In-Context Explainability 360 (ICX360)

• ICX360 is a context-sensitive explainability toolkit that links LLM outputs to user-provided prompts.
• It integrates multiple methods, including perturbation-based and gradient-based techniques, to assess prompt influence on generated texts.
• It offers a modular, 360-degree approach by addressing stakeholder-specific needs in various LLM applications like summarization and jailbreaking.

In-Context Explainability 360 (ICX360) is an open-source Python toolkit for explaining LLMs with a focus on the user-provided context, or prompts in general, that are fed to the LLMs (Wei et al., 14 Nov 2025). It was introduced for settings in which LLMs produce summaries, answers, lists, recommendations, or other generated outputs whose dependence on prompt context must be inspected in a technically grounded way (Wei et al., 14 Nov 2025). The toolkit contains implementations for three recent tools that explain LLMs using both black-box and white-box methods, via perturbations and gradients respectively, and it includes quick-start guidance materials as well as detailed tutorials covering retrieval augmented generation, natural language generation, and jailbreaking (Wei et al., 14 Nov 2025). More broadly, ICX360 also denotes a 360-degree view of explainability in which explanation is treated as a context-sensitive, audience-aware, and workflow-embedded capability rather than as a single static artifact (Wei et al., 14 Nov 2025, Beaudouin et al., 2020, Bello et al., 6 Jun 2025).

1. Definition and historical position

ICX360 explains LLM-generated text outputs in terms of the input text provided to the model, especially the user-supplied context or prompt, though explanations can also target any part of the prompt, including instructions or system prompts (Wei et al., 14 Nov 2025). Its central concern is therefore not only model internals, but the relation between prompt content and generated output in deployed LLM applications such as summarization, question answering, and safety analysis (Wei et al., 14 Nov 2025).

In historical terms, ICX360 can be read as a prompt- and context-centered successor to AI Explainability 360, an earlier open-source toolkit featuring ten diverse and state-of-the-art explainability methods and two evaluation metrics, organized around the principle that “One Explanation Does Not Fit All” (Arya et al., 2021). That earlier toolkit was built for predictive models and stakeholder diversity; ICX360 shifts the focus to generative models, prompt-conditioned behavior, and explanation of full text generations rather than only class labels or scores (Arya et al., 2021, Wei et al., 14 Nov 2025). This suggests a change in the explanatory object itself: from model decisions in general ML pipelines to context-conditioned language generation in LLM pipelines.

The toolkit was introduced against the background of increasing LLM use in higher-stakes applications, including summarizing meeting transcripts and answering doctors’ questions (Wei et al., 14 Nov 2025). Its stated purpose is therefore both practical and infrastructural: to provide a reusable explanation layer for LLM systems that are often accessed through APIs, used with long natural-language prompts, and deployed in settings where understanding prompt influence is as important as understanding model architecture (Wei et al., 14 Nov 2025).

2. Conceptual foundations of the “360” orientation

The conceptual logic behind ICX360 aligns with several earlier explainability frameworks that reject one-size-fits-all explanation. A multidisciplinary account of context-specific explainability argues that explanation should begin with contextual factors such as who the audience of the explanation is, the operational context, the level of harm that the system could cause, and the legal and regulatory framework; it then recommends selecting the right levels of global and local explanation outputs while taking into account the costs involved (Beaudouin et al., 2020). A related multi-component framework defines a progression from explicit explanation representation to alternative explanations, knowledge of the explainee, and interactive capability, thereby treating explanation as a layered system capability rather than a single output (Atakishiyev et al., 2020).

This context-sensitive orientation is sharpened by work on grounded explainability, which argues for “building a broader and deeper understanding of Explainability by ‘grounding’ it in the social contexts in which these socio-technical systems operate” and for centering “local communities” and “the worst-off” rather than treating explainability as a universal model-internal property (Singh et al., 2022). In parallel, a multilevel stakeholder framework distinguishes algorithmic and domain-based explainability, human-centered explainability, and social explainability, and assigns LLMs a particularly important role in the social layer, where technical outputs must be translated into accessible natural-language explanations (Bello et al., 6 Jun 2025).

Taken together, these lines of work suggest that the “360” in ICX360 is not merely a branding extension of earlier XAI toolkits. It denotes a broader explanatory stance: explanation should be adapted to access regime, audience, task, deployment setting, and actionability. The toolkit paper itself remains focused on input-based explanation of LLM outputs, but the surrounding literature indicates that a complete ICX360 program would also have to address interactivity, stakeholder differentiation, and governance (Wei et al., 14 Nov 2025, Atakishiyev et al., 2020, Bello et al., 6 Jun 2025).

3. Toolkit architecture and explanation methods

ICX360 is organized around explainer classes, with auxiliary abstractions for model wrappers, text perturbation, scalarization, segmentation, and evaluation (Wei et al., 14 Nov 2025). It integrates with Hugging Face transformers and vLLM-served models using the OpenAI API, and uses spaCy for segmentation (Wei et al., 14 Nov 2025). The main supported methods are MExGen, CELL, and Token Highlighter (Wei et al., 14 Nov 2025).

Method	Access type	Explanation form
MExGen	black-box perturbation-based; text-only and dark-grey-box/logits	importance scores for input segments
CELL	black-box, text-only	a contrastive prompt whose response has a desired property
Token Highlighter	white-box, gradient-based	importance score per input token

MExGen extends perturbation-based attribution methods such as LIME and SHAP to the generative LLM setting and is implemented through modified variants called C-LIME and L-SHAP (Wei et al., 14 Nov 2025). It uses scalarizers to convert full generated outputs into quantities that can be perturbed and compared: ProbScalarizedModel computes the log probability of generating the original output sequence conditioned on different inputs, while TextScalarizedModel computes various similarity metrics between the original output and outputs generated from perturbed inputs (Wei et al., 14 Nov 2025). The implementation exposes a base class MExGenExplainer and subclasses CLIME and LSHAP (Wei et al., 14 Nov 2025).

CELL provides contrastive explanations for LLMs by searching for small prompt modifications that induce a response with a desired contrastive property (Wei et al., 14 Nov 2025). The toolkit implements two search variants, CELL and mCELL, and provides infillers BART_infiller and T5_infiller (Wei et al., 14 Nov 2025). Its scalarizers include PreferenceScalarizer, ContradictionScalarizer, NLIScalarizer, and BleuScalarizer, enabling explanations framed in terms of preferability, contradiction, non-entailment, or controlled output divergence (Wei et al., 14 Nov 2025).

Token Highlighter is a white-box, gradient-based explainer. It computes the log likelihood of the response to the input, takes gradients with respect to each token embedding, and uses the Euclidean norm of that gradient as the token’s importance score (Wei et al., 14 Nov 2025). Segment-level importance can then be obtained by averaging token scores over the segment (Wei et al., 14 Nov 2025). In the toolkit’s tutorial framing, this is especially relevant to jailbreak analysis, where prompt-critical tokens must be identified (Wei et al., 14 Nov 2025).

4. Access regimes, granularity, and practical workflow

A central contribution of ICX360 is its explicit taxonomy of access regimes. The toolkit distinguishes truly black-box or text-only access, dark-grey-box access with probabilities or logits, white-box gradient-based access, and white-box attention-based access (Wei et al., 14 Nov 2025). The three included methods map onto these regimes as follows: CELL operates in the truly black-box text-only setting, MExGen supports both text-only and dark-grey-box/logits settings, and Token Highlighter requires white-box gradient access (Wei et al., 14 Nov 2025). Attention-based explanations are part of the taxonomy but are not currently implemented (Wei et al., 14 Nov 2025).

ICX360 also treats explanation granularity as a design variable. On the input side, possible units include tokens, words, phrases, sentences, paragraphs, and documents (Wei et al., 14 Nov 2025). On the output side, it distinguishes explanation of individual output tokens from explanation of larger output units, and all three current methods operate at the latter, coarser level (Wei et al., 14 Nov 2025). This is one of its main differences from earlier sequence-generation explanation libraries that fragment explanation across output tokens rather than explaining the full generation or a larger semantic unit (Wei et al., 14 Nov 2025).

The practical workflow implied by the toolkit is modular. A practitioner chooses a model wrapper such as HFModel or VLLMModel, selects an explainer, configures auxiliary components such as scalarizers, infillers, or segmenters, and then evaluates attribution quality with PerturbCurveEvaluator when appropriate (Wei et al., 14 Nov 2025). The wrappers expose a common interface including convert_input, generate, and GeneratedOutput (Wei et al., 14 Nov 2025). This modularity is important because it makes ICX360 usable across both open-weight and API-served models, and across different prompt-conditioned applications (Wei et al., 14 Nov 2025).

The tutorial coverage is itself revealing. Retrieval augmented generation positions explanation around retrieved passages and prompt context; natural language generation treats explanation at the level of full source-conditioned outputs; jailbreaking frames explanation as prompt sensitivity and attack-surface inspection (Wei et al., 14 Nov 2025). This suggests that ICX360 is designed less as a single-method library than as a context-aware explanation shell for multiple LLM workflows.

5. Adjacent research programs and expanding interpretations

ICX360 sits within a broader research movement that uses in-context conditioning, structured prompting, and judge-based evaluation to rethink explainability. “In-Context Explainers” proposes three ICL-based methods—Perturb ICL, Perturb+Guide ICL, and Explain ICL—for using LLMs to explain other black-box predictive models, and concludes that LLMs can generate faithful explanations similar to post hoc explainers (Kroeger et al., 2023). This complements ICX360 by showing that in-context prompting can serve not only as an explanation target but also as an explanation engine.

A second adjacent line appears in multimodal in-context learning. “Explaining is Harder Than Predicting Alone” evaluates concept-based explanations for frozen multimodal LLMs under few-shot ICL using five conditions of increasing formal rigour, ranging from baseline classification to Description Logics axiom generation, and reports that accuracy declines monotonically from 93.8% to 90.1% as explanation complexity increases (Quiles-Ramírez et al., 27 May 2026). The paper’s conclusion that formal syntax alone is not sufficient evidence of good explanation is highly relevant to ICX360, because it separates predictive success from explanation quality and treats judge-scored groundedness, specificity, local discriminativeness, and logical coherence as distinct axes (Quiles-Ramírez et al., 27 May 2026).

A third adjacent direction is explicitly human-centered. “Mind the XAI Gap” uses in-context learning and a benchmark-derived contextual “thesaurus” to generate one response that contains both an expert explanation and a non-expert explanation, reporting a Spearman rank correlation of 0.92 with ground-truth explanations and improved interpretability and human-friendliness in a user study with $N=56$ (Paraschou et al., 13 Jun 2025). This is closely aligned with the layered audience logic that an ICX360 system would require in practice.

Interface work reinforces the same point. IXAII, an interactive explainable intelligent system, integrates LIME, SHAP, Anchors, and DiCE, provides tailored views for five user groups, and gives users agency over explanation content and format (Speckmann et al., 26 Jun 2025). This suggests that a full ICX360 implementation is likely to require not only prompt- and context-level explanation methods, but also interactive orchestration across roles, questions, and rendering modes.

6. Limitations, open problems, and future directions

The current toolkit paper identifies several direct limitations. ICX360 does not yet implement attention-based explanations; more efficient black-box methods are needed, possibly through amortization; different infillers could be explored for CELL and adapted for MExGen; and future work may need to explain more advanced phenomena such as reasoning traces by treating them as additional generated context (Wei et al., 14 Nov 2025). These are toolkit-level limitations tied to method coverage and efficiency.

Broader research raises deeper methodological cautions. Structured, judge-evaluated work on multimodal ICL shows that greater formalism can degrade both prediction and explanation quality, and that models can produce outputs that look formally compliant while failing to encode the right visual concepts (Quiles-Ramírez et al., 27 May 2026). Interactive vision research further argues that static explanations can give a false sense of comprehension and that explanation should be an iterative process of inspection, intervention, and hypothesis testing in context (Kirchler et al., 2021). These results imply that ICX360 cannot be reduced to prompt attribution alone if the goal is faithful, user-relevant explanation in complex systems.

Human-centered deployment adds another unresolved layer. Interactive explanation systems show the importance of role-aware views, switching between reference methods, multiple formats, and a guide to explanation types (Speckmann et al., 26 Jun 2025). Grounded explainability adds that explanations must be situated in real-time socio-political contexts and designed to support local communities, democratic participation, and the well-being of the worst-off (Singh et al., 2022). A plausible implication is that future ICX360 systems will need to connect technical prompt-based explanation to stakeholder adaptation, recourse, and governance rather than stopping at context attribution.

In that broader sense, ICX360 can be understood as both a specific toolkit and an emerging research program. As a toolkit, it provides an open-source, prompt-focused infrastructure for explaining LLM outputs across black-box and white-box access regimes (Wei et al., 14 Nov 2025). As a research program, it points toward a fuller 360-degree explainability stack in which context, audience, interaction, and formal evaluation are treated as first-class components of explanation rather than as external afterthoughts (Atakishiyev et al., 2020, Bello et al., 6 Jun 2025, Singh et al., 2022).