Model Safety, Alignment & Jailbreaks
A Framework for Deploying Activation- and Logit-Based Methods Against Closed-Source LLM APIs via Instrumented Proxy Models
Generated automatically from the limitations stated in 177 papers (ACL, EMNLP, ICLR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
A very large body of techniques — safety defenses, interpretability probes, evaluation metrics, decoding controls — requires hidden states, attention maps, gradients, or token-level probabilities, and is therefore silently restricted to open-weight models. As a result, published methods are never validated on, and cannot protect or audit, the models most people actually use (GPT-4, Claude, Gemini). This is the single most common self-reported scope limitation in the literature, yet each paper treats it as an isolated footnote rather than a solvable engineering and inference problem.
Why it matters
It converts hundreds of open-weights-only methods into candidate deployable tools for closed APIs, and gives providers an evidence-based menu of which minimal exposures would enable which safety/evaluation capabilities.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Systematic transfer study: take 10–20 representative white-box methods (refusal probing, hallucination detection, jailbreak scoring, evaluation metrics), fit them on open proxies (Llama, Qwen, Mistral families), and deploy them against closed APIs using only black-box signals (text, top-k logprobs where offered, embeddings endpoints). Measure transfer of detection/calibration quality (AUROC, F1) relative to native white-box performance, and identify which internal quantities have cheap black-box surrogates.
- 2
Black-box reconstruction: train lightweight surrogate probes on API outputs (logprobs, responses to contrastive prompt sets) that predict the latent quantities the method needs (refusal probability, uncertainty, hidden-state-derived scores), and measure how much of the white-box method's discriminative power is recovered per API budget.
- 3
Provider-side instrumentation protocol: specify the minimal state exposure (e.g., top-k logprobs, one chosen hidden layer) a provider would need to release, and empirically measure, method by method, how performance degrades across exposure tiers — turning "white-box required" into a quantified requirement rather than a binary.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
Internal states may simply not transfer across model families — refusal/uncertainty geometry differs enough between the open proxy and the closed target that surrogate probes recover little signal, leaving only the (unactionable) recommendation that providers expose more.
Sub-problems
Black-Box Safety Auditing and Alignment Verification for API-Only Foundation Models
Current safety auditing methods, alignment interventions, and vulnerability evaluations rely directly on internal model representations, attention maps, or backward gradients. Because commercial frontier models are accessible only via black-box inference APIs, third-party auditors and downstream developers cannot inspect internal activations, calculate attributions, or fine-tune weights to enforce safety constraints. This leaves the most widely deployed foundation models structurally unverifiable by standard white-box safety pipelines.
Black-Box Behavioral Attribution and Interpretability for Closed-Source Language Models
Current interpretability and in-context learning diagnostic tools overwhelmingly require white-box access to attention matrices, intermediate activations, or raw token logits. Because commercial frontier models are accessible solely through text-generation APIs, these diagnostic and auditing methods cannot be deployed on the systems that handle most real-world user traffic. This leaves safety auditors and practitioners unable to verify reasoning mechanisms, explain unexpected failures, or trace factual hallucinations in closed-source models.
Behavioral and Query-Based Surrogates for White-Box LLM Activation Probing
State-of-the-art evaluation techniques—including truthfulness probes, sparse autoencoder feature analysis, and internal anomaly detectors—depend on access to internal hidden states, layer activations, or full-vocabulary log probabilities. Because leading commercial models are accessible exclusively through restricted text-generation APIs, independent researchers and regulators cannot run these diagnostic evaluations on the systems deployed at scale. Consequently, external auditing of proprietary LLMs is limited to surface-level output metrics, unable to leverage the finer-grained diagnostic power of representation-level evaluations.
Black-Box Vision-Language Model Attribution and Verification Without Internal State Access
Leading multimodal foundation models are predominantly deployed behind commercial APIs that expose only prompt inputs and output text or token probabilities. Existing techniques for vision-language tasks—such as visual grounding, hallucination detection, feature attribution, and model steering—overwhelmingly require white-box access to internal attention maps, layer activations, or gradients. Consequently, practitioners cannot apply standard interpretability, auditing, or safety-verification tools to the most capable proprietary vision-language models. This creates an operational barrier between open-weight academic methods and closed-source production deployments.
Black-Box Reasoning Steering and Diagnosis Without Hidden-State Access
Current interpretability, steering, and verification techniques for LLM reasoning rely strictly on white-box access to intermediate activations, residual streams, or unembedding matrices. Because leading commercial models (such as GPT-4 and Claude) are served exclusively via text-only or restricted-logit APIs, these internal-state methods cannot be deployed or audited on frontier reasoning engines. Structurally, developers and safety evaluators cannot monitor, steer, or verify latent reasoning trajectories without internal tensor access.
Black-Box Compression and Steering of Chain-of-Thought Reasoning
Current chain-of-thought optimization, compression, and steering techniques require internal access to hidden activations, attention weights, or gradient updates. Consequently, these methods cannot be deployed on proprietary, API-hosted frontier models where a substantial fraction of production reasoning workloads take place. Users of closed-source models are forced to pay full inference latency and token costs without access to the efficiency or calibration mechanisms developed for open-weight models. This creates a hard practical divide between white-box reasoning research and black-box deployment.
Black-Box Behavioral Steering and State Tracking for API-Based Language Agents
Current methods for auditing, steering, and editing language agents depend on white-box access to internal states, including activation vectors, attention maps, gradients, and model weights. Because the most capable frontier models are accessible only via closed-source inference APIs, these white-box techniques cannot be applied to deployed commercial agents. Practitioners are consequently unable to monitor internal reasoning failures, edit agent capabilities, or apply activation-level safety steering to proprietary models. Developing black-box equivalents is necessary to inspect and control agents operating behind API boundaries.
Black-Box Interpretability and Mechanism Auditing for API-Restricted Large Language Models
The vast majority of mechanistic interpretability and in-context learning diagnostics rely on inspecting internal activations, attention maps, and layer-wise representations. As a result, these diagnostic and hallucination-detection tools are entirely unusable on frontier commercial models accessible only via query APIs (e.g., GPT-4o, Gemini). Practitioners deploying or evaluating closed-source systems are blocked from verifying whether a model's in-context reasoning relies on genuine task induction or superficial heuristic retrieval. Developing black-box proxies that infer internal mechanism dynamics from behavioral traces and output distribution perturbations bridges this critical evaluation divide.
Black-Box Interpretability and Diagnostic Auditing for Closed-API Language Models
Current interpretability and mechanistic diagnostic methods strictly depend on internal representations—gradients, hidden states, attention maps, and predictive logits—that are unavailable for closed-source, API-served models. Consequently, third-party researchers and auditors cannot inspect, explain, or verify the reasoning and factual reliability of the most capable deployed frontier models. This leaves external safety and compliance evaluations structurally blocked unless model providers voluntarily expose internal access.
Cross-Model Generalization and Failure Boundaries of Chain-of-Thought Reasoning Frameworks
Current prompt-based and search-augmented reasoning methods are overwhelmingly validated on a single proprietary model (often GPT-3.5 or GPT-4), yet independent reports show their benefits frequently vanish or reverse when applied to different model families or capability scales. Without systematic cross-model evaluation, researchers cannot distinguish whether a proposed reasoning technique provides general algorithmic utility or merely exploits idiosyncratic behaviors of specific backbones. This blocks the reliable deployment of reasoning frameworks across open-source and alternative foundation models.
Black-Box Hallucination Detection and Auditing for Vision-Language Models Without Proprietary Judges
Current techniques for assessing and mitigating hallucinations in large vision-language models (LVLMs) depend heavily on either white-box access to internal activations or validation by proprietary API judges such as GPT-4V. This creates a structural barrier: white-box methods cannot audit commercial, closed-source models where weights and activations are inaccessible. Simultaneously, relying on proprietary APIs for evaluation introduces severe reproducibility issues caused by model deprecation, opaque backend updates, and high query costs. As a result, there is currently no standardized, reproducible way to verify multimodal generation quality on arbitrary black-box endpoints.
Black-Box Model Safety Auditing and Defense Without Internal Hidden State Access
Current activation-based safety mechanisms—including representation engineering, hidden-state linear probes, and internal steering vectors—depend strictly on white-box access to intermediate activations. Because major commercial models operate strictly behind text- or top-logprob APIs, these safety defenses and latent auditing tools cannot be deployed by downstream developers or third-party auditors. Consequently, closed-weight models remain vulnerable to safety blind spots that internal representation monitors are designed to detect, leaving a major discrepancy between academic safety tooling and commercial deployment.
Black-Box Reasoning Evaluation and Confidence Estimation Without Token Log-Probabilities
Current evaluation, verification, and uncertainty estimation frameworks for chain-of-thought reasoning depend on access to token-level log-probabilities or output distributions. Because major proprietary foundation models (e.g., GPT-4, Claude, Gemini) only expose generated text via black-box APIs, these evaluation and selection methods cannot be applied to the most capable deployed systems. This creates an empirical blind spot where open-source and proprietary models cannot be benchmarked or guided using the same evaluation machinery.
Black-Box Context Pruning and Efficiency Without Model Internals
Modern inference efficiency and long-context techniques rely heavily on white-box signals such as internal hidden states, embedding modifications, and full-vocabulary entropy or perplexity scores. Consequently, these techniques cannot be deployed on commercial proprietary LLMs (e.g., GPT-4, Claude) that only expose string-in, string-out interfaces. Developers building on commercial APIs are forced to transmit full, uncompressed context windows at maximum token cost and latency.
Strict Black-Box Safety and Alignment Auditing for Text-Only APIs
The majority of formal alignment diagnostics, safety verification methods, and red-teaming frameworks rely on access to internal activations, full-vocabulary logits, or top-k log probabilities. As commercial frontier LLM providers increasingly restrict API outputs to raw text to defend against model extraction and distillation, these safety assessment methods become unusable on deployed systems. Independent auditors, researchers, and regulators are structurally blocked from validating the safety boundaries and alignment guarantees of closed-source models using internal-state diagnostics.
Black-Box Cross-Lingual Transfer and Steering for API-Only Large Language Models
Current multilingual transfer and cross-lingual steering techniques rely on inspecting or modifying internal hidden states, forward passes, and FFN activations. Because frontier models (such as GPT-4o and Gemini) are accessible strictly via text-in/text-out APIs, these internal intervention methods cannot be deployed on top-tier commercial LLMs. Practitioners working with low-resource languages in production environments are consequently blocked from using representation-level alignment techniques without hosting their own white-box weights.
Black-Box Safety and Alignment Auditing Without Internal Representation Access
Current representation-based safety audits, latent guardrails, and jailbreak vulnerability assessments strictly require white-box access to internal hidden states, embedding lookup tables, or intermediate activations. As a direct consequence, safety evaluators cannot apply these internal auditing methods to commercial, closed-source API-only models where only text input-output interfaces are exposed. This creates a critical blind spot where safety guarantees evaluated on open-weight models cannot be verified on the frontier models actually deployed to end users.
Black-Box Jailbreak Detection and Refusal Control Without Logit or Representation Access
Existing state-of-the-art defenses against jailbreaks, over-refusal mitigations, and refusal steering methods depend fundamentally on white-box signals such as internal layer activations, latent refusal directions, or next-token log-probabilities. Commercial and proprietary large language models are increasingly served via strictly black-box text-in/text-out APIs where token probabilities and internal embeddings are completely hidden. As a result, third-party auditors, downstream application developers, and API consumers cannot deploy modern representation- or logit-guided safety interventions on closed-source systems.
Black-Box Surrogates and Behavioral Probing for Internal State LM Evaluation
A prominent class of language model evaluation and diagnostic benchmarks relies on inspecting intermediate-layer activations, residual stream dynamics, and unembedding projections. Because leading frontier models (e.g., GPT-4, Claude) are accessible exclusively via black-box text or logprob APIs, these diagnostic evaluation techniques cannot be executed on the very models most widely deployed in practice. As a result, comparative evaluations between open-weight and proprietary models remain bifurcated, unable to audit internal failure modes or calibration on closed systems.
Black-Box Retrieval-Augmented Generation Without Internal Model States
Existing RAG methods for adaptive retrieval, attribution tracking, and generation steering rely on internal model representations such as attention weights, input embedding layers, intermediate gradients, or full vocabulary logit distributions. Because commercial LLM providers expose only text outputs (or at most top-$k$ output token probabilities), these advanced retrieval-augmented techniques cannot be deployed on state-of-the-art closed-source models. Practitioners using proprietary APIs are consequently restricted to naive text-in text-out prompting without confidence-guided retrieval or internal attribution. Developing black-box proxies for these mechanisms is necessary to extend advanced RAG capabilities to closed APIs.
Systematic Evaluation and Scaffolding of Multi-Step Tool-Use Agents Across Open-Source Model Scales
Current language agent and tool-use frameworks are predominantly evaluated on proprietary, closed-source LLM APIs, leaving their generalizability to open-source models across parameter scales systematically untested. When open-source models are used, multi-step execution frequently destabilizes due to brittle formatted outputs, API syntax errors, and compounding trajectory drift. Without empirical characterization across model sizes and open architectures, research in language agents remains economically gated by commercial APIs and structurally non-reproducible.
Black-Box Watermarking and Controlled Decoding for Closed-Source Language Model APIs
The majority of production large language models are hosted behind closed-source APIs that return only raw text completions, offering no access to hidden states, logits, or decoding configurations. Existing watermarking, safety steering, and provenance-tracking techniques rely on modifying the token sampling distribution or monitoring internal representations during inference. Because of this architectural precondition, third-party developers and downstream auditors cannot apply current provenance or control mechanisms to proprietary model endpoints.
Preference Optimization and Alignment for Strictly Black-Box Language Model APIs
Current alignment and preference optimization methods fundamentally depend on white-box or gray-box access, such as updating model weights, reading internal activations, modifying vocabulary tokens, or querying full output logit distributions. Consequently, practitioners and downstream developers interacting with closed-source, API-only models cannot apply standard preference optimization techniques to customize or align model behavior to domain-specific human preferences. This locks proprietary systems into fixed alignment profiles unless providers explicitly retrain them, preventing users from executing post-hoc preference steering beyond basic context prompting.
Black-Box Adaptation and Calibration for API-Constrained Question Answering
High-performance question answering and reading comprehension techniques increasingly rely on parameter-level fine-tuning, internal activation steering, or hidden-state uncertainty estimation. These methods cannot be deployed on state-of-the-art commercial LLMs that are only accessible via text-in/text-out APIs. Consequently, practitioners using closed-source models cannot apply established internal-state interventions to improve faithfulness, calibrate answer confidence, or correct reasoning errors. Bridging this interface barrier is necessary to transfer white-box reliability enhancements to proprietary frontier models.
Evidence
Each paper's own statement of the limitation, verbatim.
- What Do Large Language Models Know About Opinions?ICLR 2026
Requires white-box access to model activations, making it inapplicable to closed-API models.
- Unmasking Backdoors: An Explainable Defense via Gradient-Attention Anomaly Scoring for Pre-trained Language ModelsICLR 2026
Requires white-box access to model gradients and attention weights, preventing deployment in black-box API settings.
- Unlocking Multilingual Reasoning Capability of LLMs and LVLMs through Representation EngineeringACL 2026
Requires white-box access to internal hidden states and the ability to modify forward passes, making it infeasible for closed-source models
- Two Pathways to Truthfulness: On the Intrinsic Encoding of LLM HallucinationsACL 2026
Requires white-box access to internal model representations (hidden states and attention), making the approach inapplicable to black-box API-only models
- Tracing the Traces: Latent Temporal Signals for Efficient and Accurate ReasoningICLR 2026
Requires full access to internal model activations, making it inapplicable to closed-box API models.
- Towards Privacy-Preserving Large Language Model: Text-free Inference Through Alignment and AdaptationACL 2026
Requires open-weight access to the server model for continuous prefix injection and LoRA tuning, making it incompatible with proprietary, black-box API-only LLMs.
- Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking TokensICML 2026
Requires access to intermediate-layer hidden states and the unembedding matrix during decoding, so it cannot be used with API-only models or without per-layer logit computation overhead.
- Thermometer of Thoughts: Enhancing LLM’s Exploration via Attention Temperature ModulationACL 2026
Requires modifying model source code (attention scaling), so it cannot be applied to API-based models like GPT-5 or any closed-weight LLM
Show all 177 papers
- The Geometric Reasoner: Manifold-Informed Latent Foresight Search for Long-Context ReasoningICML 2026
Requires access to hidden states for anchor extraction and residual injection, ruling out closed-box API deployment
- The Digital Dunning-Kruger Effect: Decoupling Hallucinations via Geometric Hidden-state Observation for Semantic TruthfulnessACL 2026
Requires white-box access to hidden states, so it cannot be applied to closed-source API-only models
- The ACUTE Protocol: Operationalizing Language Model Activations for Better Calibration, Utility, and TrustICML 2026
ACUTE requires white-box access to model internals (activations at every layer), so it cannot be applied to API-only models; the authors note only developers have this access
- TRACEDET: HALLUCINATION DETECTION FROM THE DECODING TRACE OF DIFFUSION LARGE LANGUAGE MODELSICLR 2026
Requires access to internal model signals like token-level logits and entropy, making it inapplicable to black-box D-LLM APIs.
- TCAP: Tri-Component Attention Profiling for Unsupervised Backdoor Detection in MLLM Fine-TuningICML 2026
Requires access to the model's internal cross-modal attention maps from the target MLLM, so it cannot be applied to closed-source models or models without attention outputs
- Student Guides Teacher: Weak-to-Strong Inference via Spectral Orthogonal ExplorationACL 2026
Requires white-box access to hidden states for SVD, covariance estimation, and residual scoring, so it cannot be applied to closed-source API-only models
- Stop Tracking Me! Proactive Defense Against Attribute Inference Attack in LLMsICLR 2026
RPS and MPS components require access to model log-probabilities, making them inapplicable to closed-source API-only models without proxy model transfer.
- ShinkaEvolve: Towards Open-Ended and Sample-Efficient Program EvolutionICLR 2026
Relies on closed-source LLM APIs, so large-scale API costs create economic barriers and limit reproducibility of the cost profile
- Self-Reflective Generation at Test TimeACL 2026
Requires white-box access to hidden states and gradients, so it cannot be deployed behind text-only or restricted inference APIs.
- Seeing No Evil: Blinding Large Vision-Language Models to Safety Instructions via Adversarial Attention HijackingACL 2026
Requires white-box gradient access to the full model, so it does not directly apply to closed-source models (transfer results rely on images optimized on Qwen-VL)
- SafeConstellations: Mitigating Over-Refusals in LLMs Through Task-Aware Representation SteeringACL 2026
Requires white-box access to model internals (hidden states), making it inapplicable to closed-weight APIs.
- SafeCompass: Dynamic Chain-of-Thought Steering via Inference-Time Safety SignalsICML 2026
Requires white-box access to internal hidden states, precluding direct use on closed-source models without exposed states
- STARE: Step-wise Temporal Alignment and Red-teaming Engine for Multi-modal Toxicity AttackICML 2026
Requires direct white-box access to the T2I model for RL fine-tuning (288 A100 GPU-hours vs 72 for ART), so it cannot be applied when only a black-box image generator is exposed.
- SARA: Selective and Adaptive Retrieval-augmented Generation with Context CompressionACL 2026
Requires white-box access to the LLM's input embedding layer, making it incompatible with proprietary API-only models.
- Rethinking LLM-as-a-Judge: Representation-as-a-Judge with Small Language Models via Semantic Capacity AsymmetryICLR 2026
Requires access to internal model representations, which is not possible for closed-source/API-only models.
- ReasoningGuard: Safeguarding Large Reasoning Models with Inference-time Safety Aha MomentsACL 2026
Requires white-box access to internal attention weights, so it cannot be applied to closed-source or heavily quantized models without attention access (quantization only briefly ablated)
- Real-Time Monitoring and Calibration of Chain-of-Thought Sycophancy in Large Reasoning ModelsICML 2026
Requires access to model internal activations and layer selection, so it cannot be applied to API-only models; the inference-time calibration infrastructure itself is noted as a potential misuse channel
- Reading Between the Tokens: Improving Preference Predictions through Mechanistic ForecastingICML 2026
Requires white-box access to model internals and compute for probe training, excluding API-only or closed models
- ReFL: Reflective Feedback Learning for Hallucination Detection of Large Language ModelsACL 2026
Requires white-box access to internal hidden states, so it cannot apply to closed-source models like GPT-4o or Gemini
- RFS-Guard: Detecting Reasoning Hallucinations via Cross-Phase Routing Focus in Large Reasoning ModelsACL 2026
Requires white-box access to attention weights and hidden states, so it cannot be applied to closed-source API models
- Query-Level Uncertainty in Large Language ModelsICLR 2026
Requires white-box access to internal hidden states and the unembedding matrix, so it cannot be applied to closed black-box APIs
- Programming over Thinking: Efficient and Robust Multi-Constraint PlanningACL 2026
Relies on strong LLM coding and formatted-output abilities; only closed-source models were evaluated, so applicability to open-source models is untested.
- ProSafePrune: Projected Safety Pruning for Mitigating Over-Refusal in LLMsICLR 2026
Requires white-box access to model weights and hidden representations, so it cannot be applied to API-only black-box models
- PrefixNLI: Detecting Factual Inconsistencies as Soon as They AriseACL 2026
Requires access to model logits for token distribution modification, so it cannot be applied to closed-source API models.
- Predicting LLM Output Length via Entropy-Guided RepresentationsICLR 2026
Requires access to the serving model's internal hidden states and extra forward passes over the vocabulary to compute per-token entropies, limiting applicability to black-box or API-only models
- N-GLARE: An Non-Generative Latent Representation-Efficient LLM Safety EvaluatorACL 2026
Requires white-box access to internal hidden representations, so it cannot evaluate API-only black-box models
- Mitigating Safety Fallback in Editing-based Backdoor Injection on LLMsICLR 2026
Requires full white-box access to model weights, making it inapplicable to proprietary or closed-source API-only models.
- Mitigating Multimodal Hallucinations via Gradient-based Self-ReflectionCVPR 2026
Requires white-box access to model gradients, excluding API-only or closed models
- Mitigating Action-Relation Hallucinations in LVLMs via Relation-aware Visual EnhancementACL 2026
Requires white-box access to internal attention layers and weights, so it cannot be applied to closed-source API-only models
- MPR-GUI: Benchmarking and Enhancing Multilingual Perception and Reasoning in GUI AgentsACL 2026
Inapplicable to closed-source or API-only models (e.g., Gemini) because it requires direct access and modification of internal hidden states during inference.
- Logical Consistency as a Bridge: Improving LLM Hallucination Detection via Label Constraint Modeling between Responses and Self-JudgmentsACL 2026
Requires access to LLM internals (hidden states, logits, attention), so it is only usable by the LLM provider, not third-party users
- Less Languages, Less Tokens: An Efficient Unified Logic Cross-lingual Chain-of-Thought Reasoning FrameworkACL 2026
Requires white-box access to hidden states at an internal layer, so it cannot be applied to black-box LLM APIs
- Leave My Images Alone: Preventing Multi-Modal Large Language Models from Analyzing Images via Visual Prompt InjectionACL 2026
Requires white-box access (gradients) to target MLLMs; not directly applicable to closed-source APIs
- Language Model Networks: Supervision-Efficient Learning through Dense CommunicationICML 2026
Requires white-box access to model internals and gradient optimization, so it cannot be built on black-box APIs; dense messages are not human-interpretable
- Landscape of Thoughts: Visualizing the Reasoning Process of Large Language ModelsICLR 2026
Incompatible with closed-source models (e.g., GPT-4, Gemini) that do not provide public log-likelihood/perplexity estimations.
- JULI: Jailbreak Large Language Models by Self-IntrospectionICLR 2026
Strictly requires APIs to return top-k token log probabilities, which is a feature not supported by all proprietary providers.
- How Do Medical MLLMs Fail? A Study on Visual Grounding in Medical ImagesICLR 2026
Requires access to internal attention weights, making it inapplicable to closed-source API-only models.
- Grounding or Guessing? Visual Signals for Detecting Hallucinations in Sign Language TranslationICLR 2026
Requires access to internal model states like hidden layers and cross-attention weights, precluding use with black-box APIs.
- From Token to Token Pair: Efficient Prompt Compression for Large Language Models in Clinical PredictionICML 2026
Requires direct modification of the tokeniser and embedding matrix, which is impractical for closed-source models or deployments with fixed vocabularies
- Expert Heads: Robust Evidence Identification for Large Language ModelsICLR 2026
Assumes access to internal attention weights, which is not possible for closed-source models through standard APIs.
- Enhancing Hallucination Detection through Noise InjectionICLR 2026
Needs white-box access to intermediate activations, making it incompatible with closed-source, API-only models
- Eliciting Numerical Predictive Distributions of LLMs Without Auto-RegressionICLR 2026
Requires access to internal model activations, and probes are model-specific — must be retrained for each architecture or tokenization scheme.
- Efficient LLM Moderation with Multi-Layer Latent PrototypesICML 2026
Requires full white-box access to intermediate activations, preventing use with closed-source, API-only commercial LLMs.
- Efficient Hallucination Detection for LLMs Using Uncertainty-Aware Attention HeadsICML 2026
Requires white-box access to attention weights and token probabilities, so it cannot be applied to black-box APIs
- ET-Agent: Incentivizing Effective Tool-Integrated Reasoning Agent via Behavior CalibrationACL 2026
Framework only demonstrated on 7-8B scale models; scaling to larger architectures untested due to resource constraints
- DiSRouter: Distributed Self-Routing for LLM SelectionsICLR 2026
Requires the ability to fine-tune model weights (SFT/RL), making it inapplicable to closed-source proprietary APIs.
- Detecting Misbehaviors of Large Vision-Language Models by Evidential Uncertainty QuantificationICLR 2026
Requires white-box access to internal pre-logit representations and output-head weights, so it cannot be applied to closed-source APIs (e.g. GPT-4)
- DecodeShare: Tracing the Shared Subspace of LLM Decode-Time DecisionsICML 2026
Requires full white-box access to model activations and cannot be deployed on black-box model APIs.
- Cross-Modal Attention Calibration for LVLM Hallucination MitigationCVPR 2026
Requires access to internal attention logits and value vectors, making it inapplicable to black-box API-based models.
- Critical-CoT: A Robust Defense Framework against Reasoning-Level Backdoor Attacks in Large Language ModelsACL 2026
Requires fine-tuning model parameters (SFT and DPO), making it inapplicable to black-box or API-only models and computationally intensive for very large models.
- Controlling Distributional Bias in Multi-Round LLM Generation via KL-Optimized Fine-TuningACL 2026
Requires adding tokens to the vocabulary and accessing model weights, so it cannot be applied to closed-source models
- Confidence Should Be Calibrated More Than One Turn DeepACL 2026
Requires white-box access to hidden states, so it cannot apply to closed-source models like GPT-4/GPT series
- CoSToM: Causal-oriented Steering for Intrinsic Theory-of-Mind Alignment in Large Language ModelsACL 2026
Requires access to open-source model weights, internal activations, and gradients, excluding API-only models.
- Can Reasoning Path still be Effective as Input? Bridging Post-Reasoning to Chain-of-Thought CompressionACL 2026
Relies on executor-generated CoT data and requires training both compressor and executor jointly, so it cannot be applied to models without weight access
- Bridging Internal Consistency and External Alignment: A Causal and Dynamic Interpretability Framework for LLM GenerationACL 2026
Requires token-level logit access, so it cannot be applied directly to fully closed-source models
- Beyond Logits: Metastable Latent Dynamics for Sample-Efficient Best-of-N Selection in LLMsICML 2026
Requires white-box access to hidden states and replayed forward passes, so it cannot be used with closed-source/API models
- Backdoor Collapse: Eliminating Unknown Threats Via Known Backdoor Aggregation In Language ModelsACL 2026
Requires white-box access to model weights and the ability to fine-tune, so it does not apply to API-only or inference-time settings
- Attention Sinks as Internal Signals for Hallucination Detection in Large Language ModelsICML 2026
Requires access to raw attention weights, restricting use to open-weight models; efficient deployment would need a specialized kernel to avoid materializing full attention matrices
- Attention Basin: Why Contextual Position Matters in Large Language ModelsACL 2026
Inapplicable to closed-source models that do not expose internal attention scores via API.
- Adversarial Style Optimization: Enhancing VLM Jailbreaks by GRPO-based Stylistic Triggers OptimizationCVPR 2026
Requires white-box access signals: logit-based refusal probabilities from the target model (or a classifier proxy), which is unavailable for strictly black-box APIs
- AdaFuse: Adaptive Ensemble Decoding for Large Language ModelsACL 2026
Requires access to token-level likelihoods and tokenizer outputs, so it cannot be used with closed-source or black-box LLM APIs
- AFTER: Mitigating the Object Hallucination of LVLM via Adaptive Factual-Guided Activation EditingICLR 2026
Requires access to internal activations of open-source LVLMs, so it cannot be applied to closed-source models
- When Big Models Train Small Ones: Label-Free Model Parity Alignment for Efficient Visual Question Answering using Small VLMsEMNLP 2025
Requires access to a large VLM for annotation and validation, which is itself costly or API-dependent; using proprietary APIs (GPT-4o) compromises reproducibility due to opaque and potentially changing behavior
- Weighted Multi-Prompt Learning with Description-free Large Language Model DistillationICLR 2025
Depends on closed-source OpenAI embedding APIs with undisclosed training data/architecture, limiting reproducibility and control
- WebInject: Prompt Injection Attack to Web AgentsEMNLP 2025
Requires white-box access to the MLLM parameters used by the agent; no transferability to closed-source MLLMs was evaluated
- Unveiling the Response of Large Vision-Language Models to Visually Absent TokensEMNLP 2025
Requires access to internal FFN activations, so it is inapplicable to closed-source LVLMs
- Unlocking General Long Chain-of-Thought Reasoning Capabilities of Large Language Models via Representation EngineeringACL 2025
Requires white-box access to internal hidden states, so it cannot be applied to closed-source/API LLMs
- Understanding and Mitigating Overrefusal in LLMs from an Unveiling Perspective of Safety Decision BoundaryEMNLP 2025
Requires white-box access to model hidden states, so it cannot be applied to proprietary or strictly black-box models
- TokenSkip: Controllable Chain-of-Thought Compression in LLMsEMNLP 2025
Requires a supervised fine-tuning stage for each target model, making it inapplicable to closed-source black-box APIs.
- Thought Communication in Multiagent CollaborationNeurIPS 2025
Requires white-box access to model hidden states, precluding use with closed-source API-based LLMs.
- TRACT: Regression-Aware Fine-tuning Meets Chain-of-Thought Reasoning for LLM-as-a-JudgeACL 2025
Requires access to token-level output probabilities, so the method cannot be applied to proprietary API-only models
- Steer LLM Latents for Hallucination DetectionICML 2025
Requires white-box access to internal latent states of open-source LLMs; not applicable to closed APIs, and OOD transfer loses ~4-8 AUROC (e.g., 79.8% vs 84.0% on TriviaQA)
- Sparse Activation Editing for Reliable Instruction Following in NarrativesEMNLP 2025
Requires publicly available pre-trained SAEs for the target model, so it cannot be applied to models lacking such checkpoints.
- Semantic Representation Attack against Aligned Large Language ModelsNeurIPS 2025
Requires logit/probability access to the target model, so it does not apply to closed-source API models
- SafeQuant: LLM Safety Analysis via Quantized Gradient InspectionNAACL 2025
White-box requirement: needs access to model weights and gradients, so it cannot be applied to commercial models like GPT-4 or Claude
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore MechanismNeurIPS 2025
Requires access to intermediate hidden states, so it cannot be applied to black-box models like GPT-4
- SHIFT: Smoothing Hallucinations by Information Flow Tuning for Multimodal Large Language ModelsICCV 2025
Requires white-box access to internal feature encodings, making it incompatible with closed-source API-based models
- Robust Hallucination Detection in LLMs via Adaptive Token SelectionNeurIPS 2025
Requires white-box access to the LLM's internal hidden states and predictive logits.
- Representation Consistency for Accurate and Coherent LLM Answer AggregationNeurIPS 2025
Requires white-box access to internal activations and pretrained SAEs, so it cannot be applied to proprietary API-only LLMs
- RepGuard: Adaptive Feature Decoupling for Robust Backdoor Defense in Large Language ModelsNeurIPS 2025
Requires access to the training/fine-tuning pipeline to apply the multi-objective optimization, precluding its use as a post-hoc, inference-only defense.
- Putnam-AXIOM: A Functional & Static Benchmark for Measuring Higher Level Mathematical Reasoning in LLMsICML 2025
TFA requires log-probability access, so proprietary models cannot be evaluated, and it under-scores models with valid but stylistically different solutions or tool-use/code fine-tuning
- Personality Alignment of Large Language ModelsICLR 2025
The method requires white-box access to model activations, preventing its use with closed-source API-only models.
- Neuron based Personality Trait Induction in Large Language ModelsICLR 2025
Requires access to internal neuron activations, making it inapplicable to closed-source API models
- NeuroGenPoisoning: Neuron-Guided Attacks on Retrieval-Augmented Generation of LLM via Genetic Optimization of External KnowledgeNeurIPS 2025
Requires white-box access to neuron activations and Integrated Gradients attribution (forward passes plus attribution analysis), so it does not apply to black-box or API-only models
- Mitigating Hallucinations in Vision-Language Models through Image-Guided Head SuppressionEMNLP 2025
Requires access to model weights, so it cannot be applied to closed-source or API-based models; head-selection hyperparameters must be searched per model/task
- Meta-Reflection: A Feedback-Free Reflection Learning FrameworkACL 2025
Requires parameter access to the LLM (intermediate hidden states and tunable codebook), so it cannot apply to API-only models like ChatGPT or Claude
- MULTIGUARD: An Efficient Approach for AI Safety Moderation Across Languages and ModalitiesEMNLP 2025
Requires access to internal embeddings, so it is inapplicable to closed-source/ API-only models
- MM-IFEngine: Towards Multimodal Instruction FollowingICCV 2025
Relies on proprietary models (GPT-4o) for initial task generation and as a comparative judge for subjective constraints, potentially introducing model-specific biases
- MA-GTS: A Multi-Agent Framework for Solving Complex Graph Problems in Real-World ApplicationsEMNLP 2025
Framework requires many sequential LLM API calls and substantial computational resources for large-scale problems, and open-source model tool-calling is too weak for stable operation
- Logic-Regularized Verifier Elicits Reasoning from LLMsACL 2025
Requires white-box access to hidden states, so it cannot be applied to black-box LLMs
- Lines of Thought in Large Language ModelsICLR 2025
Requires access to hidden states, restricting it to open-source models; heavily fine-tuned or re-aligned models might exhibit different patterns
- Latent Space Chain-of-Embedding Enables Output-free LLM Self-EvaluationICLR 2025
Requires white-box access to all intermediate hidden states, so it cannot be applied to closed-source models like GPT-4
- Large Language Models Have Intrinsic Meta-Cognition, but Need a Good LensEMNLP 2025
Requires access to model internals (hidden states, logits), restricting hidden-state-based lenses to open-source models; not applicable to closed-source APIs like GPT-4
- Injecting Universal Jailbreak Backdoors into LLMs in MinutesICLR 2025
Requires white-box access to model parameters and internal activations, making it inapplicable to closed-source LLMs like GPT-4
- Improving Preference Extraction In LLMs By Identifying Latent Knowledge Through Classifying ProbesACL 2025
Requires access to internal model activations/hidden states, precluding use with closed-source API-only models.
- Idiosyncrasies in Large Language ModelsICML 2025
Does not handle an unknown or very large set of candidate source LLMs; the setup assumes a fixed, known list of models to classify among
- Identifying Pre-training Data in LLMs: A Neuron Activation-Based Detection FrameworkEMNLP 2025
Requires white-box access to model weights and internal activations, so it cannot be applied to closed-source models or API-only LLMs
- ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMsACL 2025
Requires white-box access to hidden states and attention scores, so it cannot be applied to proprietary/API-only models
- Hidden No More: Attacking and Defending Private Third-Party LLM InferenceICML 2025
Requires the adversary to know the specific architecture and layer from which hidden states were extracted.
- Has this Fact been Edited? Detecting Knowledge Edits in Language ModelsNAACL 2025
Requires white-box access to hidden states or the full probability distribution, unavailable for closed-source LLMs
- GraphRouter: A Graph-based Router for LLM SelectionsICLR 2025
Relies on current LLMs to model numerical attributes like token pricing and context length, which they handle poorly; these details significantly affect selection quality
- From Sub-Ability Diagnosis to Human-Aligned Generation: Bridging the Gap for Text Length Control via MarkerGenACL 2025
Only applicable to open-source models; cannot be used with closed-source APIs since it requires injecting markers into the decoding process
- From Neurons to Semantics: Evaluating Cross-Linguistic Alignment Capabilities of Large Language Models via Neurons AlignmentACL 2025
Requires white-box access to FFN activations, so it cannot be applied to closed-source models
- FlexAC: Towards Flexible Control of Associative Reasoning in Multimodal Large Language ModelsNeurIPS 2025
Requires white-box access to hidden states; not applicable to black-box models like ChatGPT
- FacLens: Transferable Probe for Foreseeing Non-Factuality in Fact-Seeking Question Answering of Large Language ModelsEMNLP 2025
Requires white-box access to LLM parameters/hidden states, so it cannot apply to API-based black-box LLMs
- Exploiting the Shadows: Unveiling Privacy Leaks through Lower-Ranked Tokens in Large Language ModelsACL 2025
Relies on access to top-k logits or the ability to apply logit-bias, which may be restricted in some commercial APIs.
- Exploiting Edited Large Language Models as General Scientific OptimizersNAACL 2025
Requires access to model weights for model editing, so it cannot be applied to closed-source models like GPT-4 or Claude 3.5
- EverTracer: Hunting Stolen Large Language Models via Stealthy and Robust Probabilistic FingerprintEMNLP 2025
Requires gray-box access to token-level log probabilities, making it inapplicable to strict black-box APIs that only provide text output.
- Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model FeaturesICCV 2025
Requires white-box access to model internals (attention head activations), so it is not applicable to closed-source models like GPT-4 or Gemini
- Eliminating Position Bias of Language Models: A Mechanistic ApproachICLR 2025
Requires low-level access to the model's attention masks and position encoding logic, complicating deployment in closed-source API environments.
- ETF: An Entity Tracing Framework for Hallucination Detection in Code SummariesACL 2025
Depends on large LLMs (GPT-4-Omni, Gemini) for NER and verification; smaller open-source models perform near random, making the pipeline costly and hard to scale
- Dual-perspective Data Augmentation and Curriculum Learning Framework for Low-resource Complex Named Entity RecognitionSIGIR 2025
Depends on a commercial LLM (gpt-3.5-turbo) for augmentation, adding cost and reproducibility concerns
- Detoxifying Large Language Models via Autoregressive Reward Guided Representation EditingNeurIPS 2025
Requires white-box access to internal representations; not applicable to API-only models
- Detecting LLM Hallucination Through Layer-wise Information Deficiency: Analysis of Ambiguous Prompts and Unanswerable QuestionsEMNLP 2025
Requires access to per-layer hidden states and the LM head of the model, precluding closed-source API models
- Detecting High-Stakes Interactions with Activation ProbesNeurIPS 2025
Requires white-box access to internal activation tensors, precluding use with closed-source models accessed via black-box APIs.
- DeAL: Decoding-time Alignment for Large Language ModelsACL 2025
Requires access to output logits, so it cannot be applied to most proprietary API models
- DSVD: Dynamic Self-Verify Decoding for Faithful Generation in Large Language ModelsEMNLP 2025
Requires white-box access to model hidden states, making it inapplicable to closed-source API models
- Correlation Dimension of Autoregressive Large Language ModelsNeurIPS 2025
Requires access to full logits/log-probabilities, making it inapplicable to closed models via APIs as-is (authors only speculate about integration)
- Communicating Activations Between Language Model AgentsICML 2025
Requires full white-box access to model activations, making it incompatible with black-box API-only models.
- Collaborative Beam Search: Enhancing LLM Reasoning via Collective ConsensusEMNLP 2025
Relies on the availability of log-probabilities from all participating models, precluding use with black-box APIs that do not provide token-level scores.
- CoCoA: Confidence- and Context-Aware Adaptive Decoding for Resolving Knowledge Conflicts in Large Language ModelsEMNLP 2025
Requires full access to token-level probability distributions (logits), precluding use with black-box APIs that only return text.
- ChronoSense: Exploring Temporal Understanding in Large Language Models with Time Intervals of EventsACL 2025
Model coverage is small and includes closed-source APIs (GPT-4o), limiting reproducibility; prompt phrasing may influence results
- Can an Individual Manipulate the Collective Decisions of Multi-Agents?EMNLP 2025
Requires white-box access to model weights for gradient computation, limiting use against closed-source API-based agents.
- Calibrating LLM Confidence by Probing Perturbed Representation StabilityEMNLP 2025
The method requires white-box access to internal hidden states and the ability to compute gradients, making it incompatible with black-box LLM APIs.
- Bridging the Language Gaps in Large Language Models with Inference-Time Cross-Lingual InterventionACL 2025
Requires white-box access to internal layers and hidden states, so it cannot be applied to closed-source/API-only models like GPT-4o
- Beyond the Surface: Enhancing LLM-as-a-Judge Alignment with Human via Internal RepresentationsNeurIPS 2025
Requires white-box access to intermediate-layer hidden states, so it does not apply to API-only models (only the simplified E-Score variant works there)
- Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViTNeurIPS 2025
Requires white-box access to all LLM internal hidden states, so it cannot be applied to API-only models
- Beyond Input Activations: Identifying Influential Latents by Gradient Sparse AutoencodersEMNLP 2025
Requires white-box access to the model to compute gradients, preventing use with closed-source API-based LLMs
- Augmenting Multi-Agent Communication with State Delta TrajectoryEMNLP 2025
Incompatible with black-box or API-based LLMs as it requires direct access to and modification of internal activations.
- AdvWave: Stealthy Adversarial Jailbreak Attack against Large Audio-Language ModelsICLR 2025
Requires white-box access to ALM weights, encoder, and cluster centers, limiting applicability to closed APIs
- Adaptive Tool Use in Large Language Models with Meta-Cognition TriggerACL 2025
Requires white-box access to model hidden states, making it inapplicable to closed-source API-only models.
- A Training-free LLM-based Approach to General Chinese Character Error CorrectionACL 2025
Requires access to model probability distributions, so it cannot be applied directly to API-accessed LLMs
- Word Embeddings Are Steers for Language ModelsACL 2024
Requires direct access to output word embeddings, so it cannot be applied to LM APIs that hide embeddings
- Trusting Your Evidence: Hallucinate Less with Context-aware DecodingNAACL 2024
Requires access to output logits, so it cannot be applied to black-box API models like ChatGPT/GPT-4
- Transferable and Efficient Non-Factual Content Detection via Probe Training with Offline Consistency CheckingACL 2024
Requires access to internal representations, so only works with open-source LLMs, not API-only models
- Toward Adaptive Reasoning in Large Language Models with Thought RollbackICML 2024
Gains depend on LLM capability: Llama2-70b with TR improves only ~4% over baselines, and for GPT-3.5-turbo on AQuA-RAT/MATH adding CoT examples to the TR prompt decreases performance because weak LLMs cannot distinguish the different instruction types.
- Thermometer: Towards Universal Calibration for Large Language ModelsICML 2024
Requires access to the internal hidden states/feature extractor of the LLM, making it inapplicable to black-box models served via restricted APIs.
- Rethinking Tabular Data Understanding with Large Language ModelsNAACL 2024
Only GPT-3.5 was used due to budget constraints, so generalizability to other LLMs is untested
- ResumeFlow: An LLM-facilitated Pipeline for Personalized Resume Generation and RefinementSIGIR 2024
Depends on closed-source commercial LLMs (GPT-4, Gemini) with associated cost and no open-source model support yet
- Reasoning in Flux: Enhancing Large Language Models Reasoning through Uncertainty-aware Adaptive GuidanceACL 2024
Requires access to token probability distributions, so it cannot be applied to closed-source models like ChatGPT or Claude
- Null-Shot Prompting: Rethinking Prompting Large Language Models With HallucinationEMNLP 2024
Effects are highly model-dependent: GPT-4 Turbo, Claude 3 Sonnet, and Claude 3 Opus mostly lose performance (e.g., Claude 3 Opus -92.85% StrategyQA, -99.11% WinoGrande), so gains cannot be generalized across LLMs
- Multimodal Task Vectors Enable Many-Shot Multimodal In-Context LearningNeurIPS 2024
Requires white-box access to model internals, so it cannot be applied to proprietary models like GPT-4 or Gemini
- Model Internals-based Answer Attribution for Trustworthy Retrieval-Augmented GenerationEMNLP 2024
Requires white-box access to model internals (gradients and log-probabilities), making it incompatible with closed-source APIs
- MACM: Utilizing a Multi-Agent System for Condition Mining in Solving Complex Mathematical ProblemsNeurIPS 2024
Requires many more LLM queries than simpler prompting, though this buys higher error-correction ceiling (~90% of GPT-4 Turbo's original errors corrected at high query counts vs ~20% for I-O/CoT/SC-CoT)
- LoRA-Guard: Parameter-Efficient Guardrail Adaptation for Content Moderation of Large Language ModelsEMNLP 2024
Requires access to chat model weights, so it cannot be applied to black-box/proprietary LLM systems.
- Learning to Correct for QA Reasoning with Black-box LLMsEMNLP 2024
Effectiveness depends heavily on the open-source LLM used for initialization: with LLaMA-2-7B the adapter fails to beat gpt-3.5-turbo (e.g., 63.13 on StrategyQA vs 70.92 unadapted), whereas Mistral-7B works
- LTGC: Long-tail Recognition via Leveraging LLMs-driven Generated ContentCVPR 2024
Depends on proprietary models (GPT-4, GPT-4V, DALL-E), so reproducibility and cost are concerns and results hinge on those APIs
- Instructional Fingerprinting of Large Language ModelsNAACL 2024
The most effective 'adapter' variant requires white-box access to the user model's embedding weights, limiting its use against restricted APIs.
- Improving Text Embeddings with Large Language ModelsACL 2024
Relies on manual prompt engineering and proprietary LLMs (GPT-4) for data generation; no open-source model tested for generation
- Image Hijacks: Adversarial Images can Control Generative Models at RuntimeICML 2024
Requires white-box access to the model's gradients, making it difficult to apply directly against proprietary black-box APIs.
- Forgetting Curve: A Reliable Method for Evaluating Memorization Capability for Long-Context ModelsEMNLP 2024
Applying the method to closed-source commercial APIs is computationally intractable due to the requirement of simulating token-by-token teacher forcing without direct model parameter access.
- Faithful Logical Reasoning via Symbolic Chain-of-ThoughtACL 2024
Requires a capable backbone LLM — improvement is largest with GPT-4 (up to +24.3% from GPT-3.5 to GPT-4), so gains shrink with weaker models
- Explicit Memory Learning with Expectation MaximizationEMNLP 2024
Requires access to internal model outputs (token probabilities/perplexity) for entropy and validation-likelihood scoring, so it cannot be applied to commercial API models like GPT or Claude
- Exogenous and Endogenous Data Augmentation for Low-Resource Complex Named Entity RecognitionSIGIR 2024
Depends on a proprietary external LLM (gpt-3.5-turbo via OpenAI API) whose generated-data quality and self-reflection reliability are unverified beyond the LLM's own judgment
- Enhancing In-Context Learning via Implicit Demonstration AugmentationACL 2024
Requires access to the final fully-connected layer weights, so it is only applicable to open-source models, not black-box APIs
- Emulated Disalignment: Safety Alignment for Large Language Models May Backfire!ACL 2024
Requires full access to output token distributions (logits) over the entire vocabulary, making it inapplicable to most restricted-access black-box APIs.
- Embedding and Gradient Say Wrong: A White-Box Method for Hallucination DetectionEMNLP 2024
Requires full white-box access to embeddings and gradients, excluding closed API models; authors suggest LLM owners could provide this as a service
- Don’t Just Say “I don’t know”! Self-aligning Large Language Models for Responding to Unknown Questions with ExplanationsEMNLP 2024
Requires access to fine-tune the base model, so it cannot be applied to black-box LLMs like ChatGPT
- DocKD: Knowledge Distillation from LLMs for Open-World Document Understanding ModelsEMNLP 2024
Relies on external closed-source components (AWS Textract API, Claude-2), and no integration with large multimodal models like GPT-4V was explored
- DVD: Dynamic Contrastive Decoding for Knowledge Amplification in Multi-Document Question AnsweringEMNLP 2024
Requires white-box access to per-input logits, so it cannot be applied to closed-source models like ChatGPT or GPT-4
- DRAGIN: Dynamic Retrieval Augmented Generation based on the Real-time Information Needs of Large Language ModelsACL 2024
Requires access to self-attention scores and token-level output distributions, so it cannot be used with black-box APIs that do not expose internal states
- DDPrompt: Differential Diversity Prompting in Large Language ModelsACL 2024
Only evaluated on GPT-3.5-turbo; generalization to other LLMs is untested
- Confidence Under the Hood: An Investigation into the Confidence-Probability Alignment in Large Language ModelsACL 2024
Requires access to token-level log probabilities, making it inapplicable to many proprietary black-box APIs.
- Accelerating Greedy Coordinate Gradient and General Prompt Optimization via Probe SamplingNeurIPS 2024
Restricted to white-box open-source models as it requires access to internal losses and gradients.
- Training Chain-of-Thought via Latent-Variable InferenceNeurIPS 2023
Requires gradient access to the model, excluding proprietary APIs without fine-tuning endpoints (e.g., GPT-4)
- Self-Detoxifying Language Models via Toxification ReversalEMNLP 2023
Requires full white-box access to model internals during forward passes, so it cannot be applied to API-only models
- LLMScore: Unveiling the Power of Large Language Models in Text-to-Image Synthesis EvaluationNeurIPS 2023
Relies on proprietary GPT-4 via paid API, limiting accessibility and reproducibility
- Improving Transformer-based Program Repair Model through False Behavior DiagnosisEMNLP 2023
Requires high-resolution access to internal model states (attention maps and value vectors), precluding use with closed-box APIs.
- Automatically Auditing Large Language Models via Discrete OptimizationICML 2023
Requires white-box access (forward passes and gradients); auditing API-only models and certifying absence of failure modes are left unresolved
Nearest existing work
- AlienLM: Alienization of Language for API-Boundary Privacy in Black-Box LLMsICML 2026
- Interpretation Meets Safety: A Survey on Interpretation Methods and Tools for Improving LLM SafetyEMNLP 2025
- Transferable Direct Prompt Injection via Activation-Guided MCMC SamplingEMNLP 2025
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language ModelsICLR 2025
- JULI: Jailbreak Large Language Models by Self-IntrospectionICLR 2026
- Speak Easy: Eliciting Harmful Jailbreaks from LLMs with Simple InteractionsICML 2025
- Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal TrainingACL 2025
- RepIt: Steering Language Models with Concept-Specific Refusal VectorsICLR 2026
- Jailbreak-Tuning: Models Efficiently Learn Jailbreak SusceptibilityEMNLP 2025
- Jailbreak Transferability Emerges from Shared RepresentationsICLR 2026
- Detecting High-Stakes Interactions with Activation ProbesNeurIPS 2025
- Distract Large Language Models for Automatic Jailbreak AttackEMNLP 2024
- You Can't Eat Your Cake and Have It Too: The Performance Degradation of LLMs with Jailbreak DefenseTheWebConference 2025
- Tracing the Dynamics of Refusal: Exploiting Latent Refusal Trajectories for Robust Jailbreak DetectionICML 2026
- The First to Know: How Token Distributions Reveal Hidden Knowledge in Large Vision-Language Models?ECCV 2024
Related open problems
Model Safety, Alignment & Jailbreaks
Decision-Only and Logit-Free Safety Auditing and Alignment Evaluation for Black-Box Models
Current safety auditing, jailbreak discovery, and representation-based alignment diagnostics overwhelmingly rely on white-box preconditions such as gradient access, intermediate activations, or output token logit distributions. When models are deployed behind production APIs that return only generated text, practitioners cannot run these diagnostic and red-teaming pipelines directly. Relying on surrogate transfer from open-source models produces high false-negative rates due to incomplete transferability across distinct architectures and tokenizers. Consequently, safety assessments for proprietary, API-governed models remain structurally disconnected from the methods developed in the literature.
Model Safety, Alignment & Jailbreaks
Recovering Internal Attention and Activation Signals from API-Visible Model Behavior
The shared wall is a dependence on inference-time extraction of layer-wise internal states — attention weights, hidden activations, per-layer execution — which are unavailable outside open-weight deployments and costly and per-model-fragile even where weights are open. Every technique built on those signals stops working the moment the model is served through an API; the sanctioned fallbacks, transferring from an open proxy or degrading to logit-only features, demonstrably forfeit most of the signal (one published detector drops from its white-box ceiling to 0.66 AUROC). Even with open weights, each release re-breaks the tooling: layer indices must be re-selected per model and multi-pass decoding adds latency. The result is a method family whose reach shrinks exactly as the most capable models become less open, and whose findings nobody without weight access can verify or use.
Model Safety, Alignment & Jailbreaks
Robust Evaluation of Non-Binary and Phrasing-Agnostic Refusal in Safety-Aligned Language Models
Current safety benchmarks and verifiable reward mechanisms evaluate model refusals primarily through binary keyword matching, hand-crafted refusal templates, or uncalibrated model-based autoraters (which show inter-annotator agreement as low as $\alpha = 0.378$). This creates a blind spot where models that partially comply, use non-standard refusal vocabulary, or alter their stance under multi-turn pushback are misclassified as either fully compliant or safely refusing. Consequently, alignment interventions optimized against these brittle metrics suffer from artificial over-refusal on benign queries while leaking safety risks through stylistic variation and conversational pressure. Without a rigorous, non-binary evaluation standard across diverse phrasing distributions, researchers cannot reliably assess whether safety alignment generalizes beyond rigid heuristic templates.
Model Safety, Alignment & Jailbreaks
Safety Monitoring and Alignment Interventions Under Black-Box and API-Only Constraints
Current state-of-the-art alignment monitoring, jailbreak detection, and activation-steering interventions explicitly require full white-box access to hidden state activations, KV caches, attention logits, or gradient flows. In practice, third-party safety auditors, downstream system builders, and end users interact with models exclusively via black-box query APIs or restricted endpoints. Because existing methods are structurally coupled to internal state inspection and modification, they cannot be deployed or evaluated on proprietary commercial models or in privacy-preserving environments.