Research map / Large language models / Language, safety and interpretability
Faithfulness of LLM explanations
Evaluate whether natural-language explanations are faithful using counterfactuals and simulatability.
169 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Language, safety and interpretability, keyword cluster “prompt · uncertainty · shot”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
48 papers in 2023–24 and 55 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 6.9% to 3.0%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2016 · Ask, and Shall You Receive? Understanding Desire Fulfillment in Natural Language Text (AAAI 2016 · 10 citations)
- 2017 · Topic Aware Neural Response Generation (AAAI 2017 · 324 citations)
- 2018 · Emotional Chatting Machine: Emotional Conversation Generation with Internal and External Memory (AAAI 2018 · 762 citations)
- 2019 · Approximating Interactive Human Evaluation with Self-Play for Open-Domain Dialog Systems (NeurIPS 2019 · 76 citations)
- 2020 · Towards Scalable Multi-Domain Conversational Agents: The Schema-Guided Dialogue Dataset (AAAI 2020 · 462 citations)
- 2021 · UBAR: Towards Fully End-to-End Task-Oriented Dialog System with GPT-2 (AAAI 2021 · 123 citations)
- 2022 · CEM: Commonsense-Aware Empathetic Response Generation (AAAI 2022 · 145 citations)
- 2023 · Dynamic Prompt Learning via Policy Gradient for Semi-structured Mathematical Reasoning (ICLR 2023 · 41 citations)
- 2024 · Exploring Large Language Model for Graph Data Understanding in Online Job Recommendations (AAAI 2024 · 68 citations)
- 2025 · Understanding Emotional Body Expressions via Large Language Models (AAAI 2025 · 16 citations)
- 2026 · Do LLMs Feel? Teaching Emotion Recognition with Prompts, Retrieval, and Curriculum Learning (AAAI 2026 · 2 citations)
Earliest papers here
- Ask, and Shall You Receive? Understanding Desire Fulfillment in Natural Language Text (AAAI 2016 · 10 citations)
- Topic Aware Neural Response Generation (AAAI 2017 · 324 citations)
- Multiresolution Recurrent Neural Networks: An Application to Dialogue Response Generation (AAAI 2017 · 160 citations)
What has been tried
129 papers here report 1,401 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 7%, hurt 22%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add behavior meta-path prompts | RecrX point-wise Random · AUC | 0.71 to 0.88 | Helped | Exploring Large Language Model for Graph Data Understanding in Online Job Recommendations (AAAI 2024 · 68 citations) |
| Distill using clinical rationales rather than training without rationale augmentation | ADNI · accuracy | Helped | Large Language Models Are Clinical Reasoners: Reasoning-Aware Diagnosis Framework with Prompt-Generated Rationales (AAAI 2024 · 44 citations) | |
| add keypoint descriptions | COCO Keypoint val · AP | 70.8 -> 72.4 | Helped | LocLLM: Exploiting Generalizable Human Keypoint Localization via Large Language Model (CVPR 2024 · 19 citations) |
| combine spatial and temporal tokens rather than use either token type alone | Emilya, KDAE, and EGBM · accuracy | Emilya: 85.08 and 83.62 -> 86.36; KDAE: 62.99 and 61.21 -> 63.35; EGBM: 53.21 and 49.54 -> 55.96 | Helped | Understanding Emotional Body Expressions via Large Language Models (AAAI 2025 · 16 citations) |
| Add rating smoothing | Yelp · FMR | 8.66 -> 8.71 | Helped | Coherency Improved Explainable Recommendation via Large Language Model (AAAI 2025 · 11 citations) |
| Replace SentenceTransformer with RoBERTa as the backbone | CEBaB, original candidate set · L2 error | .65 -> .66 | No clear effect | Faithful Explanations of Black-box NLP Models Using LLM-generated Counterfactuals (ICLR 2024 · 7 citations) |
| add Lguide | BeerAdvocate Appearance · F1 | 81.4 -> 84.0 | Helped | Learning Robust Rationales for Model Explainability: A Guidance-Based Approach (AAAI 2024 · 6 citations) |
| Add the Situation node to the reasoning process | SSConv-test · NAvg | 1.372 -> 1.390 | Helped | SocialSim: Towards Socialized Simulation of Emotional Support Conversation (AAAI 2025 · 5 citations) |
| increase HSphNN feedback from a maximum of two rounds to a maximum of ten rounds | Syllogistic reasoning · correct decision and explanation | 84.76% to 85.16% | No clear effect | Neural Reasoning for Sure Through Constructing Explainable Models (AAAI 2025 · 3 citations) |
| remove LLM alignment loss (L^L) | Games · HR@5 | 0.0836 -> 0.0858 | Hurt | Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · 3 citations) |
| add task-specific agent instructions instead of using zero-shot CoT | AddSub · QEM | 73.2 -> 79.5 | Helped | Agent Instructs Large Language Models to be General Zero-Shot Reasoners (ICML 2024 · 3 citations) |
| Increase autoregressive generation masks from one to three | German-to-English translation (1-shot) · SacreBLEU | 10.0 -> 23.7 | Helped | BERTs are Generative In-Context Learners (NeurIPS 2024 · 3 citations) |
What reviewers will expect
Datasets and baselines named in the 27 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- GSM8K · 3 of 27 papers
Baselines
- GPT-4 · 4 of 27 papers
- gpt-3.5-turbo · 3 of 27 papers
- GPT-3.5 · 2 of 27 papers
- Llama-2-13B-chat · 2 of 27 papers
- gpt-4o · 2 of 27 papers
- Llama-3-8B-Instruct · 2 of 27 papers
- GPT-4o mini · 2 of 27 papers
- Claude · 2 of 27 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 103 papers; from OpenAlex and the papers' affiliations).
- Peking University · 6 papers
- University College London · 4 papers
- The Chinese University of Hong Kong · 4 papers
- Microsoft · 4 papers
- Google · 4 papers
- University of California, Berkeley · 4 papers
- Tsinghua University · 3 papers
- Tencent · 3 papers
- Amazon · 3 papers
- Carnegie Mellon University · 3 papers
Use it
Download BibTeX of 39 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- LLM unlearning and benchmarks (298 papers)
- LLM pruning and scaling data (285 papers)
- Controlled text generation (220 papers)
- Data and evaluation methodology (136 papers)
- In-context learning analysis (95 papers)
- Parallel and early-exit decoding (100 papers)
Elsewhere in the topic
- Uncertainty quantification for LLMs (136 papers)
- Pruning and circuits in LLMs (135 papers)
- Statistical frameworks for LLM evaluation (124 papers)
- Safety alignment of LLMs (123 papers)
- Linguistic encoding in LLM representations (123 papers)
- Reasoning mechanisms and failure modes (114 papers)
- Speech and audio language models (111 papers)
- Sign and cross-lingual translation (96 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).