lowers raw membership AUC on 13/14, with reduced retain-side utility as the main cost. A deployment variant matches these gains without a retain-trained reference.

cs.AI↗

ProGRank: Probe-Gradient Reranking to Defend Dense-Retriever RAG from Corpus Poisoning

Retrieval-Augmented Generation (RAG) improves large language model applications by grounding generation in retrieved evidence, but also introduces corpus poisoning as a new attack surface. In this setting, an adversary injects or edits passages so that they enter the Top-$K$ results for target queries and influence downstream generation. Existing defences often rely on content filtering, auxiliary models, or generator-side reasoning, which complicates deployment. We propose ProGRank, a post hoc, training-free retriever-side defence for dense-retriever RAG. ProGRank stress-tests each query--passage pair under mild randomized perturbations, extracts probe gradients from a small fixed parameter subset, and derives two instability signals: representational consistency and dispersion risk. It then combines these signals with a score gate for reranking. ProGRank preserves the original passage content, requires no retraining, and supports a surrogate-based variant when the deployed retriever is unavailable. Experiments across datasets, retrievers, attacks, and retrieval-stage and end-to-end settings show that ProGRank improves robustness and maintains a favorable robustness--utility trade-off, including under adaptive evasive attacks.

cs.AI↗

Safety-Aligned 3D Object Detection: Single-Vehicle, Cooperative, and End-to-End Perspectives

Perception plays a central role in connected and autonomous vehicles (CAVs), underpinning not only conventional modular driving stacks, but also cooperative perception systems and recent end-to-end driving models. While deep learning has greatly improved perception performance, its statistical nature makes perfect predictions difficult to attain. Meanwhile, standard training objectives and evaluation benchmarks treat all perception errors equally, even though only a subset is safety-critical. In this paper, we investigate safety-aligned evaluation and optimization for 3D object detection that explicitly characterize high-impact errors. Building on our previously proposed safety-oriented metric, NDS-USC, and safety-aware loss function, EC-IoU, we make three contributions. First, we present an expanded study of single-vehicle 3D object detection models across diverse neural network architectures and sensing modalities, showing that gains under standard metrics such as mAP and NDS may not translate to safety-oriented criteria represented by NDS-USC. With EC-IoU, we reaffirm the benefit of safety-aware fine-tuning for improving safety-critical detection performance. Second, we conduct an ego-centric, safety-oriented evaluation of AV-infrastructure cooperative object detection models, underscoring its superiority over vehicle-only models and demonstrating a safety impact analysis that illustrates the potential contribution of cooperative models to "Vision Zero." Third, we integrate EC-IoU into SparseDrive and show that safety-aware perception hardening can reduce collision rate by nearly 30% and improve system-level safety directly in an end-to-end perception-to-planning framework. Overall, our results indicate that safety-aligned perception evaluation and optimization offer a practical path toward enhancing CAV safety across single-vehicle, cooperative, and end-to-end autonomy settings.

cs.CV↗

Cumulative Consensus Score: Label-Free and Model-Agnostic Evaluation of Object Detectors in Deployment

Evaluating object detection models in deployment is challenging because ground-truth annotations are rarely available. We introduce the Cumulative Consensus Score (CCS), a label-free monitoring signal for continuous evaluation and comparison of detectors in real-world settings. CCS applies test-time data augmentation to each image and measures the spatial consistency of predicted bounding boxes across augmented views using Intersection over Union. The resulting consensus score serves as a proxy for reliability without requiring bounding box annotations. In controlled experiments on Open Images and KITTI, CCS achieved over 90% congruence with F1-score, Probabilistic Detection Quality, and Optimal Correction Cost, with qualitative consistency further confirmed on COCO and BDD100K across model pairs. The method is model-agnostic, working across single-stage and two-stage detectors, and operates at the case level to highlight under-performing scenarios. We also provide a simplified theoretical link between expected CCS and detection correctness. Altogether, CCS provides a robust foundation for DevOps-style monitoring of object detectors.

cs.CV↗

Workflow-Level Design Principles for Trustworthy GenAI in Automotive System Engineering

The adoption of large language models in safety-critical system engineering is constrained by trustworthiness, traceability, and alignment with established verification practices. We propose workflow-level design principles for trustworthy GenAI integration and demonstrate them in an end-to-end automotive pipeline, from requirement delta identification to SysML v2 architecture update and re-testing. First, we show that monolithic ("big-bang") prompting misses critical changes in large specifications, while section-wise decomposition with diversity sampling and lightweight NLP sanity checks improves completeness and correctness. Then, we propagate requirement deltas into SysML v2 models and validate updates via compilation and static analysis. Additionally, we ensure traceable regression testing by generating test cases through explicit mappings from specification variables to architectural ports and states, providing practical safeguards for GenAI used in safety-critical automotive engineering.

cs.SE↗

Quantifying Fidelity: A Decisive Feature Approach to Comparing Synthetic and Real Imagery

Virtual testing using synthetic data has become a cornerstone of autonomous vehicle (AV) safety assurance. Despite progress in improving visual realism through advanced simulators and generative AI, recent studies reveal that pixel-level fidelity alone does not ensure reliable transfer from simulation to the real world. What truly matters is whether the system-under-test (SUT) bases its decisions on consistent decision evidence in both real and simulated environments, not just whether images "look real" to humans. To this end this paper proposes a behavior-grounded fidelity measure by introducing Decisive Feature Fidelity (DFF), a new SUT-specific metric that extends the existing fidelity spectrum to capture mechanism parity, that is, agreement in the model-specific decisive evidence that drives the SUT's decisions across domains. DFF leverages explainable-AI methods to identify and compare the decisive features driving the SUT's outputs for matched real-synthetic pairs. We further propose estimators based on counterfactual explanations, along with a DFF-guided calibration scheme to enhance simulator fidelity. Experiments on 2126 matched KITTI-VirtualKITTI2 pairs demonstrate that DFF reveals discrepancies overlooked by conventional output-value fidelity. Furthermore, results show that DFF-guided calibration improves decisive-feature and input-level fidelity without sacrificing output value fidelity across diverse SUTs.

cs.AI↗

Validating Generalist Robots with Situation Calculus and STL Falsification

Generalist robots are becoming a reality, capable of interpreting natural language instructions and executing diverse operations. However, their validation remains challenging because each task induces its own operational context and correctness specification, exceeding the assumptions of traditional validation methods. We propose a two-layer validation framework that combines abstract reasoning with concrete system falsification. At the abstract layer, situation calculus models the world and derives weakest preconditions, enabling constraint-aware combinatorial testing to systematically generate diverse, semantically valid world-task configurations with controllable coverage strength. At the concrete layer, these configurations are instantiated for simulation-based falsification with STL monitoring. Experiments on tabletop manipulation tasks show that our framework effectively uncovers failure cases in the NVIDIA GR00T controller, demonstrating its promise for validating general-purpose robot autonomy.

cs.RO↗

Revisiting Out-of-Distribution Detection in Real-time Object Detection: From Benchmark Pitfalls to a New Mitigation Paradigm

Out-of-distribution (OoD) inputs pose a persistent challenge to deep learning models, often triggering overconfident predictions on non-target objects. While prior work has primarily focused on refining scoring functions and adjusting test-time thresholds, such algorithmic improvements offer only incremental gains. We argue that a rethinking of the entire development lifecycle is needed to mitigate these risks effectively. This work addresses two overlooked dimensions of OoD detection in object detection. First, we reveal fundamental flaws in widely used evaluation benchmarks: contrary to their design intent, up to 13% of objects in the OoD test sets actually belong to in-distribution classes, and vice versa. These quality issues severely distort the reported performance of existing methods and contribute to their high false positive rates. Second, we introduce a novel training-time mitigation paradigm that operates independently of external OoD detectors. Instead of relying solely on post-hoc scoring, we fine-tune the detector using a carefully synthesized OoD dataset that semantically resembles in-distribution objects. This process shapes a defensive decision boundary by suppressing objectness on OoD objects, leading to a 91% reduction in hallucination error of a YOLO model on BDD-100K. Our methodology generalizes across detection paradigms such as YOLO, Faster R-CNN, and RT-DETR, and supports few-shot adaptation. Together, these contributions offer a principled and effective way to reduce OoD-induced hallucination in object detectors. Code and data are available at: https://gricad-gitlab.univ-grenoble-alpes.fr/dnn-safety/m-hood.

cs.CV↗

Randomized Smoothing Meets Vision-Language Models

Randomized smoothing (RS) is one of the prominent techniques to ensure the correctness of machine learning models, where point-wise robustness certificates can be derived analytically. While RS is well understood for classification, its application to generative models is unclear, since their outputs are sequences rather than labels. We resolve this by connecting generative outputs to an oracle classification task and showing that RS can still be enabled: the final response can be classified as a discrete action (e.g., service-robot commands in VLAs), as harmful vs. harmless (content moderation or toxicity detection in VLMs), or even applying oracles to cluster answers into semantically equivalent ones. Provided that the error rate for the oracle classifier comparison is bounded, we develop the theory that associates the number of samples with the corresponding robustness radius. We further derive improved scaling laws analytically relating the certified radius and accuracy to the number of samples, showing that the earlier result of 2 to 3 orders of magnitude fewer samples sufficing with minimal loss remains valid even under weaker assumptions. Together, these advances make robustness certification both well-defined and computationally feasible for state-of-the-art VLMs, as validated against recent jailbreak-style adversarial attacks.

cs.LG↗

Runtime Monitoring and Enforcement of Conditional Fairness in Generative AIs

The deployment of generative AI (GenAI) models raises significant fairness concerns, addressed in this paper through novel characterization and enforcement techniques specific to GenAI. Unlike standard AI performing specific tasks, GenAI's broad functionality requires ``conditional fairness'' tailored to the context being generated, such as demographic fairness in generating images of poor people versus successful business leaders. We define two fairness levels: the first evaluates fairness in generated outputs, independent of prompts and models; the second assesses inherent fairness with neutral prompts. Given the complexity of GenAI and challenges in fairness specifications, we focus on bounding the worst case, considering a GenAI system unfair if the distance between appearances of a specific group exceeds preset thresholds. We also explore combinatorial testing for assessing relative completeness in intersectional fairness. By bounding the worst case, we develop a prompt injection scheme within an agent-based framework to enforce conditional fairness with minimal intervention, validated on state-of-the-art GenAI systems.

cs.LG↗

Trustworthy Text-to-Image Diffusion Models: A Timely and Focused Survey

Text-to-Image (T2I) Diffusion Models (DMs) have garnered widespread attention for their impressive advancements in image generation. However, their growing popularity has raised ethical and social concerns related to key non-functional properties of trustworthiness, such as robustness, fairness, security, privacy, factuality, and explainability, similar to those in traditional deep learning (DL) tasks. Conventional approaches for studying trustworthiness in DL tasks often fall short due to the unique characteristics of T2I DMs, e.g., the multi-modal nature. Given the challenge, recent efforts have been made to develop new methods for investigating trustworthiness in T2I DMs via various means, including falsification, enhancement, verification \& validation and assessment. However, there is a notable lack of in-depth analysis concerning those non-functional properties and means. In this survey, we provide a timely and focused review of the literature on trustworthy T2I DMs, covering a concise-structured taxonomy from the perspectives of property, means, benchmarks and applications. Our review begins with an introduction to essential preliminaries of T2I DMs, and then we summarise key definitions/metrics specific to T2I tasks and analyses the means proposed in recent literature based on these definitions/metrics. Additionally, we review benchmarks and domain applications of T2I DMs. Finally, we highlight the gaps in current research, discuss the limitations of existing methods, and propose future research directions to advance the development of trustworthy T2I DMs. Furthermore, we keep up-to-date updates in this field to track the latest developments and maintain our GitHub repository at: https://github.com/wellzline/Trustworthy_T2I_DMs

cs.LG↗