PaperlayerView exact revision on arXiv

Source-linked AI summary

Reliable and Responsible Foundation Models: A Comprehensive Survey

Xinyu Yang, Junlin Han, Rishi Bommasani, Jinqi Luo, Wenjie Qu, Wangchunshu Zhou, Adel Bibi, Xiyao Wang, Jaehong Yoon, Elias Stengel-Eskin, Shengbang Tong, Lingfeng Shen, Rafael Rafailov, Runjia Li, Zhaoyang Wang, Yiyang Zhou, Chenhang Cui, Yu Wang, Wenhao Zheng, Huichi Zhou, Jindong Gu, Zhaorun Chen, Peng Xia, Tony Lee, Thomas Zollo, Vikash Sehwag, Jixuan Leng, Jiuhai Chen, Yuxin Wen, Huan Zhang, Zhun Deng, Linjun Zhang, Pavel Izmailov, Pang Wei Koh, Yulia Tsvetkov, Andrew Wilson, Jiaheng Zhang, James Zou, Cihang Xie, Hao Wang, Philip Torr, Julian McAuley, David Alvarez-Melis, Florian Tramèr, Kaidi Xu, Suman Jana, Chris Callison-Burch, Rene Vidal, Filippos Kokkinos, Mohit Bansal, Beidi Chen, Huaxiu Yao

arXiv:2602.08145v1cs.LGcs.AIcs.CLcs.CVcs.CY

TL;DR

Foundation models are rapidly deployed across society, making reliable and responsible development urgent. This survey examines technical and societal challenges across foundation models, reviews existing research and future directions, and emphasizes that their interconnected risks require a holistic approach.

  • Problem

    The rapid socioeconomic impact and broad deployment of foundation models create an urgent need to ensure that these systems are reliable and responsible.

  • Method

    The survey comprehensively reviews reliability and responsibility across bias, alignment, security, privacy, hallucinations, and related areas, including evaluation and mitigation methods.

  • Results

    The survey identifies significant research progress but finds that outstanding challenges still limit the reliability and responsible development of foundation models.

  • Takeaways & Limitations

    Reliability and responsible development should be addressed holistically because the covered research areas are interconnected and influence one another.

  • Takeaways & Limitations

    Privacy attacks remain difficult to construct under weaker assumptions and more general scenarios, while personal information can leak from conversational context even after sensitive-data cleaning.

Abstract

from arXiv · show

Foundation models, including Large Language Models (LLMs), Multimodal Large Language Models (MLLMs), Image Generative Models (i.e, Text-to-Image Models and Image-Editing Models), and Video Generative Models, have become essential tools with broad applications across various domains such as law, medicine, education, finance, science, and beyond. As these models see increasing real-world deployment, ensuring their reliability and responsibility has become critical for academia, industry, and government. This survey addresses the reliable and responsible development of foundation models. We explore critical issues, including bias and fairness, security and privacy, uncertainty, explainability, and distribution shift. Our research also covers model limitations, such as hallucinations, as well as methods like alignment and Artificial Intelligence-Generated Content (AIGC) detection. For each area, we review the current state of the field and outline concrete future research directions. Additionally, we discuss the intersections between these areas, highlighting their connections and shared challenges. We hope our survey fosters the development of foundation models that are not only powerful but also ethical, trustworthy, reliable, and socially responsible.

1 Introduction

This survey examines reliable and responsible foundation models across four model categories and nine critical dimensions. It synthesizes current research, interconnections, and future directions while emphasizing intrinsic reliability and responsibility during intended use.

  • Scope and organization: The survey organizes research around bias and fairness, alignment, security, privacy, hallucination, uncertainty, distribution shift, explainability, and AIGC detection.It also connects these dimensions across LLMs, MLLMs, image generative models, and video generative models.
  • Motivation: Foundation models are increasingly integrated into applications ranging from business decision-making to everyday personal assistance.This expanding deployment motivates attention to reliable and responsible development.
  • Contribution: The survey provides a holistic analysis connecting nine dimensions across four major model classes, revealing interconnections and trade-offs overlooked by specialized reviews.The cross-cutting perspective is intended to support researchers, practitioners, and policymakers.
  • Scope and organization: Its scope focuses on intrinsic properties during intended use, complementing work on deliberate external misuse such as disinformation and cyberattacks.The reviewed literature was publicly available up to May 2025.

2 Types of Foundation Models

The survey distinguishes four foundation-model categories by their input and output modalities: LLMs, MLLMs, image generative models, and video generative models. Each category is characterized by different mappings and architectures for downstream generation or understanding.

  • Large Language Models: LLMs understand, generate, and manipulate language for tasks including summarization, translation, sentiment analysis, and dialogue.They include encoder-only, decoder-only, and encoder-decoder architectures, commonly implemented with Transformers.
  • Multimodal Large Language Models: MLLMs process multiple input modalities to generate text, commonly integrating multimodal features in a shared latent space.The survey primarily considers the image-text-to-text mapping g: I × T →T.
  • Image Generative Models: Image generative models generate images from multimodal inputs, most commonly textual instructions, using diffusion models, GANs, or autoregressive Transformers.Their applications include digital art, product design, and education.
  • Video Generative Models: Video generative models map multimodal inputs to video outputs and often extend diffusion models to model temporal relations for coherent generation.Autoregressive approaches are an alternative paradigm for modeling these temporal relations.

3 Bias and Fairness

Bias in foundation models can be inherited from large-scale pretraining data and propagated into downstream applications. The survey reviews bias and fairness concepts, evaluation and mitigation methods, and multimodal manifestations.

  • Bias and fairness: Foundation models can acquire training-data biases that propagate into downstream applications, while their sources and impacts remain poorly understood.The survey identifies further measurement and mitigation as necessary for fairness and equity.
  • Survey structure: The bias and fairness review covers definitions, consequences, measurement, mitigation, and multimodal cases involving MLLMs and image generative models.Figure 3 provides a visualization of the categorization of these concepts and methods.

3.1 Definitions

The survey defines social bias as systematic disparities, stereotypes, or inaccuracies associated with social groups and frames fairness as avoiding group disparities in model behavior and performance. It illustrates these concepts through language-generation examples and a measurement-based formulation.

  • Definitions: Social biases correlate demographic groups with stereotypical, uncharacteristic, or overly generalized traits in language technologies.Their importance follows from language’s connections to identity, social relationships, and power dynamics.
  • Definitions: LLM bias can arise from unbalanced distributions across domains and genres, such as unequal gender representation in source materials.The survey uses gendered leadership language as an illustrative example.
  • Examples: Bias may appear as local or global patterns in text generation and as gender-specific expressions in machine translation.Examples include stereotyped job associations and masculine defaults for job-related translations.
  • Fairness: Fairness requires model behavior, outputs, and decisions to respect social diversity without systematic performance disparities across groups.The survey formalizes fairness through a response-scoring function S applied to groups A and B.
  • Fairness: One bias definition compares expected scores for group-specific prompts, classifying a model as biased when the equality condition is not satisfied.The survey notes that alternative definitions and metrics can also be reasonable.

3.2 Methods for Bias Evaluation

Bias evaluation in LLMs uses generated text, feature embeddings, token-selection probabilities, and distributional comparisons to detect disparities across social groups.

  • Evaluation approaches: LLM bias evaluation commonly analyzes generated text, feature embeddings, or token-selection probabilities.These approaches assess model behavior at output, representation, or token-prediction levels.
  • Distribution metrics: Distribution metrics compare responses across group-conditioned contexts or measure demographic word frequencies against baseline distributions.Demographic Representation normalizes group-count vectors before applying total variation, KL divergence, or Wasserstein distance.
  • Generated-text evaluation: Generated-text methods use prompted outputs and metrics such as toxicity, sentiment, respect, style, and harmful-word frequencies.Classifier-based methods can compare outputs across groups, while word-level methods use harmful-word lists or precomputed bias scores.
  • Feature-embedding evaluation: Embedding methods measure associations between social-group terms and neutral attributes using vector-space distances, WEAT, or contextualized SEAT.WEAT examines word-level associations, whereas SEAT extends this analysis to sentence embeddings and contextualized templates.
  • Token-probability evaluation: Token-probability methods mask words or score pseudo-log likelihood to test preferences for biased or stereotypical alternatives.PLL evaluates the probability of individual words in a sentence and can compare stereotypical with less stereotypical sentence pairs.

3.3 Methods for Bias Mitigation

LLM bias mitigation is organized into training-based interventions and inference-time post-processing, with data quality, diversity, alignment, and external knowledge as central mechanisms.

  • Overall taxonomy: Bias mitigation methods broadly fall into training-process methods and post-processing techniques.Training interventions modify data or instruction tuning, whereas post-processing detects and corrects biased generated text during inference.
  • Training-process methods: Training data augmentation improves quality, diversity, and balance by adding instances for underrepresented or misrepresented social groups.Data filtering instead removes low-quality data or retains diverse and underrepresented examples.
  • Training-process methods: Counterfactual interpolation methods blend augmented examples with original data and labels to balance training distributions.Mixup-based approaches can align pretrained-model logits for opposing gendered words or operate across multiple corpora.
  • Training-process methods: Automated self-instructive frameworks reconstruct diverse instruction-tuning datasets from seed data using metric-based filtering.The generated prompts are produced automatically and filtered to preserve dataset diversity.
  • Post-processing techniques: Post-processing invokes external knowledge bases or word-based detectors to identify and correct biased statements in generated text.These methods can handle certain biased information but may miss biases absent from knowledge bases, introduce erroneous information, and increase latency.

3.4 Bias and Fairness in MLLMs

Bias and fairness research in MLLMs emphasizes faithful alignment with multimodal inputs, but in-depth investigation beyond image hallucination remains limited.

  • Scope: MLLM fairness requires responses to align faithfully and consistently with visual or auditory context.The focus differs from LLM fairness by emphasizing cross-modal faithfulness.
  • Current research focus: Most current MLLM research examines image hallucination, where responses contain entities, quantities, or logical information absent from the image.This phenomenon is treated as a major focus of current bias and fairness work in MLLMs.
  • Open challenges: Bias and fairness in MLLMs remain largely unexplored despite risks from scarce group-specific image-text data, biased datasets, and imbalanced distributions.These conditions may produce stereotypical impressions and biased responses for particular image-text pairs.
  • Mitigation: BEND-VLM adapts debiasing operations to each test input to avoid catastrophic forgetting during fine-tuning.The approach targets MLLM embeddings at inference time.

3.5 Bias and Fairness in Image Generative Models

Image generative models exhibit systematic social biases in generated content, while mitigation research uses prompting and fairness-aware sampling to improve representation balance.

  • Bias patterns: Text-guided diffusion models show biases involving professions, ethnicities, and social classes, with outputs diverging from real-world distributions.Studies report that generated content can amplify biases present in society and underrepresent marginalized groups.
  • Prompt-based mitigation: Prompting methods mitigate bias by injecting pronouns, optimizing soft prompt tokens, or modifying text-embedding representations.Fair Diffusion randomly adds subject pronouns to promote more balanced gender distributions.
  • Bias patterns: DALL·E exhibits a spurious correlation between gender and profession in generated images.The example illustrates how image-generation outputs can associate occupations with gender inappropriately.
  • Sampling-based mitigation: Sampling methods address bias by generating balanced classes or applying fairness-aware sampling to reduce amplification from training data.D2C uses few-shot conditional diffusion, while Fair Sampling targets amplified training-data biases.

3.6 Current Limitations and Future Directions

Bias and fairness evaluation in foundation models remains limited, especially for autoregressive bias propagation and context-dependent judgments. The survey identifies future directions spanning training, generation, evaluation, interpretability, multimodality, and fairness–utility trade-offs.

  • Current Limitations: Current bias evaluations often miss gradual bias propagation across autoregressive generation because they focus on tokens or paragraphs.Each prediction depends on previously generated tokens, allowing biases to accumulate and spread over time.
  • Current Limitations: Bias in social media and current-events content can emerge through subtle contextual, historical, social, and cultural influences rather than isolated tokens.These influences can intertwine with factual reporting and socially accepted norms, complicating mitigation.
  • Current Limitations: Unclear and culturally variable definitions of fairness complicate both bias assessment and improvement.News applications may require balancing fairness principles with historical accuracy and accurate presentation of current facts.
  • Current Limitations: Absolute fairness can conflict with factual authenticity when accurately reporting inequalities or biased social structures.The survey frames preserving information integrity while addressing ethical concerns as a significant challenge for developers.
  • Future Directions: Future research should investigate unbiased representations, fairness-aware fine-tuning and security, post-processing, auditing, mechanistic interpretability, modality-specific interventions, and fairness–utility balancing.Suggested methods include constrained RLHF, DPO, revised LoRA, post-processing of autoregressive outputs, and editing circuits associated with stereotypes.

4 Alignment

The survey presents alignment as a multi-stage process spanning SFT, preference optimization, prompt engineering, and multimodal alignment, while emphasizing vulnerabilities, data dependence, and unresolved challenges in achieving robust human-value alignment.

  • Alignment methods: Foundation models are aligned through methods including SFT, RLHF, prompt engineering, and multimodal alignment for MLLMs.SFT and RLHF are typically used during post-training for LLMs; MLLMs additionally require multimodal alignment such as Visual Instruction Tuning.
  • Supervised Fine-Tuning: SFT directly tunes pretrained LLMs to imitate desired ground-truth responses and often serves as the first alignment stage.Its objective maximizes the likelihood of a selected target response, and combining it with other methods can improve alignment stability.
  • Supervised Fine-Tuning: SFT alignment depends heavily on training-data quality and diversity, motivating curated datasets, deduplication, cleaning, and difficulty- or importance-based selection.LIMA curated 1,000 high-quality prompt-response pairs, while other methods use Minihash, LSH, filtering, instruction-following difficulty, and importance weights.
  • Current limitations and future directions: Alignment methods remain vulnerable: adversarial fine-tuning can undermine safety, RLHF faces tractability and generality challenges, and direct methods show out-of-distribution robustness issues.Reported concerns include feedback quality and bias, reward hacking, power-seeking, offline-data limitations, and insufficient exploration relative to online training.
  • Prompt Engineering: Prompt engineering mainly improves task-specific performance and contributes more indirectly to value alignment through improved output interpretability.The survey frames this performance–alignment relationship as an important consideration in developing LLMs.
  • Current limitations and future directions: Superalignment seeks to keep future superintelligent systems aligned through scalable oversight and weak-to-strong generalization, but superficial alignment may leave exploitable backdoors.The survey also identifies the difficulty of reliably evaluating whether models follow underlying human values rather than merely producing preferred responses.

5 Security

The survey examines security threats and defenses across foundation models, emphasizing attacks during training and inference and their interactions with broader reliability concerns. It also identifies formal guarantees, multimodal robustness, over-safety, and practical threat validity as unresolved challenges.

  • Threats: Security vulnerabilities span training and inference, including jailbreaks, prompt injection, poisoning, backdoors, and adversarial attacks across foundation-model classes.The survey places these threats within a unified cross-task reliability and responsibility perspective.
  • Attacks in LLMs: Jailbreaks bypass safety mechanisms, while prompt injections override original instructions through malicious prompts, including hidden instructions in external data.Jailbreaks can unlock restricted capabilities, whereas indirect prompt injection exploits third-party sources such as emails, PDFs, and web pages.
  • Attacks in LLMs: Poisoning and backdoor attacks manipulate training data so models produce predefined behavior when a specific trigger appears during inference.Instruction tuning remains vulnerable to crafted examples inserted into training datasets.
  • Defenses: Defenses include output ensembling, system-prompt wrapping, activation editing, input preprocessing, perplexity-based detection, and harmful-content removal.Image-generation defenses may be incomplete, poorly scalable, and degrade benign image quality, limiting adoption by online services.
  • Limitations and Open Challenges: Most attack methods are costly to optimize, and jailbreaks often generate only simple harmful text, leaving their real-world threat uncertain.White-box methods require substantial computation, while black-box methods incur economic costs from massive token consumption.
  • Limitations and Open Challenges: Open challenges include securing multi-agent, tool-using, and retrieval-augmented systems, while defenses still lack formal guarantees and remain vulnerable across modalities.Over-safety can also cause benign inputs to be rejected as malicious, harming user experience.

6 Privacy

The survey reviews privacy threats arising from foundation-model memorization and the techniques developed to mitigate them. It emphasizes that stronger privacy protection must be balanced against practical attack assumptions and model utility.

  • Privacy Risks: Foundation models can memorize sensitive training data, creating privacy risks that grow with model scale and motivating privacy-preserving machine learning.The survey also considers privacy constraints imposed by regulations such as GDPR and CCPA.
  • Privacy Threats: Privacy attacks include membership inference, training-data extraction, prompt disclosure, and attacks targeting decoding hyperparameters or multimodal models.Extraction attacks have revealed sensitive information such as phone numbers and email addresses from language models.
  • Privacy Protection: Differential Privacy provides quantifiable protection by limiting how much an algorithm’s output changes when one individual’s data is added or removed.DP-SGD injects sample-wise Gaussian noise into gradients during optimization.
  • Privacy Protection: Other defenses remove sensitive information from weights, filter toxic outputs, scrub personally identifiable information, exclude sampled tokens from loss computation, or perturb embeddings.These methods target specific leakage or memorization pathways rather than providing a single universal defense.
  • Limitations and Open Challenges: Membership inference is difficult because non-member data can closely resemble member data, blurring the distinction required by the attack.Training-data extraction avoids this ambiguity but currently recovers only a small fraction of training data.
  • Limitations and Open Challenges: Privacy attacks often rely on strong assumptions, while future work seeks effective attacks under weaker assumptions and contextualized privacy protections.Contextual leakage may arise from personal attributes inferred across an entire multi-turn conversation.
  • Limitations and Open Challenges: Differential Privacy offers worst-case leakage bounds but can create a gap with practical attacks and significantly degrade fine-tuning performance.Data-sanitization pipelines likewise degrade performance on main tasks, leaving privacy–utility balance unresolved.

7 Hallucination

The survey defines hallucination as ungrounded or unverifiable model output across text, images, video, and audio, and organizes its forms and mitigation strategies. It highlights unresolved validation weaknesses and calls for more reliable knowledge bases, validators, and multimodal research.

  • Overview: Hallucinations are unfaithful, insufficiently grounded outputs that can appear across text, image, video, and audio generation and impair real-world applications.They may present incorrect statements with a highly confident tone.
  • Hallucinations in LLMs: LLM hallucinations divide into factuality errors against world knowledge and faithfulness errors involving instructions, demonstrations, or preceding generations.The survey illustrates both categories with examples in Figure 17.
  • Hallucinations in LLMs: Factuality hallucination includes factual inconsistency with established knowledge and factual fabrication that appears factual but lacks verifiable evidence.These forms reflect different relationships between generated content and credible real-world sources.
  • Hallucinations in LLMs: Faithfulness hallucination includes conflicts with input instructions, in-context demonstrations, or the model’s preceding text.One example is summarizing a different text chunk from the one specified by the prompt.
  • Hallucinations in MLLMs: MLLM hallucinations also distinguish factuality and faithfulness, with factuality further separated into intrinsic and extrinsic forms.Intrinsic errors concern inaccurate image descriptions, while extrinsic errors add unverifiable or contradictory information despite correctly describing the image.
  • Reducing Hallucinations: Mitigation research for LLM hallucinations uses data-centric and model-based approaches, with training data memorization identified as a major determinant of hallucination frequency.The survey links deficient encoder comprehension with hallucination frequency and categorizes mitigation methods accordingly.
  • Current Limitations and Future Directions: Hallucination research remains incomplete because validation depends on potentially untrustworthy knowledge bases and hallucination-prone verifiers, while non-language modalities receive insufficient study.The survey notes that the statistical relationship between autoregressive hidden states and hallucination remains unclear.
  • Current Limitations and Future Directions: Future directions include verifying external knowledge, using multi-agent cross-validation, studying autoregressive hidden states, editing representations, and calibrating refusal mechanisms.These directions target both validation reliability and more faithful generation.

8 Uncertainty

Uncertainty in foundation models arises from data variability, limited model knowledge, language ambiguity, and deployment conditions such as distribution shifts. The survey reviews ways to quantify and mitigate it while identifying important limits in current evidence, cost, and applicability.

  • 8.1 Sources and Types of Uncertainty: Foundation-model uncertainty includes aleatoric uncertainty from inherent data randomness and epistemic uncertainty from limited or imperfect model knowledge.Deployment can add uncertainty through distribution shifts, novel scenarios, and unforeseen user inputs.
  • 8.1.1 Data: Natural language creates multiple plausible outputs through ambiguous inputs, open-ended requests, and semantically equivalent surface forms.For example, differently worded answers can express the same meaning even when the input is fixed.
  • 8.2 Quantifying and Addressing Uncertainty: Reliable black-box deployment requires measuring response uncertainty, commonly using predictive entropy and related estimates of confidence.The survey also discusses recalibration and using uncertainty estimates to identify cases requiring additional handling.
  • 8.2.2 Semantic Uncertainty: Semantic uncertainty methods sample multiple sequences and group them by meaning, addressing the mismatch between sequence variation and equivalent answers.This extends uncertainty measurement beyond surface-form differences to semantic equivalence.
  • 8.2.5 Distribution-free Uncertainty Quantification: Prompt Risk Control selects prompts using rigorous upper bounds on informative risk measures, unifying techniques from distribution-free uncertainty quantification.The framework considers system prompts or few-shot examples and validates candidate prompts before selection.
  • 8.3 Current Limitations and Future Directions: Current uncertainty-quantification evidence is limited by simple evaluation settings, inference costs of 2 to 20 times or more, and unclear zero-shot gains over sequence entropy.The survey calls for broader pipeline-wide evaluation, calibration benchmarking, and attention to fine-tuning effects on uncertainty estimates.

9 Distribution Shift

Foundation models often lose performance under distribution shifts, creating reliability challenges in real-world settings. The survey reviews detection, generalization, adaptation, and editing strategies, while identifying efficiency, continual learning, and multimodal integration as future directions.

  • OOD Detection: LLMs can detect OOD data without fine-tuning, while in-distribution fine-tuning can further improve OOD detection.The reported investigation evaluates LLaMA families across different model sizes and considers zero-shot and fine-tuning scenarios.
  • OOD Generalization: Adversarial training improves LLM generalization in most scenarios, but Gaussian noise and complex connective-based questions substantially challenge MLLM performance.The MLLM study tests noise addition, blurring, median filtering, and adversarial questions involving conjunctions, disjunctions, and negations.
  • Adaptation Strategies: Fine-tuning and in-context learning show different strengths: full-data fine-tuning favors ID performance, whereas ICL performs better on OOD instances and low-resource settings.The BOSS benchmark also finds that ICL effectiveness varies across models and tasks, while fine-tuning remains non-identifiable in some real-world behaviors.
  • Adaptation Strategies: Fine-tuning methods can target distribution mismatch more efficiently: TAIA updates all parameters but uses only fine-tuned attention parameters during inference.The underlying analysis reports that attention parameters significantly benefit downstream performance when training and test distributions mismatch; PEFT methods can preserve more pretrained capabilities.
  • Model Editing: Model-editing paradigms trade off updateability, locality, scalability, precision, and cross-modal generalization, with multimodal settings requiring modality-aware retrieval or cross-modal localization.The survey highlights unintended side effects for knowledge-based edits and suggests hybrid approaches combining parameter precision with external-memory or classifier mechanisms.
  • Challenges Under Distribution Shift: Foundation models often exhibit decreased performance when exposed to data that deviates from their training distribution, undermining reliability and robustness.The survey attributes these challenges to difficulties in OOD detection, generalization, and adaptation.
  • Current Limitations and Future Directions: Future work should develop lightweight OOD methods, continual or lifelong learning, and unified multimodal frameworks that adapt efficiently while preserving previously acquired knowledge.These directions aim to reduce retraining requirements and improve robustness across diverse real-world applications.

10 Explainability

Explainability research studies how foundation-model inputs, internal components, knowledge, and training samples influence outputs, while emphasizing that plausible explanations may not be faithful. The field increasingly seeks robust, faithful, and transferable explanations across modalities, supported by standardized evaluation and benchmarks.

  • Methods: Explainability methods examine raw input features, internal knowledge, and the roles of samples across pre-training, fine-tuning, and few-shot learning.The literature includes perturbation-, gradient-, attention-, probing-, and concept-based approaches.
  • Feature Attribution Methods: Perturbation-based explanations modify tokens or spans and monitor model representations or outputs to estimate their influence on predictions.Polyjuice extends this direction by generating diverse realistic counterfactuals.
  • Attention-based explanations: Attention visualizations expose token relationships and cross-layer information flow, but attention scores do not necessarily identify faithful explanations.Attention-flow methods can correlate more strongly with gradient- and ablation-based importance than raw attention.
  • Evaluation: Explainability evaluation distinguishes plausibility, which reflects human persuasiveness, from faithfulness, which reflects whether an explanation represents the model’s reasoning.ERASER operationalizes faithfulness through comprehensiveness and sufficiency of extracted rationales.
  • Improving Model Performances: Explanations can improve prompting and reasoning performance, but chain-of-thought rationales may be systematically unfaithful and misleading.Scratchpads can support multi-step reasoning, whereas plausible rationales should not be treated as guaranteed accounts of model reasoning.
  • Transferability Across Modalities: Future work should develop robust, faithful explanation frameworks that transfer across modalities and aggregate multiple explanation levels.The survey also calls for standardized cross-modal explanation benchmarks and critical assessment of interpretability methods.

11 AIGC Detection

AIGC detection classifies content as AI- or human-generated using zero-shot, watermark-based, or learnable detectors across textual and visual inputs. The survey highlights deployment and robustness challenges, including limited model access, increasingly capable generators, data scarcity, cross-domain overfitting, and paraphrasing attacks.

  • Overview: AIGC detection treats images, text, or audio as inputs to a binary classifier that labels content as AI-generated or human-created.The survey categorizes detectors as zero-shot, watermark, and learnable methods.
  • Zero-Shot Detectors: Zero-shot detectors use intrinsic content distinctions and require no additional training of the detector or generating foundation model.Traditional textual approaches use statistics such as entropy, perplexity, n-gram frequencies, and token log probabilities.
  • Zero-Shot Detectors: Statistical detection approaches are inadequate for diverse, high-quality foundation-model content, while API-level methods reduce dependence on source-model access.Surrogate models and re-prompting are among the proposed API-level strategies.
  • Future Directions: Improved generative-model quality makes intuitive AIGC indicators less reliable and increases the need to identify image origins when source-model access is restricted.The image-origin problem is especially practical when only limited source-model images are available.
  • Pre-trained LLMs: Pre-trained LLMs perform poorly as training-free detectors: ChatGPT and GPT-4 were unreliable, while GPT-3.5-turbo approached random guessing.These findings concern direct question-and-answer prompting for generated-text identification.
  • Watermark Detectors: Learnable watermarking encodes keyed payloads during generation and verifies them with a corresponding key or public verifier.Training can use adversarial discrimination or semantic-difference penalties to preserve coherence and consistency.
  • Neural Network Detectors: Reliable neural-network detectors face data scarcity, paraphrasing attacks, cross-domain overfitting, and reduced performance on content from different models.These constraints are especially difficult in diverse domains with limited training samples.

12 Intersection and Conclusion

The survey presents a comprehensive view of reliable and responsible foundation models while emphasizing that challenges across technical and societal dimensions are interconnected. It identifies persistent limitations and argues for holistic research and development across model classes and trustworthy-AI topics.

  • 12 Intersection and Conclusion: Current research has made progress, but outstanding challenges still limit how reliably and responsibly foundation models can be developed.The survey links this need for further research to the technology’s broader societal impact.
  • 12 Intersection and Conclusion: The survey unifies reliability and responsibility research across four foundation-model classes and nine trustworthy-AI dimensions.Its cross-cutting perspective connects topics that prior surveys often examine separately.
  • 12.2 Bias, Fairness, and Security: Security interventions can create fairness trade-offs, because adversarial training may worsen disparities and data poisoning can introduce or amplify bias.The survey notes disproportionate degradation for underrepresented groups and biased hate-speech detection of non-offensive African American English.
  • 12 Intersection and Conclusion: Privacy protections remain vulnerable: jailbreaking can circumvent prompts and RLHF, while sensitive information may be recoverable despite model editing.The survey also frames privacy as a legal concern, including GDPR’s Right to be Forgotten.
  • 12 Intersection and Conclusion: AIGC detection remains imperfect under adversarial attacks, while generated content can pose societal-security risks by spreading misinformation and eroding trust.The survey also identifies fairness concerns when detection or watermarking affects content associated with particular groups.
  • 12 Intersection and Conclusion: Uncertainty and distribution shift are linked to hallucinations because unfamiliar data can produce inaccurate OOD detection and overconfident predictions.This unreliability is especially concerning in safety-critical applications.
Loading 2602.08145v1…