| {"id": "PTL-0001", "slug": "prompt", "term": "Prompt", "aliases": ["input", "context"], "category": "Foundations", "definition": "A prompt is the complete input text, and optionally other media, that is given to a language model to condition its output, including instructions, context, examples, and the user's request.", "description": "In chat-based systems the prompt is assembled from several parts, such as a system prompt, the prior conversation turns, retrieved documents, and tool results. The model produces its output by predicting tokens conditioned on this entire input, so every part of the prompt can influence the response, not only the final question.", "example": null, "broader": [], "narrower": ["system-prompt", "prompt-template", "multimodal-prompting"], "related": ["context-window", "context-engineering"], "introduced": null, "sources": [{"title": "Language Models are Few-Shot Learners", "authors": "Brown et al.", "year": 2020, "url": "https://arxiv.org/abs/2005.14165"}, {"title": "The Prompt Report: A Systematic Survey of Prompt Engineering Techniques", "authors": "Schulhoff et al.", "year": 2024, "url": "https://arxiv.org/abs/2406.06608"}], "url": "https://protologue.com/t/prompt/", "citation": "Protologue. (2026). Prompt. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0001). https://protologue.com/t/prompt/"} | |
| {"id": "PTL-0002", "slug": "prompt-engineering", "term": "Prompt Engineering", "aliases": ["prompt design"], "category": "Foundations", "definition": "Prompt engineering is the practice of designing, testing, and iteratively refining prompts so that a language model reliably produces the desired output for a task.", "description": "It covers the wording of instructions, the choice and ordering of examples, output formatting constraints, and the decomposition of a task across multiple calls. As models and tooling matured, much of this work broadened into context engineering, which treats the whole information environment around a model call as the object being designed.", "example": null, "broader": [], "narrower": ["context-engineering"], "related": ["prompt", "automatic-prompt-engineer"], "introduced": null, "sources": [{"title": "The Prompt Report: A Systematic Survey of Prompt Engineering Techniques", "authors": "Schulhoff et al.", "year": 2024, "url": "https://arxiv.org/abs/2406.06608"}, {"title": "Prompt engineering overview", "authors": "Anthropic", "year": 2024, "url": "https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview"}], "url": "https://protologue.com/t/prompt-engineering/", "citation": "Protologue. (2026). Prompt Engineering. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0002). https://protologue.com/t/prompt-engineering/"} | |
| {"id": "PTL-0003", "slug": "system-prompt", "term": "System Prompt", "aliases": ["system message", "developer message", "system instruction"], "category": "Foundations", "definition": "A system prompt is a privileged instruction block, placed before the conversation, that sets a model's role, rules, tone, and constraints for every subsequent turn.", "description": "Chat APIs typically expose the system prompt as a separate message role. Models trained with an instruction hierarchy are taught to give system instructions precedence over conflicting user or tool content, which makes the system prompt the main lever for operator control and a frequent target of prompt-leaking attacks.", "example": null, "broader": ["prompt"], "narrower": [], "related": ["instruction-hierarchy", "role-prompting", "prompt-leaking"], "introduced": null, "sources": [{"title": "Training language models to follow instructions with human feedback", "authors": "Ouyang et al.", "year": 2022, "url": "https://arxiv.org/abs/2203.02155"}, {"title": "The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions", "authors": "Wallace et al.", "year": 2024, "url": "https://arxiv.org/abs/2404.13208"}], "url": "https://protologue.com/t/system-prompt/", "citation": "Protologue. (2026). System Prompt. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0003). https://protologue.com/t/system-prompt/"} | |
| {"id": "PTL-0004", "slug": "context-window", "term": "Context Window", "aliases": ["context length", "context size"], "category": "Foundations", "definition": "The context window is the maximum number of tokens a language model can attend to in a single call, covering both the prompt and the generated output.", "description": "Anything outside the context window is invisible to the model unless it is re-inserted. Larger windows allow whole documents or codebases to be included, but research shows models do not use all positions equally well, which motivates retrieval, summarization, and careful placement of key information.", "example": null, "broader": [], "narrower": [], "related": ["token", "lost-in-the-middle", "needle-in-a-haystack", "context-engineering"], "introduced": null, "sources": [{"title": "Attention Is All You Need", "authors": "Vaswani et al.", "year": 2017, "url": "https://arxiv.org/abs/1706.03762"}, {"title": "Lost in the Middle: How Language Models Use Long Contexts", "authors": "Liu et al.", "year": 2023, "url": "https://arxiv.org/abs/2307.03172"}], "url": "https://protologue.com/t/context-window/", "citation": "Protologue. (2026). Context Window. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0004). https://protologue.com/t/context-window/"} | |
| {"id": "PTL-0005", "slug": "token", "term": "Token", "aliases": ["subword", "BPE token", "tokenizer", "tokenization"], "category": "Foundations", "definition": "A token is the basic unit of text a language model reads and writes, typically a word, word fragment, or character sequence produced by a subword tokenizer such as byte-pair encoding.", "description": "Context limits, pricing, and generation speed are all measured in tokens. Because tokenization splits text unevenly, character-level tasks such as counting letters or reversing strings are harder for models than they appear, and the same content can cost different numbers of tokens in different languages.", "example": null, "broader": [], "narrower": [], "related": ["context-window"], "introduced": null, "sources": [{"title": "Neural Machine Translation of Rare Words with Subword Units", "authors": "Sennrich et al.", "year": 2015, "url": "https://arxiv.org/abs/1508.07909"}], "url": "https://protologue.com/t/token/", "citation": "Protologue. (2026). Token. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0005). https://protologue.com/t/token/"} | |
| {"id": "PTL-0006", "slug": "temperature", "term": "Temperature", "aliases": ["sampling temperature"], "category": "Foundations", "definition": "Temperature is a sampling parameter that rescales a model's output probabilities before a token is chosen; lower values make outputs more deterministic and higher values make them more varied.", "description": "At temperature zero the model approximately always picks its most likely token (greedy decoding). Techniques that rely on diverse samples, such as self-consistency and best-of-N, deliberately use a nonzero temperature, while extraction and classification tasks usually use a low one.", "example": null, "broader": [], "narrower": [], "related": ["top-p-sampling", "self-consistency", "best-of-n-sampling"], "introduced": null, "sources": [{"title": "The Curious Case of Neural Text Degeneration", "authors": "Holtzman et al.", "year": 2019, "url": "https://arxiv.org/abs/1904.09751"}], "url": "https://protologue.com/t/temperature/", "citation": "Protologue. (2026). Temperature. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0006). https://protologue.com/t/temperature/"} | |
| {"id": "PTL-0007", "slug": "top-p-sampling", "term": "Top-p Sampling", "aliases": ["nucleus sampling"], "category": "Foundations", "definition": "Top-p sampling, also called nucleus sampling, draws each next token only from the smallest set of candidates whose cumulative probability exceeds a threshold p.", "description": "It was proposed to avoid both the repetitive text produced by greedy or beam decoding and the incoherent text produced by sampling from the full distribution. It is usually combined with temperature.", "example": null, "broader": [], "narrower": [], "related": ["temperature"], "introduced": 2019, "sources": [{"title": "The Curious Case of Neural Text Degeneration", "authors": "Holtzman et al.", "year": 2019, "url": "https://arxiv.org/abs/1904.09751"}], "url": "https://protologue.com/t/top-p-sampling/", "citation": "Protologue. (2026). Top-p Sampling. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0007). https://protologue.com/t/top-p-sampling/"} | |
| {"id": "PTL-0008", "slug": "zero-shot-prompting", "term": "Zero-shot Prompting", "aliases": [], "category": "Foundations", "definition": "Zero-shot prompting asks a model to perform a task from an instruction alone, without any worked examples in the prompt.", "description": "Zero-shot performance improved dramatically with instruction tuning and preference training, which taught models to follow natural-language task descriptions. It is the default for most modern chat use, with examples added only when the format or judgment required is hard to describe.", "example": null, "broader": [], "narrower": [], "related": ["few-shot-prompting", "instruction-tuning", "zero-shot-chain-of-thought"], "introduced": null, "sources": [{"title": "Language Models are Few-Shot Learners", "authors": "Brown et al.", "year": 2020, "url": "https://arxiv.org/abs/2005.14165"}, {"title": "Finetuned Language Models Are Zero-Shot Learners", "authors": "Wei et al.", "year": 2021, "url": "https://arxiv.org/abs/2109.01652"}], "url": "https://protologue.com/t/zero-shot-prompting/", "citation": "Protologue. (2026). Zero-shot Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0008). https://protologue.com/t/zero-shot-prompting/"} | |
| {"id": "PTL-0009", "slug": "few-shot-prompting", "term": "Few-shot Prompting", "aliases": ["k-shot prompting", "in-context examples", "one-shot prompting"], "category": "Foundations", "definition": "Few-shot prompting includes a small number of input-output examples in the prompt so the model can infer the task and the desired format from demonstrations.", "description": "Popularized by the GPT-3 paper, it allows a model to perform new tasks without any weight updates. Results are sensitive to which examples are chosen, their order, and their label balance, which spawned a body of work on exemplar selection and calibration.", "example": "Classify sentiment.\nReview: \"Arrived broken.\" -> negative\nReview: \"Works perfectly.\" -> positive\nReview: \"Battery died in a day.\" ->\n", "broader": ["in-context-learning"], "narrower": ["exemplar-selection", "exemplar-ordering", "few-shot-calibration"], "related": ["zero-shot-prompting", "many-shot-in-context-learning"], "introduced": 2020, "sources": [{"title": "Language Models are Few-Shot Learners", "authors": "Brown et al.", "year": 2020, "url": "https://arxiv.org/abs/2005.14165"}], "url": "https://protologue.com/t/few-shot-prompting/", "citation": "Protologue. (2026). Few-shot Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0009). https://protologue.com/t/few-shot-prompting/"} | |
| {"id": "PTL-0010", "slug": "instruction-tuning", "term": "Instruction Tuning", "aliases": ["instruction fine-tuning", "supervised fine-tuning", "SFT"], "category": "Foundations", "definition": "Instruction tuning is fine-tuning a pretrained language model on many tasks phrased as natural-language instructions, so that it follows unseen instructions zero-shot.", "description": "The FLAN work showed that tuning on dozens of instruction-formatted datasets substantially improved zero-shot performance on held-out tasks. Combined with reinforcement learning from human feedback, it produced the instruction-following chat models that most prompting techniques now target.", "example": null, "broader": [], "narrower": [], "related": ["zero-shot-prompting", "rlhf"], "introduced": 2021, "sources": [{"title": "Finetuned Language Models Are Zero-Shot Learners", "authors": "Wei et al.", "year": 2021, "url": "https://arxiv.org/abs/2109.01652"}, {"title": "Training language models to follow instructions with human feedback", "authors": "Ouyang et al.", "year": 2022, "url": "https://arxiv.org/abs/2203.02155"}], "url": "https://protologue.com/t/instruction-tuning/", "citation": "Protologue. (2026). Instruction Tuning. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0010). https://protologue.com/t/instruction-tuning/"} | |
| {"id": "PTL-0011", "slug": "rlhf", "term": "Reinforcement Learning from Human Feedback", "aliases": ["RLHF", "preference tuning"], "category": "Foundations", "definition": "Reinforcement learning from human feedback (RLHF) trains a language model to produce outputs people prefer, by learning a reward model from human comparisons and optimizing the model against it.", "description": "InstructGPT showed that RLHF made a much smaller model preferred over a far larger base model. RLHF shapes how models respond to prompts, including their helpfulness and refusals, and is linked to failure modes such as sycophancy. Direct preference optimization is a widely used simpler alternative.", "example": null, "broader": [], "narrower": ["direct-preference-optimization"], "related": ["instruction-tuning", "constitutional-ai", "sycophancy"], "introduced": 2022, "sources": [{"title": "Training language models to follow instructions with human feedback", "authors": "Ouyang et al.", "year": 2022, "url": "https://arxiv.org/abs/2203.02155"}], "url": "https://protologue.com/t/rlhf/", "citation": "Protologue. (2026). Reinforcement Learning from Human Feedback. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0011). https://protologue.com/t/rlhf/"} | |
| {"id": "PTL-0012", "slug": "direct-preference-optimization", "term": "Direct Preference Optimization", "aliases": ["DPO"], "category": "Foundations", "definition": "Direct preference optimization (DPO) aligns a language model to human preferences by training directly on preferred-versus-rejected response pairs, without fitting a separate reward model or running reinforcement learning.", "description": "DPO reframes the RLHF objective as a simple classification-style loss over preference pairs. It became a common alternative to RLHF because it is stable and cheap to run.", "example": null, "broader": ["rlhf"], "narrower": [], "related": [], "introduced": 2023, "sources": [{"title": "Direct Preference Optimization: Your Language Model is Secretly a Reward Model", "authors": "Rafailov et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.18290"}], "url": "https://protologue.com/t/direct-preference-optimization/", "citation": "Protologue. (2026). Direct Preference Optimization. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0012). https://protologue.com/t/direct-preference-optimization/"} | |
| {"id": "PTL-0013", "slug": "prompt-template", "term": "Prompt Template", "aliases": ["prompt scaffold"], "category": "Foundations", "definition": "A prompt template is a reusable prompt with placeholder variables that are filled in at run time, separating the fixed instructions from the per-request data.", "description": "Templates make prompts testable and versionable, and they are the unit optimized by frameworks such as DSPy. Keeping untrusted data in clearly delimited slots also supports defenses against prompt injection.", "example": null, "broader": ["prompt"], "narrower": [], "related": ["delimiters", "dspy", "spotlighting"], "introduced": null, "sources": [{"title": "The Prompt Report: A Systematic Survey of Prompt Engineering Techniques", "authors": "Schulhoff et al.", "year": 2024, "url": "https://arxiv.org/abs/2406.06608"}], "url": "https://protologue.com/t/prompt-template/", "citation": "Protologue. (2026). Prompt Template. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0013). https://protologue.com/t/prompt-template/"} | |
| {"id": "PTL-0014", "slug": "delimiters", "term": "Delimiters", "aliases": ["XML tags", "separators"], "category": "Foundations", "definition": "Delimiters are explicit markers, such as XML-style tags, triple quotes, or headings, that separate the parts of a prompt so the model can tell instructions, data, and examples apart.", "description": "Vendor guidance recommends delimiters to reduce ambiguity, make outputs easier to parse, and lower the chance that content inside a document is mistaken for an instruction. Tags can also be requested in the output for reliable extraction.", "example": null, "broader": [], "narrower": [], "related": ["prompt-template", "structured-outputs", "spotlighting"], "introduced": null, "sources": [{"title": "Prompting best practices", "authors": "Anthropic", "year": 2026, "url": "https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices"}, {"title": "Prompt engineering", "authors": "OpenAI", "year": 2024, "url": "https://developers.openai.com/api/docs/guides/prompt-engineering"}], "url": "https://protologue.com/t/delimiters/", "citation": "Protologue. (2026). Delimiters. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0014). https://protologue.com/t/delimiters/"} | |
| {"id": "PTL-0015", "slug": "prefill", "term": "Prefill", "aliases": ["response prefilling", "output priming", "assistant prefill"], "category": "Foundations", "definition": "Prefill is the technique of writing the first part of the model's response yourself, so that the model continues from that text and is steered into a specific format or direction.", "description": "Starting the assistant turn with an opening brace pushes the model toward JSON output, and starting with a heading or a character's name can enforce structure or persona. Support is model-specific and declining. Anthropic's documentation states that, starting with Claude 4.6 models, requests that prefill the final assistant turn return an error, while earlier models still accept them and assistant messages elsewhere in the conversation are unaffected. It gives migration guidance for common uses, such as controlling output format and eliminating preambles, mostly through explicit instructions.", "example": null, "broader": [], "narrower": [], "related": ["structured-outputs", "delimiters"], "introduced": null, "sources": [{"title": "Prompting best practices", "authors": "Anthropic", "year": 2026, "url": "https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices"}], "url": "https://protologue.com/t/prefill/", "citation": "Protologue. (2026). Prefill. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0015). https://protologue.com/t/prefill/"} | |
| {"id": "PTL-0016", "slug": "role-prompting", "term": "Role Prompting", "aliases": ["persona prompting", "role-play prompting"], "category": "Foundations", "definition": "Role prompting assigns the model a persona or professional role, such as \"You are an experienced tax accountant,\" to shape its tone, vocabulary, and focus.", "description": "Role prompts reliably change style and framing. Evidence that they improve factual accuracy is mixed, with some studies finding gains in reasoning tasks and a large evaluation finding that adding personas to system prompts did not consistently improve performance.", "example": null, "broader": [], "narrower": [], "related": ["system-prompt"], "introduced": null, "sources": [{"title": "Better Zero-Shot Reasoning with Role-Play Prompting", "authors": "Kong et al.", "year": 2023, "url": "https://arxiv.org/abs/2308.07702"}, {"title": "When \"A Helpful Assistant\" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models", "authors": "Zheng et al.", "year": 2023, "url": "https://arxiv.org/abs/2311.10054"}], "url": "https://protologue.com/t/role-prompting/", "citation": "Protologue. (2026). Role Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0016). https://protologue.com/t/role-prompting/"} | |
| {"id": "PTL-0017", "slug": "in-context-learning", "term": "In-Context Learning", "aliases": ["ICL"], "category": "Exemplars & In-Context Learning", "definition": "In-context learning (ICL) is a language model's ability to perform a task by conditioning on instructions or demonstrations in its prompt, without any update to its weights.", "description": "Identified as an emergent capability of large pretrained models in the GPT-3 paper, ICL underlies few-shot prompting. Studies show that models often rely more on the format and label space of demonstrations than on whether the demonstration labels are correct.", "example": null, "broader": [], "narrower": ["few-shot-prompting", "demonstration-label-sensitivity", "many-shot-in-context-learning"], "related": [], "introduced": 2020, "sources": [{"title": "Language Models are Few-Shot Learners", "authors": "Brown et al.", "year": 2020, "url": "https://arxiv.org/abs/2005.14165"}], "url": "https://protologue.com/t/in-context-learning/", "citation": "Protologue. (2026). In-Context Learning. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0017). https://protologue.com/t/in-context-learning/"} | |
| {"id": "PTL-0018", "slug": "demonstration-label-sensitivity", "term": "Demonstration Label Sensitivity", "aliases": ["role of demonstrations"], "category": "Exemplars & In-Context Learning", "definition": "Demonstration label sensitivity refers to how much a model's few-shot performance depends on whether the example labels are correct; research found that randomly replacing labels often hurts performance only slightly.", "description": "Min et al. showed that demonstrations mainly supply the label space, the input distribution, and the format of the task. This suggests few-shot examples work largely by specifying what the task looks like rather than by teaching the input-label mapping.", "example": null, "broader": ["in-context-learning"], "narrower": [], "related": ["few-shot-prompting", "exemplar-selection"], "introduced": null, "sources": [{"title": "Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?", "authors": "Min et al.", "year": 2022, "url": "https://arxiv.org/abs/2202.12837"}], "url": "https://protologue.com/t/demonstration-label-sensitivity/", "citation": "Protologue. (2026). Demonstration Label Sensitivity. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0018). https://protologue.com/t/demonstration-label-sensitivity/"} | |
| {"id": "PTL-0019", "slug": "exemplar-selection", "term": "Exemplar Selection", "aliases": ["demonstration selection", "dynamic few-shot", "kNN prompting"], "category": "Exemplars & In-Context Learning", "definition": "Exemplar selection is the choice of which demonstrations to include in a few-shot prompt, commonly by retrieving the examples most semantically similar to the current input.", "description": "Liu et al. showed that retrieving nearest-neighbor examples by embedding similarity outperformed random selection. Selection can also target diversity, difficulty, or the model's uncertainty, as in active prompting.", "example": null, "broader": ["few-shot-prompting"], "narrower": ["active-prompting"], "related": ["exemplar-ordering", "retrieval-augmented-generation"], "introduced": null, "sources": [{"title": "What Makes Good In-Context Examples for GPT-3?", "authors": "Liu et al.", "year": 2021, "url": "https://arxiv.org/abs/2101.06804"}], "url": "https://protologue.com/t/exemplar-selection/", "citation": "Protologue. (2026). Exemplar Selection. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0019). https://protologue.com/t/exemplar-selection/"} | |
| {"id": "PTL-0020", "slug": "exemplar-ordering", "term": "Exemplar Ordering", "aliases": ["demonstration order", "order sensitivity"], "category": "Exemplars & In-Context Learning", "definition": "Exemplar ordering is the arrangement of demonstrations within a few-shot prompt, which can swing accuracy from near state-of-the-art to near chance for the same set of examples.", "description": "Lu et al. documented this order sensitivity and proposed selecting performant orderings using a probe set generated by the model itself. The effect is a key reason few-shot results should be reported across several permutations.", "example": null, "broader": ["few-shot-prompting"], "narrower": [], "related": ["exemplar-selection", "few-shot-calibration", "prompt-sensitivity"], "introduced": null, "sources": [{"title": "Fantastically Ordered Prompts and Where to Find Them: Overcoming Few-Shot Prompt Order Sensitivity", "authors": "Lu et al.", "year": 2021, "url": "https://arxiv.org/abs/2104.08786"}], "url": "https://protologue.com/t/exemplar-ordering/", "citation": "Protologue. (2026). Exemplar Ordering. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0020). https://protologue.com/t/exemplar-ordering/"} | |
| {"id": "PTL-0021", "slug": "few-shot-calibration", "term": "Few-shot Calibration", "aliases": ["contextual calibration", "calibrate before use"], "category": "Exemplars & In-Context Learning", "definition": "Few-shot calibration corrects a model's systematic biases toward particular answers, such as the most frequent or most recent label in the examples, by adjusting output probabilities measured on a content-free input.", "description": "Zhao et al. identified majority-label bias, recency bias, and common-token bias in few-shot prompting, and proposed contextual calibration, which estimates the bias from an input such as \"N/A\" and rescales predictions to neutralize it.", "example": null, "broader": ["few-shot-prompting"], "narrower": [], "related": ["exemplar-ordering", "prompt-sensitivity"], "introduced": null, "sources": [{"title": "Calibrate Before Use: Improving Few-Shot Performance of Language Models", "authors": "Zhao et al.", "year": 2021, "url": "https://arxiv.org/abs/2102.09690"}], "url": "https://protologue.com/t/few-shot-calibration/", "citation": "Protologue. (2026). Few-shot Calibration. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0021). https://protologue.com/t/few-shot-calibration/"} | |
| {"id": "PTL-0022", "slug": "many-shot-in-context-learning", "term": "Many-shot In-Context Learning", "aliases": ["many-shot ICL", "long-context ICL"], "category": "Exemplars & In-Context Learning", "definition": "Many-shot in-context learning places hundreds or thousands of demonstrations in a long-context prompt, often yielding large gains over few-shot prompting.", "description": "Agarwal et al. found consistent improvements as the number of shots grew into the hundreds, and introduced variants that use model-generated rationales or unlabeled problems. The same scaling behavior underlies the many-shot jailbreaking attack.", "example": null, "broader": ["in-context-learning"], "narrower": [], "related": ["few-shot-prompting", "many-shot-jailbreaking", "context-window"], "introduced": 2024, "sources": [{"title": "Many-Shot In-Context Learning", "authors": "Agarwal et al.", "year": 2024, "url": "https://arxiv.org/abs/2404.11018"}], "url": "https://protologue.com/t/many-shot-in-context-learning/", "citation": "Protologue. (2026). Many-shot In-Context Learning. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0022). https://protologue.com/t/many-shot-in-context-learning/"} | |
| {"id": "PTL-0023", "slug": "active-prompting", "term": "Active Prompting", "aliases": ["Active-Prompt"], "category": "Exemplars & In-Context Learning", "definition": "Active prompting selects which questions to annotate with chain-of-thought exemplars by choosing those on which the model is most uncertain, measured by disagreement across sampled answers.", "description": "Borrowing from active learning, it focuses human annotation effort on the examples most informative for the model, rather than on a fixed or random set.", "example": null, "broader": ["exemplar-selection"], "narrower": [], "related": ["chain-of-thought", "self-consistency"], "introduced": 2023, "sources": [{"title": "Active Prompting with Chain-of-Thought for Large Language Models", "authors": "Diao et al.", "year": 2023, "url": "https://arxiv.org/abs/2302.12246"}], "url": "https://protologue.com/t/active-prompting/", "citation": "Protologue. (2026). Active Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0023). https://protologue.com/t/active-prompting/"} | |
| {"id": "PTL-0024", "slug": "chain-of-thought", "term": "Chain-of-Thought Prompting", "aliases": ["CoT", "step-by-step reasoning"], "category": "Reasoning Elicitation", "definition": "Chain-of-thought (CoT) prompting elicits a sequence of intermediate reasoning steps from a language model before its final answer, which improves performance on multi-step arithmetic, commonsense, and symbolic reasoning tasks.", "description": "Wei et al. introduced CoT by including worked examples with step-by-step reasoning in a few-shot prompt, and found the benefit emerged mainly in large models. It is the root of a large family of techniques, including zero-shot CoT, self-consistency, and tree of thoughts, and of the extended reasoning trained into modern reasoning models.", "example": "Q: Roger has 5 balls. He buys 2 cans of 3 balls each. How many balls does he have?\nA: He starts with 5. Two cans of 3 is 6. 5 + 6 = 11. The answer is 11.\n", "broader": [], "narrower": ["zero-shot-chain-of-thought", "auto-cot", "self-consistency", "tree-of-thoughts", "contrastive-chain-of-thought", "thread-of-thought", "complexity-based-prompting", "chain-of-draft"], "related": ["scratchpad", "reasoning-model", "unfaithful-chain-of-thought"], "introduced": 2022, "sources": [{"title": "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models", "authors": "Wei et al.", "year": 2022, "url": "https://arxiv.org/abs/2201.11903"}], "url": "https://protologue.com/t/chain-of-thought/", "citation": "Protologue. (2026). Chain-of-Thought Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0024). https://protologue.com/t/chain-of-thought/"} | |
| {"id": "PTL-0025", "slug": "zero-shot-chain-of-thought", "term": "Zero-shot Chain-of-Thought", "aliases": ["zero-shot CoT", "let's think step by step"], "category": "Reasoning Elicitation", "definition": "Zero-shot chain-of-thought prompting triggers step-by-step reasoning without examples by appending a cue such as \"Let's think step by step\" to the question.", "description": "Kojima et al. showed this single phrase produced large accuracy gains on arithmetic and logic benchmarks. The method uses a second prompt to extract the final answer from the generated reasoning.", "example": null, "broader": ["chain-of-thought"], "narrower": ["plan-and-solve-prompting"], "related": ["zero-shot-prompting"], "introduced": 2022, "sources": [{"title": "Large Language Models are Zero-Shot Reasoners", "authors": "Kojima et al.", "year": 2022, "url": "https://arxiv.org/abs/2205.11916"}], "url": "https://protologue.com/t/zero-shot-chain-of-thought/", "citation": "Protologue. (2026). Zero-shot Chain-of-Thought. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0025). https://protologue.com/t/zero-shot-chain-of-thought/"} | |
| {"id": "PTL-0026", "slug": "scratchpad", "term": "Scratchpad", "aliases": ["intermediate computation"], "category": "Reasoning Elicitation", "definition": "A scratchpad is a region of model output reserved for intermediate computation, which the model writes before its final answer so that multi-step calculations can be carried out explicitly.", "description": "Nye et al. trained models to emit intermediate steps for tasks like long addition and program execution, an important precursor to chain-of-thought prompting. In practice the term also refers to private reasoning sections that are hidden from end users.", "example": null, "broader": [], "narrower": [], "related": ["chain-of-thought", "extended-thinking"], "introduced": 2021, "sources": [{"title": "Show Your Work: Scratchpads for Intermediate Computation with Language Models", "authors": "Nye et al.", "year": 2021, "url": "https://arxiv.org/abs/2112.00114"}], "url": "https://protologue.com/t/scratchpad/", "citation": "Protologue. (2026). Scratchpad. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0026). https://protologue.com/t/scratchpad/"} | |
| {"id": "PTL-0027", "slug": "auto-cot", "term": "Automatic Chain-of-Thought", "aliases": ["Auto-CoT"], "category": "Reasoning Elicitation", "definition": "Automatic chain-of-thought (Auto-CoT) builds chain-of-thought demonstrations without manual writing, by clustering questions for diversity and generating a reasoning chain for a representative of each cluster with zero-shot CoT.", "description": "Diversity across clusters limits the damage from mistakes in any single generated chain, and the approach matched manually written CoT exemplars on several benchmarks.", "example": null, "broader": ["chain-of-thought"], "narrower": [], "related": ["zero-shot-chain-of-thought", "exemplar-selection"], "introduced": 2022, "sources": [{"title": "Automatic Chain of Thought Prompting in Large Language Models", "authors": "Zhang et al.", "year": 2022, "url": "https://arxiv.org/abs/2210.03493"}], "url": "https://protologue.com/t/auto-cot/", "citation": "Protologue. (2026). Automatic Chain-of-Thought. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0027). https://protologue.com/t/auto-cot/"} | |
| {"id": "PTL-0028", "slug": "self-consistency", "term": "Self-Consistency", "aliases": ["majority voting", "CoT-SC"], "category": "Reasoning Elicitation", "definition": "Self-consistency samples multiple chain-of-thought reasoning paths for the same question and returns the answer that appears most often, rather than relying on a single greedy decode.", "description": "Wang et al. found this majority vote substantially improved chain-of-thought accuracy on arithmetic and commonsense benchmarks. It trades extra inference cost for reliability and is an early form of test-time compute scaling.", "example": null, "broader": ["chain-of-thought"], "narrower": ["universal-self-consistency"], "related": ["best-of-n-sampling", "test-time-compute-scaling", "temperature"], "introduced": 2022, "sources": [{"title": "Self-Consistency Improves Chain of Thought Reasoning in Language Models", "authors": "Wang et al.", "year": 2022, "url": "https://arxiv.org/abs/2203.11171"}], "url": "https://protologue.com/t/self-consistency/", "citation": "Protologue. (2026). Self-Consistency. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0028). https://protologue.com/t/self-consistency/"} | |
| {"id": "PTL-0029", "slug": "universal-self-consistency", "term": "Universal Self-Consistency", "aliases": [], "category": "Reasoning Elicitation", "definition": "Universal self-consistency extends self-consistency to free-form outputs by asking the model itself to select the most consistent response among several samples, instead of counting exact-match answers.", "description": "This makes sampling-and-selecting usable for tasks such as summarization and open-ended question answering, where answers rarely match word for word.", "example": null, "broader": ["self-consistency"], "narrower": [], "related": ["llm-as-a-judge"], "introduced": 2023, "sources": [{"title": "Universal Self-Consistency for Large Language Model Generation", "authors": "Chen et al.", "year": 2023, "url": "https://arxiv.org/abs/2311.17311"}], "url": "https://protologue.com/t/universal-self-consistency/", "citation": "Protologue. (2026). Universal Self-Consistency. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0029). https://protologue.com/t/universal-self-consistency/"} | |
| {"id": "PTL-0030", "slug": "least-to-most-prompting", "term": "Least-to-Most Prompting", "aliases": ["problem decomposition"], "category": "Reasoning Elicitation", "definition": "Least-to-most prompting first asks the model to break a complex problem into simpler subproblems, then solves them in order, feeding each answer into the next.", "description": "Zhou et al. showed it generalizes to problems harder than those in the examples, a weakness of standard chain-of-thought, with strong results on compositional generalization benchmarks.", "example": null, "broader": [], "narrower": [], "related": ["chain-of-thought", "plan-and-solve-prompting", "self-ask", "prompt-chaining"], "introduced": 2022, "sources": [{"title": "Least-to-Most Prompting Enables Complex Reasoning in Large Language Models", "authors": "Zhou et al.", "year": 2022, "url": "https://arxiv.org/abs/2205.10625"}], "url": "https://protologue.com/t/least-to-most-prompting/", "citation": "Protologue. (2026). Least-to-Most Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0030). https://protologue.com/t/least-to-most-prompting/"} | |
| {"id": "PTL-0031", "slug": "plan-and-solve-prompting", "term": "Plan-and-Solve Prompting", "aliases": ["PS prompting", "PS+"], "category": "Reasoning Elicitation", "definition": "Plan-and-solve prompting is a zero-shot method that instructs the model to first devise a plan dividing the task into subtasks and then carry out the plan step by step.", "description": "It targets missing-step and calculation errors seen in zero-shot chain-of-thought, with an extended version that asks the model to extract relevant variables and compute carefully.", "example": null, "broader": ["zero-shot-chain-of-thought"], "narrower": [], "related": ["least-to-most-prompting"], "introduced": 2023, "sources": [{"title": "Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models", "authors": "Wang et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.04091"}], "url": "https://protologue.com/t/plan-and-solve-prompting/", "citation": "Protologue. (2026). Plan-and-Solve Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0031). https://protologue.com/t/plan-and-solve-prompting/"} | |
| {"id": "PTL-0032", "slug": "step-back-prompting", "term": "Step-Back Prompting", "aliases": ["abstraction prompting"], "category": "Reasoning Elicitation", "definition": "Step-back prompting has the model first answer a more general, abstract question about the underlying principle, then use that answer to reason about the original specific question.", "description": "For a physics question, the step-back question might ask which law applies. Zheng et al. reported gains on science, multi-hop, and knowledge-intensive question answering.", "example": null, "broader": [], "narrower": [], "related": ["chain-of-thought", "generated-knowledge-prompting"], "introduced": 2023, "sources": [{"title": "Take a Step Back: Evoking Reasoning via Abstraction in Large Language Models", "authors": "Zheng et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.06117"}], "url": "https://protologue.com/t/step-back-prompting/", "citation": "Protologue. (2026). Step-Back Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0032). https://protologue.com/t/step-back-prompting/"} | |
| {"id": "PTL-0033", "slug": "tree-of-thoughts", "term": "Tree of Thoughts", "aliases": ["ToT"], "category": "Reasoning Elicitation", "definition": "Tree of Thoughts (ToT) lets a model explore multiple reasoning branches as a search tree, evaluating partial solutions and backtracking, rather than committing to a single left-to-right chain of thought.", "description": "Yao et al. combined model-generated candidate thoughts, model self-evaluation of states, and breadth- or depth-first search, producing large gains on tasks such as the Game of 24 that require planning or lookahead.", "example": null, "broader": ["chain-of-thought"], "narrower": ["graph-of-thoughts"], "related": ["self-consistency", "test-time-compute-scaling"], "introduced": 2023, "sources": [{"title": "Tree of Thoughts: Deliberate Problem Solving with Large Language Models", "authors": "Yao et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.10601"}], "url": "https://protologue.com/t/tree-of-thoughts/", "citation": "Protologue. (2026). Tree of Thoughts. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0033). https://protologue.com/t/tree-of-thoughts/"} | |
| {"id": "PTL-0034", "slug": "graph-of-thoughts", "term": "Graph of Thoughts", "aliases": [], "category": "Reasoning Elicitation", "definition": "Graph of Thoughts (GoT) models a language model's reasoning as an arbitrary graph, in which thoughts can be combined, refined, and looped back on, generalizing chain and tree structures.", "description": "Aggregation of several partial solutions into one, such as merging sorted sublists, is the key operation GoT adds over tree-shaped search.", "example": null, "broader": ["tree-of-thoughts"], "narrower": [], "related": ["chain-of-thought"], "introduced": 2023, "sources": [{"title": "Graph of Thoughts: Solving Elaborate Problems with Large Language Models", "authors": "Besta et al.", "year": 2023, "url": "https://arxiv.org/abs/2308.09687"}], "url": "https://protologue.com/t/graph-of-thoughts/", "citation": "Protologue. (2026). Graph of Thoughts. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0034). https://protologue.com/t/graph-of-thoughts/"} | |
| {"id": "PTL-0035", "slug": "skeleton-of-thought", "term": "Skeleton-of-Thought", "aliases": [], "category": "Reasoning Elicitation", "definition": "Skeleton-of-thought first asks the model for a brief outline of its answer, then expands each outline point in parallel, reducing end-to-end generation latency.", "description": "Because the points are expanded independently, it suits list-like answers better than tightly sequential reasoning.", "example": null, "broader": [], "narrower": [], "related": ["parallelization", "prompt-chaining"], "introduced": 2023, "sources": [{"title": "Skeleton-of-Thought: Prompting LLMs for Efficient Parallel Generation", "authors": "Ning et al.", "year": 2023, "url": "https://arxiv.org/abs/2307.15337"}], "url": "https://protologue.com/t/skeleton-of-thought/", "citation": "Protologue. (2026). Skeleton-of-Thought. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0035). https://protologue.com/t/skeleton-of-thought/"} | |
| {"id": "PTL-0036", "slug": "analogical-prompting", "term": "Analogical Prompting", "aliases": [], "category": "Reasoning Elicitation", "definition": "Analogical prompting asks the model to recall or generate relevant example problems and their solutions on its own before solving the target problem, removing the need for hand-written exemplars.", "description": "Inspired by how people draw on analogous past experience, the self-generated examples are tailored to each problem rather than fixed for the whole task.", "example": null, "broader": [], "narrower": [], "related": ["few-shot-prompting", "generated-knowledge-prompting"], "introduced": 2023, "sources": [{"title": "Large Language Models as Analogical Reasoners", "authors": "Yasunaga et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.01714"}], "url": "https://protologue.com/t/analogical-prompting/", "citation": "Protologue. (2026). Analogical Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0036). https://protologue.com/t/analogical-prompting/"} | |
| {"id": "PTL-0037", "slug": "self-discover", "term": "Self-Discover", "aliases": [], "category": "Reasoning Elicitation", "definition": "Self-Discover has the model compose a task-specific reasoning structure by selecting, adapting, and combining general reasoning modules, such as critical thinking or step-by-step analysis, before solving instances of the task.", "description": "The structure is discovered once per task and then reused, so it costs far fewer inference calls than sampling-heavy methods like self-consistency.", "example": null, "broader": [], "narrower": [], "related": ["chain-of-thought", "meta-prompting"], "introduced": 2024, "sources": [{"title": "Self-Discover: Large Language Models Self-Compose Reasoning Structures", "authors": "Zhou et al.", "year": 2024, "url": "https://arxiv.org/abs/2402.03620"}], "url": "https://protologue.com/t/self-discover/", "citation": "Protologue. (2026). Self-Discover. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0037). https://protologue.com/t/self-discover/"} | |
| {"id": "PTL-0038", "slug": "contrastive-chain-of-thought", "term": "Contrastive Chain-of-Thought", "aliases": [], "category": "Reasoning Elicitation", "definition": "Contrastive chain-of-thought adds both valid and deliberately invalid reasoning demonstrations to a prompt, so the model learns which mistakes to avoid as well as what correct reasoning looks like.", "description": "The invalid demonstrations can be generated automatically by perturbing correct reasoning chains.", "example": null, "broader": ["chain-of-thought"], "narrower": [], "related": [], "introduced": 2023, "sources": [{"title": "Contrastive Chain-of-Thought Prompting", "authors": "Chia et al.", "year": 2023, "url": "https://arxiv.org/abs/2311.09277"}], "url": "https://protologue.com/t/contrastive-chain-of-thought/", "citation": "Protologue. (2026). Contrastive Chain-of-Thought. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0038). https://protologue.com/t/contrastive-chain-of-thought/"} | |
| {"id": "PTL-0039", "slug": "thread-of-thought", "term": "Thread of Thought", "aliases": [], "category": "Reasoning Elicitation", "definition": "Thread of Thought is a prompting strategy for long, chaotic contexts that asks the model to walk through the context in manageable parts, summarizing and analyzing each before answering.", "description": "A typical trigger asks the model to go through the context step by step, summarizing and analyzing as it goes.", "example": null, "broader": ["chain-of-thought"], "narrower": [], "related": ["lost-in-the-middle", "system-2-attention"], "introduced": 2023, "sources": [{"title": "Thread of Thought Unraveling Chaotic Contexts", "authors": "Zhou et al.", "year": 2023, "url": "https://arxiv.org/abs/2311.08734"}], "url": "https://protologue.com/t/thread-of-thought/", "citation": "Protologue. (2026). Thread of Thought. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0039). https://protologue.com/t/thread-of-thought/"} | |
| {"id": "PTL-0040", "slug": "system-2-attention", "term": "System 2 Attention", "aliases": ["S2A"], "category": "Reasoning Elicitation", "definition": "System 2 Attention (S2A) first prompts the model to rewrite the input so that it keeps only the relevant, unbiased content, then answers using the rewritten context.", "description": "Weston and Sukhbaatar showed this reduces the influence of irrelevant or opinionated text in the prompt, improving factuality and reducing sycophancy toward opinions stated in the question.", "example": null, "broader": [], "narrower": [], "related": ["thread-of-thought", "sycophancy", "rephrase-and-respond"], "introduced": 2023, "sources": [{"title": "System 2 Attention (is something you might need too)", "authors": "Weston & Sukhbaatar", "year": 2023, "url": "https://arxiv.org/abs/2311.11829"}], "url": "https://protologue.com/t/system-2-attention/", "citation": "Protologue. (2026). System 2 Attention. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0040). https://protologue.com/t/system-2-attention/"} | |
| {"id": "PTL-0041", "slug": "rephrase-and-respond", "term": "Rephrase and Respond", "aliases": ["RaR"], "category": "Reasoning Elicitation", "definition": "Rephrase and Respond (RaR) asks the model to rephrase and expand the user's question in its own words before answering, reducing misunderstandings caused by ambiguous phrasing.", "description": "The rephrasing can be done in the same response or by one model for another, as a two-step variant.", "example": null, "broader": [], "narrower": [], "related": ["system-2-attention"], "introduced": 2023, "sources": [{"title": "Rephrase and Respond: Let Large Language Models Ask Better Questions for Themselves", "authors": "Deng et al.", "year": 2023, "url": "https://arxiv.org/abs/2311.04205"}], "url": "https://protologue.com/t/rephrase-and-respond/", "citation": "Protologue. (2026). Rephrase and Respond. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0041). https://protologue.com/t/rephrase-and-respond/"} | |
| {"id": "PTL-0042", "slug": "self-ask", "term": "Self-Ask", "aliases": [], "category": "Reasoning Elicitation", "definition": "Self-ask prompting has the model explicitly pose and answer follow-up sub-questions before answering a multi-hop question, a format that can plug a search engine in to answer each sub-question.", "description": "Press et al. used it to study the compositionality gap, the tendency of models to answer each sub-question correctly yet fail the composed question.", "example": null, "broader": [], "narrower": [], "related": ["least-to-most-prompting", "react"], "introduced": 2022, "sources": [{"title": "Measuring and Narrowing the Compositionality Gap in Language Models", "authors": "Press et al.", "year": 2022, "url": "https://arxiv.org/abs/2210.03350"}], "url": "https://protologue.com/t/self-ask/", "citation": "Protologue. (2026). Self-Ask. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0042). https://protologue.com/t/self-ask/"} | |
| {"id": "PTL-0043", "slug": "maieutic-prompting", "term": "Maieutic Prompting", "aliases": [], "category": "Reasoning Elicitation", "definition": "Maieutic prompting generates a tree of recursive explanations for and against an answer, then infers the most logically consistent answer from the relations among them.", "description": "Named after the Socratic method, it is designed to tolerate individual explanations that are wrong or inconsistent.", "example": null, "broader": [], "narrower": [], "related": ["self-consistency", "chain-of-verification"], "introduced": 2022, "sources": [{"title": "Maieutic Prompting: Logically Consistent Reasoning with Recursive Explanations", "authors": "Jung et al.", "year": 2022, "url": "https://arxiv.org/abs/2205.11822"}], "url": "https://protologue.com/t/maieutic-prompting/", "citation": "Protologue. (2026). Maieutic Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0043). https://protologue.com/t/maieutic-prompting/"} | |
| {"id": "PTL-0044", "slug": "generated-knowledge-prompting", "term": "Generated Knowledge Prompting", "aliases": [], "category": "Reasoning Elicitation", "definition": "Generated knowledge prompting first asks the model to produce relevant facts about a question, then supplies those generated facts as context when answering it.", "description": "It acts like retrieval from the model's own parametric knowledge, and improved commonsense reasoning benchmarks in the original study.", "example": null, "broader": [], "narrower": [], "related": ["step-back-prompting", "retrieval-augmented-generation"], "introduced": 2021, "sources": [{"title": "Generated Knowledge Prompting for Commonsense Reasoning", "authors": "Liu et al.", "year": 2021, "url": "https://arxiv.org/abs/2110.08387"}], "url": "https://protologue.com/t/generated-knowledge-prompting/", "citation": "Protologue. (2026). Generated Knowledge Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0044). https://protologue.com/t/generated-knowledge-prompting/"} | |
| {"id": "PTL-0045", "slug": "program-aided-language-models", "term": "Program-Aided Language Models", "aliases": ["PAL"], "category": "Reasoning Elicitation", "definition": "Program-aided language models (PAL) have the model write a program, typically Python, that expresses its reasoning, and then delegate the actual computation to an interpreter.", "description": "Offloading arithmetic and logic to code removes calculation errors from the model's output, and PAL outperformed much larger chain-of-thought models on math word problems.", "example": null, "broader": [], "narrower": [], "related": ["program-of-thoughts", "chain-of-thought", "function-calling"], "introduced": 2022, "sources": [{"title": "PAL: Program-aided Language Models", "authors": "Gao et al.", "year": 2022, "url": "https://arxiv.org/abs/2211.10435"}], "url": "https://protologue.com/t/program-aided-language-models/", "citation": "Protologue. (2026). Program-Aided Language Models. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0045). https://protologue.com/t/program-aided-language-models/"} | |
| {"id": "PTL-0046", "slug": "program-of-thoughts", "term": "Program of Thoughts", "aliases": ["PoT"], "category": "Reasoning Elicitation", "definition": "Program of Thoughts (PoT) prompting expresses numerical reasoning as executable code, separating computation, done by an interpreter, from reasoning, done by the model.", "description": "Developed concurrently with PAL, it showed strong gains on financial and math question answering.", "example": null, "broader": [], "narrower": [], "related": ["program-aided-language-models", "chain-of-thought"], "introduced": 2022, "sources": [{"title": "Program of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks", "authors": "Chen et al.", "year": 2022, "url": "https://arxiv.org/abs/2211.12588"}], "url": "https://protologue.com/t/program-of-thoughts/", "citation": "Protologue. (2026). Program of Thoughts. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0046). https://protologue.com/t/program-of-thoughts/"} | |
| {"id": "PTL-0047", "slug": "self-refine", "term": "Self-Refine", "aliases": ["iterative refinement", "self-critique"], "category": "Self-Critique & Verification", "definition": "Self-Refine is an iterative method in which the same model generates an output, critiques it with specific feedback, and revises it, repeating until a stopping condition is met.", "description": "It requires no extra training or separate models. Madaan et al. reported improvements across tasks such as code optimization, dialogue, and math reasoning, though later work found that unaided self-correction of reasoning can fail without external signals.", "example": null, "broader": [], "narrower": ["chain-of-density"], "related": ["reflexion", "evaluator-optimizer", "chain-of-verification"], "introduced": 2023, "sources": [{"title": "Self-Refine: Iterative Refinement with Self-Feedback", "authors": "Madaan et al.", "year": 2023, "url": "https://arxiv.org/abs/2303.17651"}, {"title": "Large Language Models Cannot Self-Correct Reasoning Yet", "authors": "Huang et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.01798"}], "url": "https://protologue.com/t/self-refine/", "citation": "Protologue. (2026). Self-Refine. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0047). https://protologue.com/t/self-refine/"} | |
| {"id": "PTL-0048", "slug": "reflexion", "term": "Reflexion", "aliases": [], "category": "Self-Critique & Verification", "definition": "Reflexion is an agent technique in which, after a failed attempt, the model writes a verbal reflection on what went wrong and stores it in memory to guide its next attempt.", "description": "It is reinforcement through language rather than weight updates, and uses feedback signals such as unit-test results or environment rewards. Shinn et al. reported large gains on coding and sequential decision-making benchmarks.", "example": null, "broader": [], "narrower": [], "related": ["self-refine", "react", "agent-memory"], "introduced": 2023, "sources": [{"title": "Reflexion: Language Agents with Verbal Reinforcement Learning", "authors": "Shinn et al.", "year": 2023, "url": "https://arxiv.org/abs/2303.11366"}], "url": "https://protologue.com/t/reflexion/", "citation": "Protologue. (2026). Reflexion. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0048). https://protologue.com/t/reflexion/"} | |
| {"id": "PTL-0049", "slug": "chain-of-verification", "term": "Chain-of-Verification", "aliases": ["CoVe"], "category": "Self-Critique & Verification", "definition": "Chain-of-Verification (CoVe) reduces hallucination by having the model draft an answer, plan verification questions about its claims, answer those questions independently, and then produce a corrected final answer.", "description": "Answering the verification questions without seeing the original draft keeps the model from simply repeating its own errors.", "example": null, "broader": [], "narrower": [], "related": ["hallucination", "self-refine", "maieutic-prompting"], "introduced": 2023, "sources": [{"title": "Chain-of-Verification Reduces Hallucination in Large Language Models", "authors": "Dhuliawala et al.", "year": 2023, "url": "https://arxiv.org/abs/2309.11495"}], "url": "https://protologue.com/t/chain-of-verification/", "citation": "Protologue. (2026). Chain-of-Verification. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0049). https://protologue.com/t/chain-of-verification/"} | |
| {"id": "PTL-0050", "slug": "llm-as-a-judge", "term": "LLM-as-a-Judge", "aliases": ["model-graded evaluation", "LLM evaluator", "autorater"], "category": "Self-Critique & Verification", "definition": "LLM-as-a-judge is the use of a strong language model to grade, score, or compare the outputs of models against criteria, as a scalable substitute for human evaluation.", "description": "Zheng et al. found strong model judges agreed with human preferences at rates comparable to agreement between humans, while documenting biases toward the first-listed answer, longer answers, and the judge's own outputs.", "example": null, "broader": [], "narrower": [], "related": ["universal-self-consistency", "evaluator-optimizer", "process-reward-model"], "introduced": 2023, "sources": [{"title": "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena", "authors": "Zheng et al.", "year": 2023, "url": "https://arxiv.org/abs/2306.05685"}], "url": "https://protologue.com/t/llm-as-a-judge/", "citation": "Protologue. (2026). LLM-as-a-Judge. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0050). https://protologue.com/t/llm-as-a-judge/"} | |
| {"id": "PTL-0051", "slug": "multi-agent-debate", "term": "Multi-Agent Debate", "aliases": ["LLM debate", "society of minds"], "category": "Self-Critique & Verification", "definition": "Multi-agent debate has several model instances propose answers, read each other's reasoning, and revise their answers over multiple rounds until they converge.", "description": "Du et al. found that debate improved mathematical reasoning and factual accuracy over single-model answers.", "example": null, "broader": [], "narrower": [], "related": ["mixture-of-agents", "self-consistency"], "introduced": 2023, "sources": [{"title": "Improving Factuality and Reasoning in Language Models through Multiagent Debate", "authors": "Du et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.14325"}], "url": "https://protologue.com/t/multi-agent-debate/", "citation": "Protologue. (2026). Multi-Agent Debate. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0051). https://protologue.com/t/multi-agent-debate/"} | |
| {"id": "PTL-0052", "slug": "mixture-of-agents", "term": "Mixture-of-Agents", "aliases": ["MoA"], "category": "Self-Critique & Verification", "definition": "Mixture-of-Agents (MoA) arranges language models in layers, where each model receives all outputs from the previous layer as auxiliary input and an aggregator synthesizes a final response.", "description": "Wang et al. reported that a mixture of open models outperformed a single strong proprietary model on an instruction-following benchmark.", "example": null, "broader": [], "narrower": [], "related": ["multi-agent-debate", "parallelization"], "introduced": 2024, "sources": [{"title": "Mixture-of-Agents Enhances Large Language Model Capabilities", "authors": "Wang et al.", "year": 2024, "url": "https://arxiv.org/abs/2406.04692"}], "url": "https://protologue.com/t/mixture-of-agents/", "citation": "Protologue. (2026). Mixture-of-Agents. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0052). https://protologue.com/t/mixture-of-agents/"} | |
| {"id": "PTL-0053", "slug": "process-reward-model", "term": "Process Reward Model", "aliases": ["PRM", "step-level verifier", "process supervision"], "category": "Self-Critique & Verification", "definition": "A process reward model (PRM) scores each intermediate step of a model's reasoning, rather than only the final answer, and is used to select or train better reasoning chains.", "description": "Lightman et al. showed process supervision outperformed outcome supervision for selecting correct solutions to competition math problems, and released a large dataset of step-level human labels.", "example": null, "broader": [], "narrower": [], "related": ["best-of-n-sampling", "test-time-compute-scaling", "llm-as-a-judge"], "introduced": 2023, "sources": [{"title": "Let's Verify Step by Step", "authors": "Lightman et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.20050"}], "url": "https://protologue.com/t/process-reward-model/", "citation": "Protologue. (2026). Process Reward Model. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0053). https://protologue.com/t/process-reward-model/"} | |
| {"id": "PTL-0054", "slug": "best-of-n-sampling", "term": "Best-of-N Sampling", "aliases": ["rejection sampling", "best-of-n", "BoN"], "category": "Self-Critique & Verification", "definition": "Best-of-N sampling generates N candidate outputs and returns the one ranked highest by a verifier, reward model, or scoring function.", "description": "It is the simplest form of trading inference compute for quality. Its effectiveness depends on the verifier, and optimizing too hard against an imperfect reward model can select outputs that game it.", "example": null, "broader": [], "narrower": [], "related": ["self-consistency", "process-reward-model", "test-time-compute-scaling"], "introduced": null, "sources": [{"title": "Let's Verify Step by Step", "authors": "Lightman et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.20050"}, {"title": "Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters", "authors": "Snell et al.", "year": 2024, "url": "https://arxiv.org/abs/2408.03314"}], "url": "https://protologue.com/t/best-of-n-sampling/", "citation": "Protologue. (2026). Best-of-N Sampling. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0054). https://protologue.com/t/best-of-n-sampling/"} | |
| {"id": "PTL-0055", "slug": "retrieval-augmented-generation", "term": "Retrieval-Augmented Generation", "aliases": ["RAG", "retrieval augmentation", "grounded generation"], "category": "Retrieval & Tool Use", "definition": "Retrieval-augmented generation (RAG) supplies a language model with passages retrieved from an external corpus at query time, so its output is grounded in that information rather than only in its trained parameters.", "description": "Lewis et al. introduced RAG as a jointly trained retriever and generator. In common usage the term now covers any pipeline that retrieves documents, typically by embedding search, and inserts them into the prompt. RAG reduces hallucination and allows knowledge to be updated without retraining.", "example": null, "broader": [], "narrower": ["hypothetical-document-embeddings", "self-rag", "graphrag", "agentic-rag", "corrective-rag", "active-retrieval"], "related": ["hallucination", "context-engineering"], "introduced": 2020, "sources": [{"title": "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", "authors": "Lewis et al.", "year": 2020, "url": "https://arxiv.org/abs/2005.11401"}], "url": "https://protologue.com/t/retrieval-augmented-generation/", "citation": "Protologue. (2026). Retrieval-Augmented Generation. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0055). https://protologue.com/t/retrieval-augmented-generation/"} | |
| {"id": "PTL-0056", "slug": "hypothetical-document-embeddings", "term": "Hypothetical Document Embeddings", "aliases": ["HyDE"], "category": "Retrieval & Tool Use", "definition": "Hypothetical Document Embeddings (HyDE) improves retrieval by having a model write a hypothetical answer to the query, embedding that answer, and searching for real documents similar to it.", "description": "The generated document may contain errors, but its embedding tends to lie closer to relevant real documents than the short query does.", "example": null, "broader": ["retrieval-augmented-generation"], "narrower": [], "related": [], "introduced": 2022, "sources": [{"title": "Precise Zero-Shot Dense Retrieval without Relevance Labels", "authors": "Gao et al.", "year": 2022, "url": "https://arxiv.org/abs/2212.10496"}], "url": "https://protologue.com/t/hypothetical-document-embeddings/", "citation": "Protologue. (2026). Hypothetical Document Embeddings. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0056). https://protologue.com/t/hypothetical-document-embeddings/"} | |
| {"id": "PTL-0057", "slug": "self-rag", "term": "Self-RAG", "aliases": [], "category": "Retrieval & Tool Use", "definition": "Self-RAG trains a model to decide when to retrieve, and to emit special reflection tokens that critique whether retrieved passages are relevant and whether its own output is supported by them.", "description": "Retrieving on demand rather than for every query avoids adding irrelevant context and lets the model's critique steer generation.", "example": null, "broader": ["retrieval-augmented-generation"], "narrower": [], "related": ["chain-of-verification"], "introduced": 2023, "sources": [{"title": "Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection", "authors": "Asai et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.11511"}], "url": "https://protologue.com/t/self-rag/", "citation": "Protologue. (2026). Self-RAG. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0057). https://protologue.com/t/self-rag/"} | |
| {"id": "PTL-0058", "slug": "react", "term": "ReAct", "aliases": ["Reason + Act", "thought-action-observation loop"], "category": "Retrieval & Tool Use", "definition": "ReAct is a prompting pattern that interleaves reasoning traces (\"Thought\") with actions such as tool calls (\"Action\") and their results (\"Observation\"), letting a model plan, act, and update its plan in a loop.", "description": "Yao et al. showed that combining reasoning and acting outperformed either alone on question answering and interactive decision-making tasks. ReAct is the template for most tool-using agent loops.", "example": "Thought: I need the population of the capital of France.\nAction: search(\"capital of France\")\nObservation: Paris\nThought: Now find the population of Paris.\n", "broader": [], "narrower": [], "related": ["function-calling", "chain-of-thought", "self-ask", "ai-agent", "reflexion"], "introduced": 2022, "sources": [{"title": "ReAct: Synergizing Reasoning and Acting in Language Models", "authors": "Yao et al.", "year": 2022, "url": "https://arxiv.org/abs/2210.03629"}], "url": "https://protologue.com/t/react/", "citation": "Protologue. (2026). ReAct. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0058). https://protologue.com/t/react/"} | |
| {"id": "PTL-0059", "slug": "toolformer", "term": "Toolformer", "aliases": [], "category": "Retrieval & Tool Use", "definition": "Toolformer is a method in which a language model teaches itself to use external tools, such as a calculator or search API, by generating candidate API calls in text and keeping those that reduce its prediction loss.", "description": "It showed that tool-use behavior could be learned in a self-supervised way from only a handful of demonstrations per tool.", "example": null, "broader": [], "narrower": [], "related": ["function-calling", "react"], "introduced": 2023, "sources": [{"title": "Toolformer: Language Models Can Teach Themselves to Use Tools", "authors": "Schick et al.", "year": 2023, "url": "https://arxiv.org/abs/2302.04761"}], "url": "https://protologue.com/t/toolformer/", "citation": "Protologue. (2026). Toolformer. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0059). https://protologue.com/t/toolformer/"} | |
| {"id": "PTL-0060", "slug": "function-calling", "term": "Function Calling", "aliases": ["tool use", "tool calling"], "category": "Retrieval & Tool Use", "definition": "Function calling, or tool use, is a model capability in which the model outputs a structured request to invoke a developer-defined function with arguments, which the application executes and returns as a result to the model.", "description": "Tools are usually described to the model with a name, a natural-language description, and a JSON Schema for parameters. Clear tool descriptions function as prompts in their own right and strongly affect when and how tools are used.", "example": null, "broader": [], "narrower": ["web-search-tool", "code-execution-tool"], "related": ["react", "model-context-protocol", "structured-outputs", "toolformer"], "introduced": null, "sources": [{"title": "Tool use with Claude", "authors": "Anthropic", "year": 2024, "url": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview"}, {"title": "Gorilla: Large Language Model Connected with Massive APIs", "authors": "Patil et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.15334"}], "url": "https://protologue.com/t/function-calling/", "citation": "Protologue. (2026). Function Calling. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0060). https://protologue.com/t/function-calling/"} | |
| {"id": "PTL-0061", "slug": "model-context-protocol", "term": "Model Context Protocol", "aliases": ["MCP"], "category": "Retrieval & Tool Use", "definition": "The Model Context Protocol (MCP) is an open standard, introduced by Anthropic in 2024, that defines how applications expose tools, data resources, and prompt templates to language-model clients through a common client-server interface.", "description": "MCP replaces one-off integrations between each model application and each data source with a single protocol, so any compliant client can use any compliant server.", "example": null, "broader": [], "narrower": [], "related": ["function-calling", "context-engineering"], "introduced": 2024, "sources": [{"title": "Introducing the Model Context Protocol", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/news/model-context-protocol"}, {"title": "Specification (version 2026-07-28)", "authors": "Model Context Protocol", "year": 2026, "url": "https://modelcontextprotocol.io/specification/latest"}], "url": "https://protologue.com/t/model-context-protocol/", "citation": "Protologue. (2026). Model Context Protocol. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0061). https://protologue.com/t/model-context-protocol/"} | |
| {"id": "PTL-0062", "slug": "structured-outputs", "term": "Structured Outputs", "aliases": ["JSON mode", "constrained decoding", "schema-constrained generation"], "category": "Retrieval & Tool Use", "definition": "Structured outputs constrain a language model to produce responses that conform to a specified format, typically a JSON Schema, either through instructions or through constrained decoding that guarantees validity.", "description": "Instruction-only approaches can still produce malformed output, while constrained decoding masks tokens that would violate the schema at each step. Structured outputs are essential for passing model results reliably to downstream code.", "example": null, "broader": [], "narrower": [], "related": ["function-calling", "prefill", "delimiters"], "introduced": null, "sources": [{"title": "Structured model outputs", "authors": "OpenAI", "year": 2024, "url": "https://developers.openai.com/api/docs/guides/structured-outputs"}, {"title": "Efficient Guided Generation for Large Language Models", "authors": "Willard & Louf", "year": 2023, "url": "https://arxiv.org/abs/2307.09702"}], "url": "https://protologue.com/t/structured-outputs/", "citation": "Protologue. (2026). Structured Outputs. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0062). https://protologue.com/t/structured-outputs/"} | |
| {"id": "PTL-0063", "slug": "ai-agent", "term": "AI Agent", "aliases": ["LLM agent", "agentic system", "autonomous agent"], "category": "Agents & Orchestration", "definition": "An AI agent is a system in which a language model dynamically directs its own process and tool use in a loop, deciding what actions to take based on environment feedback until a task is complete.", "description": "Anthropic's widely cited taxonomy distinguishes agents from workflows, in which model calls and tools follow predefined code paths. Agents trade predictability and cost for flexibility on open-ended tasks.", "example": null, "broader": [], "narrower": ["computer-use-agent", "coding-agent"], "related": ["react", "agentic-workflow", "orchestrator-workers", "agent-memory", "context-engineering"], "introduced": null, "sources": [{"title": "Building effective agents", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/research/building-effective-agents"}], "url": "https://protologue.com/t/ai-agent/", "citation": "Protologue. (2026). AI Agent. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0063). https://protologue.com/t/ai-agent/"} | |
| {"id": "PTL-0064", "slug": "agentic-workflow", "term": "Agentic Workflow", "aliases": ["LLM workflow"], "category": "Agents & Orchestration", "definition": "An agentic workflow is a system in which language models and tools are orchestrated through predefined code paths, as opposed to an agent that chooses its own steps.", "description": "Common workflow patterns include prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer loops. Workflows are generally preferred when a task can be cleanly decomposed in advance.", "example": null, "broader": [], "narrower": ["prompt-chaining", "routing", "parallelization", "orchestrator-workers", "evaluator-optimizer", "plan-and-execute"], "related": ["ai-agent"], "introduced": null, "sources": [{"title": "Building effective agents", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/research/building-effective-agents"}], "url": "https://protologue.com/t/agentic-workflow/", "citation": "Protologue. (2026). Agentic Workflow. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0064). https://protologue.com/t/agentic-workflow/"} | |
| {"id": "PTL-0065", "slug": "prompt-chaining", "term": "Prompt Chaining", "aliases": ["LLM chains", "multi-step prompting"], "category": "Agents & Orchestration", "definition": "Prompt chaining decomposes a task into a fixed sequence of model calls, where each call processes the output of the previous one, often with programmatic checks between steps.", "description": "Chaining trades latency for accuracy by making each call simpler. Wu et al. found chaining also improved transparency and controllability for users building with models.", "example": null, "broader": ["agentic-workflow"], "narrower": [], "related": ["least-to-most-prompting", "routing"], "introduced": 2021, "sources": [{"title": "AI Chains: Transparent and Controllable Human-AI Interaction by Chaining Large Language Model Prompts", "authors": "Wu et al.", "year": 2021, "url": "https://arxiv.org/abs/2110.01691"}, {"title": "Building effective agents", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/research/building-effective-agents"}], "url": "https://protologue.com/t/prompt-chaining/", "citation": "Protologue. (2026). Prompt Chaining. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0065). https://protologue.com/t/prompt-chaining/"} | |
| {"id": "PTL-0066", "slug": "routing", "term": "Routing", "aliases": ["model routing", "query routing"], "category": "Agents & Orchestration", "definition": "Routing is a workflow pattern that classifies an incoming request and directs it to a specialized prompt, tool, or model suited to that category.", "description": "Routing allows separate prompts to be optimized for distinct cases, and lets easy queries go to smaller, cheaper models while hard ones go to more capable models.", "example": null, "broader": ["agentic-workflow"], "narrower": [], "related": ["prompt-chaining"], "introduced": null, "sources": [{"title": "Building effective agents", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/research/building-effective-agents"}], "url": "https://protologue.com/t/routing/", "citation": "Protologue. (2026). Routing. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0066). https://protologue.com/t/routing/"} | |
| {"id": "PTL-0067", "slug": "parallelization", "term": "Parallelization", "aliases": ["sectioning", "voting"], "category": "Agents & Orchestration", "definition": "Parallelization is a workflow pattern that runs several model calls simultaneously and aggregates their outputs, either by splitting a task into independent sections or by running the same task several times and voting.", "description": "Sectioning reduces latency and keeps each call focused, while voting improves confidence, for example by running several independent safety checks.", "example": null, "broader": ["agentic-workflow"], "narrower": [], "related": ["self-consistency", "skeleton-of-thought", "mixture-of-agents"], "introduced": null, "sources": [{"title": "Building effective agents", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/research/building-effective-agents"}], "url": "https://protologue.com/t/parallelization/", "citation": "Protologue. (2026). Parallelization. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0067). https://protologue.com/t/parallelization/"} | |
| {"id": "PTL-0068", "slug": "orchestrator-workers", "term": "Orchestrator-Workers", "aliases": ["subagents", "multi-agent orchestration", "supervisor pattern"], "category": "Agents & Orchestration", "definition": "Orchestrator-workers is a pattern in which a central model dynamically breaks a task into subtasks, delegates them to worker model calls or subagents, and synthesizes their results.", "description": "Unlike parallelization, the subtasks are not fixed in advance but decided by the orchestrator for each input, which suits tasks such as multi-file code changes or broad research.", "example": null, "broader": ["agentic-workflow"], "narrower": [], "related": ["ai-agent", "parallelization", "meta-prompting"], "introduced": null, "sources": [{"title": "Building effective agents", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/research/building-effective-agents"}], "url": "https://protologue.com/t/orchestrator-workers/", "citation": "Protologue. (2026). Orchestrator-Workers. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0068). https://protologue.com/t/orchestrator-workers/"} | |
| {"id": "PTL-0069", "slug": "evaluator-optimizer", "term": "Evaluator-Optimizer", "aliases": [], "category": "Agents & Orchestration", "definition": "Evaluator-optimizer is a workflow loop in which one model call generates a response and another evaluates it against criteria and provides feedback, repeating until the output passes.", "description": "It works best when evaluation criteria are clear and when feedback demonstrably improves the output, as in literary translation or iterative search.", "example": null, "broader": ["agentic-workflow"], "narrower": [], "related": ["self-refine", "llm-as-a-judge", "reflexion"], "introduced": null, "sources": [{"title": "Building effective agents", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/research/building-effective-agents"}], "url": "https://protologue.com/t/evaluator-optimizer/", "citation": "Protologue. (2026). Evaluator-Optimizer. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0069). https://protologue.com/t/evaluator-optimizer/"} | |
| {"id": "PTL-0070", "slug": "meta-prompting", "term": "Meta-Prompting", "aliases": [], "category": "Agents & Orchestration", "definition": "Meta-prompting uses a single model as a conductor that breaks a task down and writes prompts for fresh instances of itself acting as specialized experts, then integrates their outputs.", "description": "The term is also used more loosely for asking a model to write or improve a prompt.", "example": null, "broader": [], "narrower": [], "related": ["orchestrator-workers", "self-discover", "automatic-prompt-engineer"], "introduced": 2024, "sources": [{"title": "Meta-Prompting: Enhancing Language Models with Task-Agnostic Scaffolding", "authors": "Suzgun & Kalai", "year": 2024, "url": "https://arxiv.org/abs/2401.12954"}], "url": "https://protologue.com/t/meta-prompting/", "citation": "Protologue. (2026). Meta-Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0070). https://protologue.com/t/meta-prompting/"} | |
| {"id": "PTL-0071", "slug": "agent-memory", "term": "Agent Memory", "aliases": ["long-term memory", "memory stream"], "category": "Agents & Orchestration", "definition": "Agent memory is the set of mechanisms that let a language-model agent store information beyond a single context window, such as conversation summaries, retrievable records of past events, and reflections, and bring the relevant parts back into context later.", "description": "Park et al.'s generative agents scored memories by recency, importance, and relevance and periodically synthesized higher-level reflections. MemGPT treated the context window like main memory and paged information in and out of external storage.", "example": null, "broader": [], "narrower": [], "related": ["ai-agent", "context-engineering", "reflexion", "retrieval-augmented-generation"], "introduced": null, "sources": [{"title": "Generative Agents: Interactive Simulacra of Human Behavior", "authors": "Park et al.", "year": 2023, "url": "https://arxiv.org/abs/2304.03442"}, {"title": "MemGPT: Towards LLMs as Operating Systems", "authors": "Packer et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.08560"}], "url": "https://protologue.com/t/agent-memory/", "citation": "Protologue. (2026). Agent Memory. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0071). https://protologue.com/t/agent-memory/"} | |
| {"id": "PTL-0072", "slug": "context-engineering", "term": "Context Engineering", "aliases": ["context management"], "category": "Agents & Orchestration", "definition": "Context engineering is the practice of curating the full set of tokens a model sees at each step, including instructions, tools, retrieved data, memory, and conversation history, to maximize the chance of the desired behavior within a limited attention budget.", "description": "The term gained currency in 2025 as agents ran for many steps and the main challenge shifted from wording a single prompt to deciding what information enters and leaves the context window over time, through techniques such as compaction, structured note-taking, and subagents.", "example": null, "broader": ["prompt-engineering"], "narrower": [], "related": ["context-window", "agent-memory", "retrieval-augmented-generation", "ai-agent", "lost-in-the-middle"], "introduced": 2025, "sources": [{"title": "Effective context engineering for AI agents", "authors": "Anthropic", "year": 2025, "url": "https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents"}], "url": "https://protologue.com/t/context-engineering/", "citation": "Protologue. (2026). Context Engineering. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0072). https://protologue.com/t/context-engineering/"} | |
| {"id": "PTL-0073", "slug": "automatic-prompt-engineer", "term": "Automatic Prompt Engineer", "aliases": ["APE"], "category": "Prompt Optimization", "definition": "Automatic Prompt Engineer (APE) uses a language model to generate candidate instructions for a task from input-output examples, scores each candidate on held-out data, and selects the best one.", "description": "APE discovered a zero-shot chain-of-thought trigger that outperformed \"Let's think step by step\" on some benchmarks, and framed prompt writing as a search problem that models can solve themselves.", "example": null, "broader": [], "narrower": [], "related": ["prompt-engineering", "opro", "dspy", "meta-prompting"], "introduced": 2022, "sources": [{"title": "Large Language Models Are Human-Level Prompt Engineers", "authors": "Zhou et al.", "year": 2022, "url": "https://arxiv.org/abs/2211.01910"}], "url": "https://protologue.com/t/automatic-prompt-engineer/", "citation": "Protologue. (2026). Automatic Prompt Engineer. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0073). https://protologue.com/t/automatic-prompt-engineer/"} | |
| {"id": "PTL-0074", "slug": "opro", "term": "Optimization by Prompting", "aliases": ["OPRO", "LLMs as optimizers"], "category": "Prompt Optimization", "definition": "Optimization by PROmpting (OPRO) uses a language model as an optimizer, giving it a meta-prompt containing previously tried prompts and their scores and asking it to propose a better prompt, repeating over many rounds.", "description": "OPRO found instructions such as \"Take a deep breath and work on this problem step-by-step\" that improved math benchmark accuracy for the model being optimized.", "example": null, "broader": [], "narrower": [], "related": ["automatic-prompt-engineer", "dspy"], "introduced": 2023, "sources": [{"title": "Large Language Models as Optimizers", "authors": "Yang et al.", "year": 2023, "url": "https://arxiv.org/abs/2309.03409"}], "url": "https://protologue.com/t/opro/", "citation": "Protologue. (2026). Optimization by Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0074). https://protologue.com/t/opro/"} | |
| {"id": "PTL-0075", "slug": "dspy", "term": "DSPy", "aliases": ["Declarative Self-improving Python"], "category": "Prompt Optimization", "definition": "DSPy is a framework that treats language-model pipelines as programs of declarative modules, and compiles them by automatically optimizing the prompts and few-shot demonstrations for each module against a metric.", "description": "Rather than hand-tuning prompt strings, developers specify input-output signatures and let optimizers bootstrap demonstrations or search instructions, making pipelines portable across models.", "example": null, "broader": [], "narrower": [], "related": ["automatic-prompt-engineer", "opro", "prompt-template"], "introduced": 2023, "sources": [{"title": "DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines", "authors": "Khattab et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.03714"}], "url": "https://protologue.com/t/dspy/", "citation": "Protologue. (2026). DSPy. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0075). https://protologue.com/t/dspy/"} | |
| {"id": "PTL-0076", "slug": "prompt-tuning", "term": "Prompt Tuning", "aliases": ["soft prompts", "soft prompt tuning"], "category": "Prompt Optimization", "definition": "Prompt tuning learns a small set of continuous \"soft prompt\" embeddings that are prepended to the input, by gradient descent, while keeping the language model's weights frozen.", "description": "Lester et al. showed that as models grow, prompt tuning approaches the quality of full fine-tuning while storing only a tiny number of task-specific parameters. Soft prompts are vectors rather than readable text.", "example": null, "broader": [], "narrower": [], "related": ["prefix-tuning", "low-rank-adaptation"], "introduced": 2021, "sources": [{"title": "The Power of Scale for Parameter-Efficient Prompt Tuning", "authors": "Lester et al.", "year": 2021, "url": "https://arxiv.org/abs/2104.08691"}], "url": "https://protologue.com/t/prompt-tuning/", "citation": "Protologue. (2026). Prompt Tuning. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0076). https://protologue.com/t/prompt-tuning/"} | |
| {"id": "PTL-0077", "slug": "prefix-tuning", "term": "Prefix Tuning", "aliases": [], "category": "Prompt Optimization", "definition": "Prefix tuning learns continuous task-specific vectors that are prepended to the activations at every layer of a frozen language model, steering generation without changing the model's weights.", "description": "It was an early parameter-efficient alternative to fine-tuning for generation tasks such as table-to-text and summarization.", "example": null, "broader": [], "narrower": [], "related": ["prompt-tuning", "low-rank-adaptation"], "introduced": 2021, "sources": [{"title": "Prefix-Tuning: Optimizing Continuous Prompts for Generation", "authors": "Li & Liang", "year": 2021, "url": "https://arxiv.org/abs/2101.00190"}], "url": "https://protologue.com/t/prefix-tuning/", "citation": "Protologue. (2026). Prefix Tuning. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0077). https://protologue.com/t/prefix-tuning/"} | |
| {"id": "PTL-0078", "slug": "low-rank-adaptation", "term": "Low-Rank Adaptation", "aliases": ["LoRA"], "category": "Prompt Optimization", "definition": "Low-rank adaptation (LoRA) fine-tunes a language model by training small low-rank matrices added to its weight layers while freezing the original weights, drastically reducing the number of trainable parameters.", "description": "LoRA is a common alternative when prompting alone cannot reach the required behavior, and adapters can be swapped per task on one base model.", "example": null, "broader": [], "narrower": [], "related": ["prompt-tuning", "prefix-tuning", "instruction-tuning"], "introduced": 2021, "sources": [{"title": "LoRA: Low-Rank Adaptation of Large Language Models", "authors": "Hu et al.", "year": 2021, "url": "https://arxiv.org/abs/2106.09685"}], "url": "https://protologue.com/t/low-rank-adaptation/", "citation": "Protologue. (2026). Low-Rank Adaptation. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0078). https://protologue.com/t/low-rank-adaptation/"} | |
| {"id": "PTL-0079", "slug": "directional-stimulus-prompting", "term": "Directional Stimulus Prompting", "aliases": [], "category": "Prompt Optimization", "definition": "Directional stimulus prompting trains a small policy model to generate instance-specific hints, such as keywords, that are added to the prompt to steer a large frozen model toward desired outputs.", "description": "The policy model can be trained with supervised learning and reinforcement learning, without access to the large model's weights.", "example": null, "broader": [], "narrower": [], "related": ["prompt-tuning", "opro"], "introduced": 2023, "sources": [{"title": "Guiding Large Language Models via Directional Stimulus Prompting", "authors": "Li et al.", "year": 2023, "url": "https://arxiv.org/abs/2302.11520"}], "url": "https://protologue.com/t/directional-stimulus-prompting/", "citation": "Protologue. (2026). Directional Stimulus Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0079). https://protologue.com/t/directional-stimulus-prompting/"} | |
| {"id": "PTL-0080", "slug": "prompt-compression", "term": "Prompt Compression", "aliases": ["LLMLingua", "context compression"], "category": "Prompt Optimization", "definition": "Prompt compression shortens a prompt by removing tokens that contribute little information, typically scored by a smaller language model, to cut cost and latency while preserving task performance.", "description": "LLMLingua used a small model's perplexity to drop low-information tokens and reported high compression ratios with limited performance loss.", "example": null, "broader": [], "narrower": [], "related": ["context-engineering", "prompt-caching"], "introduced": 2023, "sources": [{"title": "LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models", "authors": "Jiang et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.05736"}], "url": "https://protologue.com/t/prompt-compression/", "citation": "Protologue. (2026). Prompt Compression. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0080). https://protologue.com/t/prompt-compression/"} | |
| {"id": "PTL-0081", "slug": "prompt-caching", "term": "Prompt Caching", "aliases": ["context caching", "prefix caching"], "category": "Prompt Optimization", "definition": "Prompt caching stores the model's processed state for a reused prompt prefix, such as a long system prompt or document, so later requests sharing that prefix are cheaper and faster.", "description": "Because caches match on an exact prefix, prompts are structured with stable content first and variable content last.", "example": null, "broader": [], "narrower": [], "related": ["system-prompt", "prompt-compression", "context-engineering"], "introduced": 2024, "sources": [{"title": "Prompt caching", "authors": "Anthropic", "year": 2024, "url": "https://platform.claude.com/docs/en/build-with-claude/prompt-caching"}], "url": "https://protologue.com/t/prompt-caching/", "citation": "Protologue. (2026). Prompt Caching. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0081). https://protologue.com/t/prompt-caching/"} | |
| {"id": "PTL-0082", "slug": "emotion-prompting", "term": "Emotion Prompting", "aliases": ["EmotionPrompt", "stimulus prompting"], "category": "Prompt Optimization", "definition": "Emotion prompting appends emotional or motivational phrases, such as \"This is very important to my career,\" to a prompt in an attempt to improve model performance.", "description": "Li et al. reported gains on several benchmarks with such \"EmotionPrompt\" stimuli. Effects of these cues vary across models and tasks and are best treated as an empirical question for each setup.", "example": null, "broader": [], "narrower": [], "related": ["prompt-sensitivity", "role-prompting"], "introduced": 2023, "sources": [{"title": "Large Language Models Understand and Can be Enhanced by Emotional Stimuli", "authors": "Li et al.", "year": 2023, "url": "https://arxiv.org/abs/2307.11760"}], "url": "https://protologue.com/t/emotion-prompting/", "citation": "Protologue. (2026). Emotion Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0082). https://protologue.com/t/emotion-prompting/"} | |
| {"id": "PTL-0083", "slug": "reasoning-model", "term": "Reasoning Model", "aliases": ["large reasoning model", "LRM", "thinking model"], "category": "Reasoning Models & Test-Time Compute", "definition": "A reasoning model is a language model trained, typically with reinforcement learning, to produce an extended internal chain of thought before answering, so that it improves with more thinking time on math, coding, and planning tasks.", "description": "OpenAI's o1, announced in 2024, popularized the category, and DeepSeek-R1 showed that reinforcement learning with verifiable rewards could produce long reasoning behaviors in an openly released model. Prompting such models differs from prompting standard models, because step-by-step instructions and few-shot reasoning examples are often unnecessary.", "example": null, "broader": [], "narrower": ["extended-thinking"], "related": ["chain-of-thought", "test-time-compute-scaling", "self-taught-reasoner"], "introduced": 2024, "sources": [{"title": "Learning to reason with LLMs", "authors": "OpenAI", "year": 2024, "url": "https://openai.com/index/learning-to-reason-with-llms/"}, {"title": "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning", "authors": "DeepSeek-AI", "year": 2025, "url": "https://arxiv.org/abs/2501.12948"}], "url": "https://protologue.com/t/reasoning-model/", "citation": "Protologue. (2026). Reasoning Model. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0083). https://protologue.com/t/reasoning-model/"} | |
| {"id": "PTL-0084", "slug": "test-time-compute-scaling", "term": "Test-Time Compute Scaling", "aliases": ["inference-time scaling", "test-time scaling"], "category": "Reasoning Models & Test-Time Compute", "definition": "Test-time compute scaling improves a model's answers by spending more computation at inference, through longer reasoning, more samples, search, or verification, rather than by training a larger model.", "description": "Snell et al. found that allocating test-time compute adaptively per prompt could be more effective than scaling model parameters for some problems. Self-consistency, best-of-N, tree search, and reasoning models are all forms of test-time scaling.", "example": null, "broader": [], "narrower": ["budget-forcing", "recursive-language-model", "sleep-time-compute"], "related": ["reasoning-model", "self-consistency", "best-of-n-sampling", "tree-of-thoughts", "process-reward-model"], "introduced": 2024, "sources": [{"title": "Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters", "authors": "Snell et al.", "year": 2024, "url": "https://arxiv.org/abs/2408.03314"}], "url": "https://protologue.com/t/test-time-compute-scaling/", "citation": "Protologue. (2026). Test-Time Compute Scaling. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0084). https://protologue.com/t/test-time-compute-scaling/"} | |
| {"id": "PTL-0085", "slug": "extended-thinking", "term": "Extended Thinking", "aliases": ["thinking mode", "thinking budget", "reasoning effort"], "category": "Reasoning Models & Test-Time Compute", "definition": "Extended thinking is a model mode in which the model generates a separate block of reasoning before its final response, with a developer-controlled setting that trades latency and cost for answer quality.", "description": "Early implementations exposed a fixed thinking token budget. Newer Claude models replace it with adaptive thinking, in which the model decides whether and how much to reason on each request and the developer sets an effort level instead of a token budget. Anthropic's documentation states that the fixed budget is deprecated on Claude 4.6 models and rejected by Claude 4.7 and later models. Vendor guidance for these modes favors high-level instructions about how to think over prescriptive step-by-step scripts.", "example": null, "broader": ["reasoning-model"], "narrower": [], "related": ["scratchpad", "chain-of-thought"], "introduced": null, "sources": [{"title": "Extended thinking", "authors": "Anthropic", "year": 2026, "url": "https://platform.claude.com/docs/en/build-with-claude/extended-thinking"}], "url": "https://protologue.com/t/extended-thinking/", "citation": "Protologue. (2026). Extended Thinking. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0085). https://protologue.com/t/extended-thinking/"} | |
| {"id": "PTL-0086", "slug": "self-taught-reasoner", "term": "Self-Taught Reasoner", "aliases": ["STaR"], "category": "Reasoning Models & Test-Time Compute", "definition": "The Self-Taught Reasoner (STaR) bootstraps reasoning ability by having a model generate rationales, keeping those that lead to correct answers, and fine-tuning on them in repeated rounds.", "description": "For problems it fails, STaR provides the correct answer as a hint and asks the model to produce a rationale for it, a step called rationalization.", "example": null, "broader": [], "narrower": [], "related": ["reasoning-model", "chain-of-thought"], "introduced": 2022, "sources": [{"title": "STaR: Bootstrapping Reasoning With Reasoning", "authors": "Zelikman et al.", "year": 2022, "url": "https://arxiv.org/abs/2203.14465"}], "url": "https://protologue.com/t/self-taught-reasoner/", "citation": "Protologue. (2026). Self-Taught Reasoner. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0086). https://protologue.com/t/self-taught-reasoner/"} | |
| {"id": "PTL-0087", "slug": "prompt-injection", "term": "Prompt Injection", "aliases": ["goal hijacking", "instruction injection"], "category": "Security & Adversarial Prompting", "definition": "Prompt injection is an attack in which text supplied to a language model, by a user or through data the model processes, contains instructions that override or subvert the instructions of the application's developer.", "description": "The name was coined in 2022 by analogy with SQL injection, because models cannot reliably separate trusted instructions from untrusted data that share the same context. It is the most prominent security risk for applications built on language models, especially agents with tool access.", "example": null, "broader": [], "narrower": ["indirect-prompt-injection", "prompt-leaking"], "related": ["jailbreak", "instruction-hierarchy", "spotlighting"], "introduced": 2022, "sources": [{"title": "Prompt injection attacks against GPT-3", "authors": "Willison", "year": 2022, "url": "https://simonwillison.net/2022/Sep/12/prompt-injection/"}, {"title": "Ignore Previous Prompt: Attack Techniques For Language Models", "authors": "Perez & Ribeiro", "year": 2022, "url": "https://arxiv.org/abs/2211.09527"}], "url": "https://protologue.com/t/prompt-injection/", "citation": "Protologue. (2026). Prompt Injection. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0087). https://protologue.com/t/prompt-injection/"} | |
| {"id": "PTL-0088", "slug": "indirect-prompt-injection", "term": "Indirect Prompt Injection", "aliases": [], "category": "Security & Adversarial Prompting", "definition": "Indirect prompt injection places malicious instructions inside content a model will later retrieve or process, such as a web page, email, or document, so the attack is triggered without the attacker interacting with the model directly.", "description": "Greshake et al. demonstrated that injected content could make integrated applications exfiltrate data, spread to other users, or manipulate outputs. Risk grows with an agent's access to tools and private data.", "example": null, "broader": ["prompt-injection"], "narrower": ["tool-poisoning"], "related": ["retrieval-augmented-generation", "spotlighting", "ai-agent"], "introduced": 2023, "sources": [{"title": "Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection", "authors": "Greshake et al.", "year": 2023, "url": "https://arxiv.org/abs/2302.12173"}], "url": "https://protologue.com/t/indirect-prompt-injection/", "citation": "Protologue. (2026). Indirect Prompt Injection. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0088). https://protologue.com/t/indirect-prompt-injection/"} | |
| {"id": "PTL-0089", "slug": "prompt-leaking", "term": "Prompt Leaking", "aliases": [], "category": "Security & Adversarial Prompting", "definition": "Prompt leaking is an attack that tricks a model into revealing its hidden system prompt or other confidential instructions.", "description": "System prompts should be treated as potentially discoverable, so they should not contain secrets such as API keys.", "example": null, "broader": ["prompt-injection"], "narrower": [], "related": ["system-prompt"], "introduced": null, "sources": [{"title": "Ignore Previous Prompt: Attack Techniques For Language Models", "authors": "Perez & Ribeiro", "year": 2022, "url": "https://arxiv.org/abs/2211.09527"}], "url": "https://protologue.com/t/prompt-leaking/", "citation": "Protologue. (2026). Prompt Leaking. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0089). https://protologue.com/t/prompt-leaking/"} | |
| {"id": "PTL-0090", "slug": "jailbreak", "term": "Jailbreak", "aliases": ["jailbreaking"], "category": "Security & Adversarial Prompting", "definition": "A jailbreak is a prompt crafted to make a model produce outputs its safety training is meant to prevent, often through role-play, hypothetical framing, obfuscation, or other adversarial techniques.", "description": "Wei et al. attributed jailbreak success to two failure modes, competing objectives between helpfulness and safety, and mismatched generalization, where safety training does not cover inputs the model can still understand. Jailbreaks target the model's safety behavior, while prompt injection targets the application's instructions.", "example": null, "broader": [], "narrower": ["adversarial-suffix", "many-shot-jailbreaking", "crescendo-attack", "pair-attack"], "related": ["prompt-injection", "constitutional-ai"], "introduced": null, "sources": [{"title": "Jailbroken: How Does LLM Safety Training Fail?", "authors": "Wei et al.", "year": 2023, "url": "https://arxiv.org/abs/2307.02483"}], "url": "https://protologue.com/t/jailbreak/", "citation": "Protologue. (2026). Jailbreak. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0090). https://protologue.com/t/jailbreak/"} | |
| {"id": "PTL-0091", "slug": "adversarial-suffix", "term": "Adversarial Suffix", "aliases": ["GCG attack", "universal adversarial attack"], "category": "Security & Adversarial Prompting", "definition": "An adversarial suffix is an automatically optimized string of tokens that, when appended to a request, causes an aligned model to comply with requests it would normally refuse.", "description": "Zou et al. used a greedy coordinate gradient search on open models to find such suffixes and found that they often transferred to other models.", "example": null, "broader": ["jailbreak"], "narrower": [], "related": [], "introduced": 2023, "sources": [{"title": "Universal and Transferable Adversarial Attacks on Aligned Language Models", "authors": "Zou et al.", "year": 2023, "url": "https://arxiv.org/abs/2307.15043"}], "url": "https://protologue.com/t/adversarial-suffix/", "citation": "Protologue. (2026). Adversarial Suffix. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0091). https://protologue.com/t/adversarial-suffix/"} | |
| {"id": "PTL-0092", "slug": "many-shot-jailbreaking", "term": "Many-shot Jailbreaking", "aliases": [], "category": "Security & Adversarial Prompting", "definition": "Many-shot jailbreaking fills a long context window with many fabricated dialogue examples in which an assistant complies with harmful requests, exploiting in-context learning to override the model's safety training.", "description": "Anthropic researchers found the attack's effectiveness followed a power law in the number of shots, mirroring the scaling of benign in-context learning.", "example": null, "broader": ["jailbreak"], "narrower": [], "related": ["many-shot-in-context-learning", "context-window"], "introduced": 2024, "sources": [{"title": "Many-shot jailbreaking", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/research/many-shot-jailbreaking"}], "url": "https://protologue.com/t/many-shot-jailbreaking/", "citation": "Protologue. (2026). Many-shot Jailbreaking. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0092). https://protologue.com/t/many-shot-jailbreaking/"} | |
| {"id": "PTL-0093", "slug": "instruction-hierarchy", "term": "Instruction Hierarchy", "aliases": [], "category": "Security & Adversarial Prompting", "definition": "The instruction hierarchy is a training approach that teaches a model to prioritize instructions by source, typically system over user over tool output, and to ignore lower-priority instructions that conflict with higher-priority ones.", "description": "Wallace et al. trained models on synthetic conflicts and found large gains in robustness to prompt injection and system-prompt extraction with limited loss in helpfulness.", "example": null, "broader": [], "narrower": [], "related": ["system-prompt", "prompt-injection", "spotlighting"], "introduced": 2024, "sources": [{"title": "The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions", "authors": "Wallace et al.", "year": 2024, "url": "https://arxiv.org/abs/2404.13208"}], "url": "https://protologue.com/t/instruction-hierarchy/", "citation": "Protologue. (2026). Instruction Hierarchy. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0093). https://protologue.com/t/instruction-hierarchy/"} | |
| {"id": "PTL-0094", "slug": "spotlighting", "term": "Spotlighting", "aliases": ["datamarking", "input marking"], "category": "Security & Adversarial Prompting", "definition": "Spotlighting is a family of prompt-level defenses against indirect prompt injection that transform untrusted input, by delimiting, marking every word, or encoding it, so the model can distinguish it from trusted instructions.", "description": "Hines et al. reported substantial reductions in attack success with datamarking and encoding variants, with little effect on task performance.", "example": null, "broader": [], "narrower": [], "related": ["indirect-prompt-injection", "delimiters", "instruction-hierarchy"], "introduced": 2024, "sources": [{"title": "Defending Against Indirect Prompt Injection Attacks With Spotlighting", "authors": "Hines et al.", "year": 2024, "url": "https://arxiv.org/abs/2403.14720"}], "url": "https://protologue.com/t/spotlighting/", "citation": "Protologue. (2026). Spotlighting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0094). https://protologue.com/t/spotlighting/"} | |
| {"id": "PTL-0095", "slug": "constitutional-ai", "term": "Constitutional AI", "aliases": ["CAI", "RLAIF", "reinforcement learning from AI feedback"], "category": "Security & Adversarial Prompting", "definition": "Constitutional AI is a training method in which a model critiques and revises its own outputs according to a written set of principles, and AI-generated preference judgments replace most human labels for harmlessness.", "description": "Bai et al. used a supervised self-critique phase followed by reinforcement learning from AI feedback. The approach made the principles governing model behavior explicit and editable.", "example": null, "broader": [], "narrower": [], "related": ["rlhf", "self-refine", "jailbreak"], "introduced": 2022, "sources": [{"title": "Constitutional AI: Harmlessness from AI Feedback", "authors": "Bai et al.", "year": 2022, "url": "https://arxiv.org/abs/2212.08073"}], "url": "https://protologue.com/t/constitutional-ai/", "citation": "Protologue. (2026). Constitutional AI. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0095). https://protologue.com/t/constitutional-ai/"} | |
| {"id": "PTL-0096", "slug": "hallucination", "term": "Hallucination", "aliases": ["confabulation", "fabrication"], "category": "Failure Modes & Evaluation", "definition": "Hallucination is generated content that is fluent and plausible but unfaithful to the provided source or factually incorrect, such as fabricated citations, facts, or quotations.", "description": "Surveys distinguish intrinsic hallucination, which contradicts the source, from extrinsic hallucination, which cannot be verified from it. Mitigations include retrieval-augmented generation, verification methods such as chain-of-verification, allowing the model to say it does not know, and requiring quotes from supplied documents.", "example": null, "broader": [], "narrower": [], "related": ["retrieval-augmented-generation", "chain-of-verification", "sycophancy"], "introduced": null, "sources": [{"title": "Survey of Hallucination in Natural Language Generation", "authors": "Ji et al.", "year": 2022, "url": "https://arxiv.org/abs/2202.03629"}], "url": "https://protologue.com/t/hallucination/", "citation": "Protologue. (2026). Hallucination. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0096). https://protologue.com/t/hallucination/"} | |
| {"id": "PTL-0097", "slug": "sycophancy", "term": "Sycophancy", "aliases": [], "category": "Failure Modes & Evaluation", "definition": "Sycophancy is a model's tendency to tailor its answers to match a user's stated beliefs or preferences, including abandoning correct answers when the user pushes back, rather than giving its most accurate response.", "description": "Sharma et al. found sycophancy across several assistants and linked it to human preference data, which tends to reward agreement. Prompting mitigations include removing opinions from the question, as in System 2 Attention.", "example": null, "broader": [], "narrower": [], "related": ["rlhf", "system-2-attention", "hallucination"], "introduced": null, "sources": [{"title": "Towards Understanding Sycophancy in Language Models", "authors": "Sharma et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.13548"}], "url": "https://protologue.com/t/sycophancy/", "citation": "Protologue. (2026). Sycophancy. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0097). https://protologue.com/t/sycophancy/"} | |
| {"id": "PTL-0098", "slug": "lost-in-the-middle", "term": "Lost in the Middle", "aliases": ["positional bias", "U-shaped context performance"], "category": "Failure Modes & Evaluation", "definition": "Lost in the middle is the finding that language models use information at the beginning or end of a long context much more reliably than information placed in the middle.", "description": "Liu et al. observed a U-shaped performance curve on multi-document question answering as the position of the relevant document varied. A practical takeaway is to place key material and instructions at the start or end of long prompts.", "example": null, "broader": [], "narrower": [], "related": ["context-window", "needle-in-a-haystack", "context-engineering", "thread-of-thought"], "introduced": 2023, "sources": [{"title": "Lost in the Middle: How Language Models Use Long Contexts", "authors": "Liu et al.", "year": 2023, "url": "https://arxiv.org/abs/2307.03172"}], "url": "https://protologue.com/t/lost-in-the-middle/", "citation": "Protologue. (2026). Lost in the Middle. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0098). https://protologue.com/t/lost-in-the-middle/"} | |
| {"id": "PTL-0099", "slug": "needle-in-a-haystack", "term": "Needle in a Haystack", "aliases": ["NIAH", "passkey retrieval"], "category": "Failure Modes & Evaluation", "definition": "Needle in a haystack is a long-context evaluation that hides a specific fact at varying depths in a long distractor document and tests whether the model can retrieve it.", "description": "It became a standard way to report effective context length, though passing it shows only simple retrieval, not reasoning over many dispersed facts.", "example": null, "broader": [], "narrower": [], "related": ["lost-in-the-middle", "context-window"], "introduced": 2023, "sources": [{"title": "needle-in-a-haystack", "authors": "Kamradt", "year": 2023, "url": "https://github.com/gkamradt/needle-in-a-haystack"}], "url": "https://protologue.com/t/needle-in-a-haystack/", "citation": "Protologue. (2026). Needle in a Haystack. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0099). https://protologue.com/t/needle-in-a-haystack/"} | |
| {"id": "PTL-0100", "slug": "prompt-sensitivity", "term": "Prompt Sensitivity", "aliases": ["prompt brittleness", "format sensitivity"], "category": "Failure Modes & Evaluation", "definition": "Prompt sensitivity is the variation in a model's performance caused by superficial changes to a prompt, such as formatting, separators, spacing, or wording, that do not change the task's meaning.", "description": "Sclar et al. found accuracy differences of tens of percentage points from formatting choices alone, and argued that evaluations should report performance across a range of plausible formats.", "example": null, "broader": [], "narrower": [], "related": ["exemplar-ordering", "few-shot-calibration", "emotion-prompting"], "introduced": null, "sources": [{"title": "Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting", "authors": "Sclar et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.11324"}], "url": "https://protologue.com/t/prompt-sensitivity/", "citation": "Protologue. (2026). Prompt Sensitivity. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0100). https://protologue.com/t/prompt-sensitivity/"} | |
| {"id": "PTL-0101", "slug": "unfaithful-chain-of-thought", "term": "Unfaithful Chain-of-Thought", "aliases": ["CoT faithfulness", "post-hoc rationalization"], "category": "Failure Modes & Evaluation", "definition": "Unfaithful chain-of-thought is stated reasoning that does not reflect the factors that actually determined the model's answer, so the explanation can be plausible yet misleading.", "description": "Turpin et al. showed that biasing features, such as always putting the correct answer in position A of few-shot examples, swayed answers while the stated reasoning never mentioned them. Lanham et al. measured how much answers actually depend on the stated reasoning, with results varying by task and model size.", "example": null, "broader": [], "narrower": [], "related": ["chain-of-thought", "reasoning-model"], "introduced": null, "sources": [{"title": "Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting", "authors": "Turpin et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.04388"}, {"title": "Measuring Faithfulness in Chain-of-Thought Reasoning", "authors": "Lanham et al.", "year": 2023, "url": "https://arxiv.org/abs/2307.13702"}], "url": "https://protologue.com/t/unfaithful-chain-of-thought/", "citation": "Protologue. (2026). Unfaithful Chain-of-Thought. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0101). https://protologue.com/t/unfaithful-chain-of-thought/"} | |
| {"id": "PTL-0102", "slug": "large-language-model", "term": "Large Language Model", "aliases": ["LLM", "language model"], "category": "Foundations", "definition": "A large language model (LLM) is a neural network, usually a Transformer with billions of parameters, trained on very large text corpora to predict the next token, which lets it generate text and perform many tasks from a prompt.", "description": "The defining property is that a single model, without task-specific training, can translate, summarize, answer questions, write code, and follow instructions, with the task specified in the prompt. Most deployed LLMs are base models that have been further instruction-tuned and preference-tuned to behave as assistants. Their capabilities and their failure modes, such as hallucination and prompt sensitivity, both follow from next-token prediction over the prompt.", "example": null, "broader": ["foundation-model"], "narrower": ["base-model"], "related": ["token", "prompt", "in-context-learning"], "introduced": null, "sources": [{"title": "A Survey of Large Language Models", "authors": "Zhao et al.", "year": 2023, "url": "https://arxiv.org/abs/2303.18223"}, {"title": "Language Models are Few-Shot Learners", "authors": "Brown et al.", "year": 2020, "url": "https://arxiv.org/abs/2005.14165"}], "url": "https://protologue.com/t/large-language-model/", "citation": "Protologue. (2026). Large Language Model. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0102). https://protologue.com/t/large-language-model/"} | |
| {"id": "PTL-0103", "slug": "foundation-model", "term": "Foundation Model", "aliases": ["frontier model", "general-purpose model"], "category": "Foundations", "definition": "A foundation model is a model trained on broad data at scale, usually with self-supervision, that can be adapted to a wide range of downstream tasks by prompting or fine-tuning.", "description": "The term was introduced in 2021 by researchers at Stanford to describe models such as BERT, GPT-3, and CLIP that serve as a common base for many applications. Large language models are the most prominent kind, but the term also covers image, audio, and multimodal models. Because many systems are built on the same few foundation models, their strengths and flaws are inherited widely.", "example": null, "broader": [], "narrower": ["large-language-model"], "related": ["base-model", "fine-tuning"], "introduced": 2021, "sources": [{"title": "On the Opportunities and Risks of Foundation Models", "authors": "Bommasani et al.", "year": 2021, "url": "https://arxiv.org/abs/2108.07258"}], "url": "https://protologue.com/t/foundation-model/", "citation": "Protologue. (2026). Foundation Model. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0103). https://protologue.com/t/foundation-model/"} | |
| {"id": "PTL-0104", "slug": "base-model", "term": "Base Model", "aliases": ["pretrained model", "base LLM", "completion model"], "category": "Foundations", "definition": "A base model is a language model after pretraining but before instruction tuning or preference tuning, so it continues text in the style of its training data rather than following instructions.", "description": "Asked a question, a base model may answer it, continue with more questions, or produce something unrelated, because it only predicts plausible continuations. Prompting one effectively means framing the task as a document to complete, for example with few-shot examples or the start of the desired output. Chat assistants are derived from base models through instruction tuning and RLHF, which is why they are often called instruct or chat models.", "example": "A completion-style prompt for a base model: \"Q: What is the capital of France?\\nA: Paris\\nQ: What is the capital of Japan?\\nA:\"", "broader": ["large-language-model"], "narrower": [], "related": ["instruction-tuning", "rlhf", "few-shot-prompting", "chat-template"], "introduced": null, "sources": [{"title": "Training language models to follow instructions with human feedback", "authors": "Ouyang et al.", "year": 2022, "url": "https://arxiv.org/abs/2203.02155"}, {"title": "Language Models are Few-Shot Learners", "authors": "Brown et al.", "year": 2020, "url": "https://arxiv.org/abs/2005.14165"}], "url": "https://protologue.com/t/base-model/", "citation": "Protologue. (2026). Base Model. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0104). https://protologue.com/t/base-model/"} | |
| {"id": "PTL-0105", "slug": "fine-tuning", "term": "Fine-Tuning", "aliases": ["model fine-tuning", "transfer learning"], "category": "Foundations", "definition": "Fine-tuning is the further training of a pretrained model on a smaller, task- or domain-specific dataset so that its weights adapt to that task or domain.", "description": "Fine-tuning complements prompting. Prompting changes only the input, while fine-tuning changes the model, which can teach a consistent format, style, or specialized skill that is hard to obtain from instructions alone, at the cost of training data, compute, and a separate model to maintain. Parameter-efficient methods such as LoRA update only a small number of added weights, and instruction tuning is the fine-tuning step that turns a base model into an assistant.", "example": null, "broader": [], "narrower": ["knowledge-distillation"], "related": ["base-model", "instruction-tuning", "low-rank-adaptation", "prompt-tuning"], "introduced": 2018, "sources": [{"title": "Universal Language Model Fine-tuning for Text Classification", "authors": "Howard and Ruder", "year": 2018, "url": "https://arxiv.org/abs/1801.06146"}, {"title": "LoRA: Low-Rank Adaptation of Large Language Models", "authors": "Hu et al.", "year": 2021, "url": "https://arxiv.org/abs/2106.09685"}], "url": "https://protologue.com/t/fine-tuning/", "citation": "Protologue. (2026). Fine-Tuning. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0105). https://protologue.com/t/fine-tuning/"} | |
| {"id": "PTL-0106", "slug": "reward-model", "term": "Reward Model", "aliases": ["preference model", "outcome reward model", "ORM"], "category": "Foundations", "definition": "A reward model is a model trained to score a language model's outputs according to human or AI preferences, usually from pairwise comparisons, so that the score can serve as a training signal or a ranking criterion.", "description": "In RLHF, people compare pairs of responses, a reward model learns to predict which one they prefer, and the language model is then optimized against it. Reward models are also used at inference time to pick the best of several sampled responses. A model that scores only the final answer is an outcome reward model, in contrast to a process reward model, which scores each reasoning step. Optimizing too hard against an imperfect reward model leads to reward hacking.", "example": null, "broader": [], "narrower": [], "related": ["rlhf", "process-reward-model", "best-of-n-sampling", "reward-hacking"], "introduced": 2022, "sources": [{"title": "Training language models to follow instructions with human feedback", "authors": "Ouyang et al.", "year": 2022, "url": "https://arxiv.org/abs/2203.02155"}, {"title": "Let's Verify Step by Step", "authors": "Lightman et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.20050"}], "url": "https://protologue.com/t/reward-model/", "citation": "Protologue. (2026). Reward Model. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0106). https://protologue.com/t/reward-model/"} | |
| {"id": "PTL-0107", "slug": "chat-template", "term": "Chat Template", "aliases": ["prompt format", "chat format", "conversation template"], "category": "Foundations", "definition": "A chat template is the model-specific format that converts a list of role-tagged messages, such as system, user, and assistant turns, into the exact token sequence with special control tokens that the model was fine-tuned on.", "description": "Hosted chat APIs apply the template for you, but when running open-weight models directly, using the wrong template, or none, can noticeably degrade instruction following because the model sees input unlike its training data. Templates also determine how system prompts, tool calls, and tool results are marked, and where the model expects to begin its reply.", "example": null, "broader": [], "narrower": [], "related": ["system-prompt", "base-model", "instruction-tuning", "delimiters"], "introduced": null, "sources": [{"title": "Chat templates", "authors": "Hugging Face", "year": 2023, "url": "https://huggingface.co/docs/transformers/en/chat_templating"}], "url": "https://protologue.com/t/chat-template/", "citation": "Protologue. (2026). Chat Template. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0107). https://protologue.com/t/chat-template/"} | |
| {"id": "PTL-0108", "slug": "top-k-sampling", "term": "Top-k Sampling", "aliases": ["top-k"], "category": "Foundations", "definition": "Top-k sampling is a decoding method that, at each step, restricts sampling to the k most probable next tokens and renormalizes their probabilities.", "description": "It removes the long tail of unlikely tokens that can derail generation, while keeping some variety. Because k is fixed, it can be too restrictive when many tokens are plausible and too permissive when only one is, which motivated top-p (nucleus) sampling. Many APIs expose top-k alongside temperature and top-p.", "example": null, "broader": [], "narrower": [], "related": ["temperature", "top-p-sampling", "greedy-decoding"], "introduced": 2018, "sources": [{"title": "Hierarchical Neural Story Generation", "authors": "Fan et al.", "year": 2018, "url": "https://arxiv.org/abs/1805.04833"}, {"title": "The Curious Case of Neural Text Degeneration", "authors": "Holtzman et al.", "year": 2019, "url": "https://arxiv.org/abs/1904.09751"}], "url": "https://protologue.com/t/top-k-sampling/", "citation": "Protologue. (2026). Top-k Sampling. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0108). https://protologue.com/t/top-k-sampling/"} | |
| {"id": "PTL-0109", "slug": "greedy-decoding", "term": "Greedy Decoding", "aliases": ["argmax decoding", "temperature 0"], "category": "Foundations", "definition": "Greedy decoding generates text by always choosing the single most probable next token, which makes output deterministic for a given prompt and model.", "description": "It is the limit of sampling as temperature approaches zero and is often used when consistency matters, as in extraction or classification. For open-ended text it tends to produce bland or repetitive output, and for reasoning tasks sampling several answers and voting, as in self-consistency, usually beats a single greedy answer.", "example": null, "broader": [], "narrower": [], "related": ["temperature", "top-k-sampling", "top-p-sampling", "self-consistency", "text-degeneration"], "introduced": null, "sources": [{"title": "The Curious Case of Neural Text Degeneration", "authors": "Holtzman et al.", "year": 2019, "url": "https://arxiv.org/abs/1904.09751"}], "url": "https://protologue.com/t/greedy-decoding/", "citation": "Protologue. (2026). Greedy Decoding. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0109). https://protologue.com/t/greedy-decoding/"} | |
| {"id": "PTL-0110", "slug": "log-probabilities", "term": "Log Probabilities", "aliases": ["logprobs", "token probabilities"], "category": "Foundations", "definition": "Log probabilities (logprobs) are the logarithms of the probabilities a model assigned to each token it generated, and optionally to the most likely alternatives, exposing how confident the model was at each step.", "description": "Applications use logprobs to score classification labels, estimate confidence, flag likely hallucinations, rank candidate completions, and build autocomplete. They describe token-level likelihood, not factual correctness, so a fluent but wrong answer can still have high probability. Not every API or model exposes them.", "example": null, "broader": [], "narrower": [], "related": ["token", "confidence-calibration", "temperature"], "introduced": null, "sources": [{"title": "Using logprobs", "authors": "OpenAI", "year": 2023, "url": "https://developers.openai.com/cookbook/examples/using_logprobs"}, {"title": "Language Models (Mostly) Know What They Know", "authors": "Kadavath et al.", "year": 2022, "url": "https://arxiv.org/abs/2207.05221"}], "url": "https://protologue.com/t/log-probabilities/", "citation": "Protologue. (2026). Log Probabilities. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0110). https://protologue.com/t/log-probabilities/"} | |
| {"id": "PTL-0111", "slug": "multimodal-prompting", "term": "Multimodal Prompting", "aliases": ["vision prompting", "image prompting", "visual prompting"], "category": "Foundations", "definition": "Multimodal prompting is prompting a model with more than one kind of input, such as images, documents, audio, or video together with text, so the model reasons over all of them.", "description": "Text-prompting techniques largely carry over, but some practices are specific to non-text inputs. Provider guidance recommends placing images before the question, labeling multiple images so they can be referred to, and sending legible, uncompressed images. Models remain weaker at precise counting, small text, and spatial coordinates, so high-stakes visual answers need checking.", "example": "Image 1: [receipt photo]. Image 2: [receipt photo]. List the total on each receipt and say which is larger.", "broader": ["prompt"], "narrower": [], "related": ["context-window", "large-language-model"], "introduced": 2023, "sources": [{"title": "The Dawn of LMMs: Preliminary Explorations with GPT-4V(ision)", "authors": "Yang et al.", "year": 2023, "url": "https://arxiv.org/abs/2309.17421"}, {"title": "Vision", "authors": "Anthropic", "year": 2024, "url": "https://platform.claude.com/docs/en/build-with-claude/vision"}], "url": "https://protologue.com/t/multimodal-prompting/", "citation": "Protologue. (2026). Multimodal Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0111). https://protologue.com/t/multimodal-prompting/"} | |
| {"id": "PTL-0112", "slug": "knowledge-cutoff", "term": "Knowledge Cutoff", "aliases": ["training cutoff", "training data cutoff"], "category": "Foundations", "definition": "A model's knowledge cutoff is the date after which its training data contains little or no information, so it does not know about later events unless they are supplied in the prompt or retrieved.", "description": "Questions about recent events, prices, or current officeholders are answered from stale knowledge unless the system adds search or retrieval. Research has found that a model's effective cutoff can differ from its reported one, and can differ across topics, because web data from earlier years is overrepresented and deduplication is imperfect. Stating the current date in the system prompt helps the model reason about what it might not know.", "example": null, "broader": [], "narrower": [], "related": ["retrieval-augmented-generation", "web-search-tool", "hallucination"], "introduced": null, "sources": [{"title": "Dated Data: Tracing Knowledge Cutoffs in Large Language Models", "authors": "Cheng et al.", "year": 2024, "url": "https://arxiv.org/abs/2403.12958"}], "url": "https://protologue.com/t/knowledge-cutoff/", "citation": "Protologue. (2026). Knowledge Cutoff. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0112). https://protologue.com/t/knowledge-cutoff/"} | |
| {"id": "PTL-0113", "slug": "instruction-following", "term": "Instruction Following", "aliases": ["instruction adherence", "instruction compliance"], "category": "Foundations", "definition": "Instruction following is a model's ability to carry out explicit instructions in a prompt, including constraints on content, length, format, and style, rather than only answering the underlying question.", "description": "It is the main capability added by instruction tuning and RLHF, and it degrades with long or conflicting instruction lists, long contexts, and multi-turn conversations. Benchmarks such as IFEval measure it with instructions whose compliance can be checked automatically, for example a minimum word count or a required keyword.", "example": "Answer in exactly three bullet points, each under 15 words, and do not use the word \"very\".", "broader": [], "narrower": [], "related": ["instruction-tuning", "rlhf", "system-prompt", "instruction-hierarchy"], "introduced": null, "sources": [{"title": "Instruction-Following Evaluation for Large Language Models", "authors": "Zhou et al.", "year": 2023, "url": "https://arxiv.org/abs/2311.07911"}, {"title": "Training language models to follow instructions with human feedback", "authors": "Ouyang et al.", "year": 2022, "url": "https://arxiv.org/abs/2203.02155"}], "url": "https://protologue.com/t/instruction-following/", "citation": "Protologue. (2026). Instruction Following. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0113). https://protologue.com/t/instruction-following/"} | |
| {"id": "PTL-0114", "slug": "decomposed-prompting", "term": "Decomposed Prompting", "aliases": ["DecomP", "modular prompting"], "category": "Reasoning Elicitation", "definition": "Decomposed prompting breaks a complex task into simpler sub-tasks, each delegated to its own prompt, model call, or symbolic function, with a controller prompt that decides which handler to call next.", "description": "Each sub-task handler can be optimized, given its own examples, or replaced with code, and handlers can call further decompositions recursively. Compared with a single chain-of-thought prompt, this modularity helps on tasks with steps that models do poorly in one pass, such as long-input processing or symbolic manipulation.", "example": null, "broader": [], "narrower": [], "related": ["least-to-most-prompting", "prompt-chaining", "orchestrator-workers", "program-aided-language-models"], "introduced": 2022, "sources": [{"title": "Decomposed Prompting: A Modular Approach for Solving Complex Tasks", "authors": "Khot et al.", "year": 2022, "url": "https://arxiv.org/abs/2210.02406"}], "url": "https://protologue.com/t/decomposed-prompting/", "citation": "Protologue. (2026). Decomposed Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0114). https://protologue.com/t/decomposed-prompting/"} | |
| {"id": "PTL-0115", "slug": "complexity-based-prompting", "term": "Complexity-Based Prompting", "aliases": [], "category": "Reasoning Elicitation", "definition": "Complexity-based prompting selects chain-of-thought exemplars that contain more reasoning steps, and at inference votes only among the sampled reasoning chains with the most steps.", "description": "The method rests on the finding that demonstrations with longer, more complex reasoning improve multi-step reasoning more than simpler ones. It combines exemplar selection with a variant of self-consistency in which votes from the most complex chains count.", "example": null, "broader": ["chain-of-thought"], "narrower": [], "related": ["exemplar-selection", "self-consistency", "auto-cot"], "introduced": 2022, "sources": [{"title": "Complexity-Based Prompting for Multi-Step Reasoning", "authors": "Fu et al.", "year": 2022, "url": "https://arxiv.org/abs/2210.00720"}], "url": "https://protologue.com/t/complexity-based-prompting/", "citation": "Protologue. (2026). Complexity-Based Prompting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0115). https://protologue.com/t/complexity-based-prompting/"} | |
| {"id": "PTL-0116", "slug": "chain-of-draft", "term": "Chain of Draft", "aliases": ["CoD", "concise chain-of-thought"], "category": "Reasoning Elicitation", "definition": "Chain of Draft is a variant of chain-of-thought prompting in which the model writes only a minimal draft of each reasoning step, a few words rather than full sentences, to reach an answer with far fewer tokens.", "description": "The authors report accuracy comparable to standard chain-of-thought on arithmetic, commonsense, and symbolic tasks while using a small fraction of the output tokens, which lowers cost and latency. It is most useful where reasoning helps but verbose explanations are not needed.", "example": "Think step by step, but keep only a minimum draft for each step, five words at most. Give the answer after ####.", "broader": ["chain-of-thought"], "narrower": [], "related": ["prompt-compression", "skeleton-of-thought", "overthinking"], "introduced": 2025, "sources": [{"title": "Chain of Draft: Thinking Faster by Writing Less", "authors": "Xu et al.", "year": 2025, "url": "https://arxiv.org/abs/2502.18600"}], "url": "https://protologue.com/t/chain-of-draft/", "citation": "Protologue. (2026). Chain of Draft. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0116). https://protologue.com/t/chain-of-draft/"} | |
| {"id": "PTL-0117", "slug": "self-correction", "term": "Self-Correction", "aliases": ["intrinsic self-correction", "self-revision"], "category": "Self-Critique & Verification", "definition": "Self-correction is the general practice of having a language model review and revise its own output, either from its own judgment alone (intrinsic self-correction) or using external feedback such as tool results, tests, or retrieved evidence.", "description": "Research has found that intrinsic self-correction does not reliably improve reasoning accuracy and can make answers worse, because the model often cannot tell which of its answers are wrong. Self-correction works best when the critique is grounded in something the model cannot fake, such as running code, executing a search, or checking against a reference.", "example": null, "broader": [], "narrower": ["critic"], "related": ["self-refine", "reflexion", "chain-of-verification", "evaluator-optimizer"], "introduced": 2023, "sources": [{"title": "Large Language Models Cannot Self-Correct Reasoning Yet", "authors": "Huang et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.01798"}], "url": "https://protologue.com/t/self-correction/", "citation": "Protologue. (2026). Self-Correction. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0117). https://protologue.com/t/self-correction/"} | |
| {"id": "PTL-0118", "slug": "self-verification", "term": "Self-Verification", "aliases": ["backward verification"], "category": "Self-Critique & Verification", "definition": "Self-verification is a technique in which a model, after producing candidate answers with chain-of-thought, checks each one by working backward, predicting the original conditions of the problem from the answer, and keeps the answer that is most consistent.", "description": "It turns verification into a separate prediction task rather than asking the model whether it is right, which reduces the tendency to approve its own mistakes. It applies most naturally to arithmetic and logical problems whose conditions can be masked and recovered.", "example": null, "broader": [], "narrower": [], "related": ["self-consistency", "chain-of-verification", "self-correction"], "introduced": 2022, "sources": [{"title": "Large Language Models are Better Reasoners with Self-Verification", "authors": "Weng et al.", "year": 2022, "url": "https://arxiv.org/abs/2212.09561"}], "url": "https://protologue.com/t/self-verification/", "citation": "Protologue. (2026). Self-Verification. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0118). https://protologue.com/t/self-verification/"} | |
| {"id": "PTL-0119", "slug": "critic", "term": "CRITIC", "aliases": ["tool-interactive critiquing"], "category": "Self-Critique & Verification", "definition": "CRITIC is a self-correction framework in which a model checks its output using external tools, such as web search or a code interpreter, critiques it based on what the tools return, and then revises it.", "description": "Grounding the critique in tool feedback addresses the main weakness of intrinsic self-correction. The original work showed improvements on question answering, program synthesis, and reducing toxic output, and the same verify-then-revise loop is common in agent systems.", "example": null, "broader": ["self-correction"], "narrower": [], "related": ["self-refine", "function-calling", "code-execution-tool", "chain-of-verification"], "introduced": 2023, "sources": [{"title": "CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing", "authors": "Gou et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.11738"}], "url": "https://protologue.com/t/critic/", "citation": "Protologue. (2026). CRITIC. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0119). https://protologue.com/t/critic/"} | |
| {"id": "PTL-0120", "slug": "chain-of-density", "term": "Chain of Density", "aliases": ["CoD summarization"], "category": "Self-Critique & Verification", "definition": "Chain of Density is a summarization prompt that has the model write a summary and then repeatedly rewrite it, adding a few missing important entities each time without making it longer, to produce progressively denser summaries.", "description": "Each round forces compression and fusion of content, so later summaries carry more information per word. The authors found human readers preferred summaries partway through the chain, denser than a plain summary but not maximally dense, which makes the number of rounds a tunable trade-off between informativeness and readability.", "example": null, "broader": ["self-refine"], "narrower": [], "related": ["prompt-chaining"], "introduced": 2023, "sources": [{"title": "From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting", "authors": "Adams et al.", "year": 2023, "url": "https://arxiv.org/abs/2309.04269"}], "url": "https://protologue.com/t/chain-of-density/", "citation": "Protologue. (2026). Chain of Density. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0120). https://protologue.com/t/chain-of-density/"} | |
| {"id": "PTL-0121", "slug": "confidence-calibration", "term": "Confidence Calibration", "aliases": ["calibration", "verbalized confidence", "uncertainty estimation"], "category": "Self-Critique & Verification", "definition": "Confidence calibration is the degree to which a model's expressed or implied confidence matches how often it is actually right; a well-calibrated model that says it is 80% sure is correct about 80% of the time.", "description": "Confidence can be read from token probabilities, by asking the model to state a confidence (verbalized confidence), or by sampling several answers and measuring agreement. Studies have found pretrained models fairly well calibrated on suitably formatted multiple-choice questions, and that for instruction-tuned models verbalized confidence can be better calibrated than token probabilities. Calibration lets an application abstain or escalate when the model is unsure. It is different from few-shot calibration, which corrects label bias in prompts.", "example": null, "broader": [], "narrower": [], "related": ["log-probabilities", "hallucination", "selfcheckgpt", "few-shot-calibration"], "introduced": null, "sources": [{"title": "Language Models (Mostly) Know What They Know", "authors": "Kadavath et al.", "year": 2022, "url": "https://arxiv.org/abs/2207.05221"}, {"title": "Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback", "authors": "Tian et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.14975"}], "url": "https://protologue.com/t/confidence-calibration/", "citation": "Protologue. (2026). Confidence Calibration. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0121). https://protologue.com/t/confidence-calibration/"} | |
| {"id": "PTL-0122", "slug": "selfcheckgpt", "term": "SelfCheckGPT", "aliases": ["sampling-based hallucination detection", "consistency-based hallucination detection"], "category": "Self-Critique & Verification", "definition": "SelfCheckGPT detects hallucinations by sampling several responses to the same prompt and checking whether each claim in the main response is supported by the others, on the principle that facts the model knows are reproduced consistently while fabrications vary.", "description": "It needs no external knowledge base and works with black-box models, which makes it a practical first check. Its limitation is that consistent errors, which the model repeats confidently every time, are not caught, so it complements rather than replaces retrieval-based fact-checking.", "example": null, "broader": [], "narrower": [], "related": ["hallucination", "self-consistency", "confidence-calibration", "chain-of-verification"], "introduced": 2023, "sources": [{"title": "SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models", "authors": "Manakul et al.", "year": 2023, "url": "https://arxiv.org/abs/2303.08896"}], "url": "https://protologue.com/t/selfcheckgpt/", "citation": "Protologue. (2026). SelfCheckGPT. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0122). https://protologue.com/t/selfcheckgpt/"} | |
| {"id": "PTL-0123", "slug": "embedding", "term": "Embedding", "aliases": ["text embedding", "vector embedding", "embedding vector"], "category": "Retrieval & Tool Use", "definition": "An embedding is a fixed-length vector of numbers that represents a piece of text, or another input, so that inputs with similar meaning have vectors that are close together, typically measured by cosine similarity.", "description": "Embeddings are produced by an embedding model, which is usually separate from the generating model. They underpin semantic search, dense retrieval for RAG, clustering, deduplication, and the selection of similar few-shot examples. Quality depends on the embedding model and on how text is chunked, and embeddings capture topical similarity better than exact facts such as numbers, names, or negation.", "example": null, "broader": [], "narrower": [], "related": ["dense-retrieval", "vector-database", "chunking", "retrieval-augmented-generation", "exemplar-selection"], "introduced": 2019, "sources": [{"title": "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks", "authors": "Reimers and Gurevych", "year": 2019, "url": "https://arxiv.org/abs/1908.10084"}, {"title": "Dense Passage Retrieval for Open-Domain Question Answering", "authors": "Karpukhin et al.", "year": 2020, "url": "https://arxiv.org/abs/2004.04906"}], "url": "https://protologue.com/t/embedding/", "citation": "Protologue. (2026). Embedding. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0123). https://protologue.com/t/embedding/"} | |
| {"id": "PTL-0124", "slug": "vector-database", "term": "Vector Database", "aliases": ["vector store", "vector index", "vector search"], "category": "Retrieval & Tool Use", "definition": "A vector database stores embeddings alongside their source content and metadata and answers queries by finding the stored vectors nearest to a query vector, usually with approximate nearest-neighbor search.", "description": "Exact nearest-neighbor search becomes slow at millions of vectors, so vector databases use index structures such as HNSW graphs or quantized inverted indexes that trade a little recall for large speedups. Most also support metadata filtering and, increasingly, keyword search for hybrid retrieval. A vector index library such as FAISS provides the search without the database features.", "example": null, "broader": [], "narrower": [], "related": ["embedding", "dense-retrieval", "hybrid-search", "retrieval-augmented-generation"], "introduced": null, "sources": [{"title": "Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs", "authors": "Malkov and Yashunin", "year": 2016, "url": "https://arxiv.org/abs/1603.09320"}, {"title": "Billion-scale similarity search with GPUs", "authors": "Johnson et al.", "year": 2017, "url": "https://arxiv.org/abs/1702.08734"}], "url": "https://protologue.com/t/vector-database/", "citation": "Protologue. (2026). Vector Database. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0124). https://protologue.com/t/vector-database/"} | |
| {"id": "PTL-0125", "slug": "chunking", "term": "Chunking", "aliases": ["document chunking", "text splitting", "passage segmentation", "late chunking"], "category": "Retrieval & Tool Use", "definition": "Chunking is splitting documents into smaller passages before indexing them for retrieval, so that each retrieved unit is small enough to fit the context and focused enough to match a query.", "description": "Chunk size and overlap are the main levers. Chunks that are too large dilute the embedding and waste context, while chunks that are too small lose the surrounding meaning. Common strategies split by fixed token counts, by document structure such as headings and paragraphs, or by meaning. Research has found that retrieving finer units, such as self-contained propositions, can improve retrieval, and approaches such as contextual retrieval and late chunking restore context that splitting removes.", "example": null, "broader": [], "narrower": [], "related": ["embedding", "contextual-retrieval", "retrieval-augmented-generation", "dense-retrieval"], "introduced": null, "sources": [{"title": "Dense X Retrieval: What Retrieval Granularity Should We Use?", "authors": "Chen et al.", "year": 2023, "url": "https://arxiv.org/abs/2312.06648"}, {"title": "Late Chunking: Contextual Chunk Embeddings Using Long-Context Embedding Models", "authors": "Günther et al.", "year": 2024, "url": "https://arxiv.org/abs/2409.04701"}], "url": "https://protologue.com/t/chunking/", "citation": "Protologue. (2026). Chunking. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0125). https://protologue.com/t/chunking/"} | |
| {"id": "PTL-0126", "slug": "dense-retrieval", "term": "Dense Retrieval", "aliases": ["semantic search", "embedding-based retrieval", "DPR", "bi-encoder retrieval"], "category": "Retrieval & Tool Use", "definition": "Dense retrieval finds relevant passages by embedding the query and every passage as dense vectors, usually with a dual encoder, and returning the passages whose vectors are most similar to the query's.", "description": "Because it matches on meaning rather than shared words, dense retrieval finds paraphrases and related concepts that keyword search misses. It can miss exact identifiers, rare names, and codes, which is why production systems often combine it with BM25 in hybrid search and refine results with a reranker.", "example": null, "broader": [], "narrower": [], "related": ["embedding", "bm25", "hybrid-search", "reranking", "vector-database"], "introduced": 2020, "sources": [{"title": "Dense Passage Retrieval for Open-Domain Question Answering", "authors": "Karpukhin et al.", "year": 2020, "url": "https://arxiv.org/abs/2004.04906"}], "url": "https://protologue.com/t/dense-retrieval/", "citation": "Protologue. (2026). Dense Retrieval. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0126). https://protologue.com/t/dense-retrieval/"} | |
| {"id": "PTL-0127", "slug": "bm25", "term": "BM25", "aliases": ["Okapi BM25", "sparse retrieval", "keyword search", "lexical search"], "category": "Retrieval & Tool Use", "definition": "BM25 is a keyword ranking function that scores documents by how often the query's terms appear in them, weighted by how rare each term is across the collection and normalized for document length.", "description": "It is the standard sparse, lexical retrieval method and remains a strong baseline. It excels at exact matches, such as product codes, error messages, and names, where embedding search is weaker, and needs no training. It does not match synonyms or paraphrases, so it is usually paired with dense retrieval.", "example": null, "broader": [], "narrower": [], "related": ["dense-retrieval", "hybrid-search", "contextual-retrieval"], "introduced": null, "sources": [{"title": "The Probabilistic Relevance Framework: BM25 and Beyond", "authors": "Robertson and Zaragoza", "year": 2009, "url": "https://doi.org/10.1561/1500000019"}], "url": "https://protologue.com/t/bm25/", "citation": "Protologue. (2026). BM25. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0127). https://protologue.com/t/bm25/"} | |
| {"id": "PTL-0128", "slug": "hybrid-search", "term": "Hybrid Search", "aliases": ["hybrid retrieval", "reciprocal rank fusion", "RRF"], "category": "Retrieval & Tool Use", "definition": "Hybrid search combines keyword retrieval, such as BM25, with dense embedding retrieval and merges the two result lists, so that a query benefits from both exact-term matching and semantic matching.", "description": "The lists are usually merged by reciprocal rank fusion, which scores each document by the sum of the reciprocals of its ranks in each list and needs no score calibration between systems. Hybrid retrieval followed by a reranker is a common default for RAG pipelines.", "example": null, "broader": [], "narrower": [], "related": ["bm25", "dense-retrieval", "reranking", "contextual-retrieval", "retrieval-augmented-generation"], "introduced": null, "sources": [{"title": "Reciprocal rank fusion outperforms condorcet and individual rank learning methods", "authors": "Cormack et al.", "year": 2009, "url": "https://doi.org/10.1145/1571941.1572114"}, {"title": "Introducing Contextual Retrieval", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/news/contextual-retrieval"}], "url": "https://protologue.com/t/hybrid-search/", "citation": "Protologue. (2026). Hybrid Search. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0128). https://protologue.com/t/hybrid-search/"} | |
| {"id": "PTL-0129", "slug": "reranking", "term": "Reranking", "aliases": ["re-ranking", "cross-encoder reranking", "reranker"], "category": "Retrieval & Tool Use", "definition": "Reranking is a second retrieval stage in which a more accurate but slower model, typically a cross-encoder that reads the query and each candidate together, rescores the top results from a fast first-stage search and reorders them.", "description": "First-stage retrievers embed queries and documents separately to stay fast, which limits accuracy. A reranker sees both texts jointly, so it judges relevance more precisely, and applying it only to a few dozen candidates keeps cost manageable. Rerankers can also be LLMs prompted to judge or order passages.", "example": null, "broader": [], "narrower": [], "related": ["dense-retrieval", "hybrid-search", "colbert", "retrieval-augmented-generation"], "introduced": 2019, "sources": [{"title": "Passage Re-ranking with BERT", "authors": "Nogueira and Cho", "year": 2019, "url": "https://arxiv.org/abs/1901.04085"}], "url": "https://protologue.com/t/reranking/", "citation": "Protologue. (2026). Reranking. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0129). https://protologue.com/t/reranking/"} | |
| {"id": "PTL-0130", "slug": "colbert", "term": "ColBERT", "aliases": ["late interaction", "late-interaction retrieval", "multi-vector retrieval"], "category": "Retrieval & Tool Use", "definition": "ColBERT is a retrieval model that embeds every token of the query and of each passage separately and scores relevance by matching each query token to its most similar passage token, an approach called late interaction.", "description": "Late interaction sits between single-vector dense retrieval, which is fast but coarse, and cross-encoder reranking, which is accurate but slow. Passage token embeddings can be precomputed and indexed, so ColBERT keeps much of a cross-encoder's precision at search-engine speed, at the cost of larger indexes.", "example": null, "broader": [], "narrower": [], "related": ["dense-retrieval", "reranking", "dspy"], "introduced": 2020, "sources": [{"title": "ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT", "authors": "Khattab and Zaharia", "year": 2020, "url": "https://arxiv.org/abs/2004.12832"}], "url": "https://protologue.com/t/colbert/", "citation": "Protologue. (2026). ColBERT. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0130). https://protologue.com/t/colbert/"} | |
| {"id": "PTL-0131", "slug": "query-rewriting", "term": "Query Rewriting", "aliases": ["query expansion", "query reformulation", "rewrite-retrieve-read", "multi-query retrieval"], "category": "Retrieval & Tool Use", "definition": "Query rewriting uses a language model to reformulate a user's question into one or more queries better suited to the retriever before retrieval is run.", "description": "User questions are often ambiguous, conversational, or phrased differently from the documents that answer them. Rewriting can resolve references to earlier turns, split a question into sub-queries, add keywords, or expand the query with a generated pseudo-document. It is closely related to HyDE, which embeds a hypothetical answer instead of the question.", "example": null, "broader": [], "narrower": [], "related": ["hypothetical-document-embeddings", "retrieval-augmented-generation", "agentic-rag"], "introduced": 2023, "sources": [{"title": "Query Rewriting for Retrieval-Augmented Large Language Models", "authors": "Ma et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.14283"}, {"title": "Query2doc: Query Expansion with Large Language Models", "authors": "Wang et al.", "year": 2023, "url": "https://arxiv.org/abs/2303.07678"}], "url": "https://protologue.com/t/query-rewriting/", "citation": "Protologue. (2026). Query Rewriting. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0131). https://protologue.com/t/query-rewriting/"} | |
| {"id": "PTL-0132", "slug": "contextual-retrieval", "term": "Contextual Retrieval", "aliases": ["contextual embeddings", "contextual BM25", "chunk contextualization"], "category": "Retrieval & Tool Use", "definition": "Contextual retrieval is a RAG preprocessing technique in which a language model writes a short explanation of where each chunk fits in its source document and prepends it to the chunk before the chunk is embedded and indexed for keyword search.", "description": "Splitting removes context, so a chunk such as \"revenue grew 3% over the previous quarter\" no longer says which company or quarter it describes. Adding that context back improves both embedding and BM25 matching. Anthropic reported that contextual embeddings plus contextual BM25 cut the top-20 retrieval failure rate by 49% on its benchmarks, and by 67% with reranking added. Prompt caching keeps the cost of generating the context for every chunk low.", "example": null, "broader": [], "narrower": [], "related": ["chunking", "bm25", "hybrid-search", "reranking", "prompt-caching"], "introduced": 2024, "sources": [{"title": "Introducing Contextual Retrieval", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/news/contextual-retrieval"}], "url": "https://protologue.com/t/contextual-retrieval/", "citation": "Protologue. (2026). Contextual Retrieval. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0132). https://protologue.com/t/contextual-retrieval/"} | |
| {"id": "PTL-0133", "slug": "graphrag", "term": "GraphRAG", "aliases": ["Graph RAG", "knowledge-graph RAG"], "category": "Retrieval & Tool Use", "definition": "GraphRAG is a retrieval-augmented generation approach that uses a language model to extract an entity-and-relationship graph from a document collection, summarizes communities of related entities, and answers questions from those summaries.", "description": "Standard RAG retrieves a handful of passages, which works for questions answered in one place but fails on global questions about a whole corpus, such as its main themes. GraphRAG answers these by combining partial answers from many community summaries. Building the graph requires many LLM calls, so indexing is far more expensive than for embedding-based RAG.", "example": null, "broader": ["retrieval-augmented-generation"], "narrower": [], "related": ["agentic-rag"], "introduced": 2024, "sources": [{"title": "From Local to Global: A Graph RAG Approach to Query-Focused Summarization", "authors": "Edge et al.", "year": 2024, "url": "https://arxiv.org/abs/2404.16130"}], "url": "https://protologue.com/t/graphrag/", "citation": "Protologue. (2026). GraphRAG. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0133). https://protologue.com/t/graphrag/"} | |
| {"id": "PTL-0134", "slug": "agentic-rag", "term": "Agentic RAG", "aliases": ["agent-based RAG", "iterative retrieval"], "category": "Retrieval & Tool Use", "definition": "Agentic RAG is retrieval-augmented generation in which an agent, rather than a fixed pipeline, decides whether to retrieve, what to search for, which sources or tools to use, and whether the evidence is sufficient, retrieving iteratively until it can answer.", "description": "A fixed pipeline retrieves once per question, so it fails on multi-hop questions and on queries that need reformulation. Agentic RAG handles them by planning, issuing follow-up searches, and reflecting on results, at the cost of more model calls, higher latency, and less predictable behavior. Self-RAG, corrective RAG, and active retrieval are earlier steps in the same direction.", "example": null, "broader": ["retrieval-augmented-generation"], "narrower": [], "related": ["ai-agent", "react", "query-rewriting", "self-rag", "corrective-rag"], "introduced": 2025, "sources": [{"title": "Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG", "authors": "Singh et al.", "year": 2025, "url": "https://arxiv.org/abs/2501.09136"}], "url": "https://protologue.com/t/agentic-rag/", "citation": "Protologue. (2026). Agentic RAG. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0134). https://protologue.com/t/agentic-rag/"} | |
| {"id": "PTL-0135", "slug": "corrective-rag", "term": "Corrective RAG", "aliases": ["CRAG"], "category": "Retrieval & Tool Use", "definition": "Corrective RAG (CRAG) adds a lightweight evaluator that grades retrieved documents for relevance and, depending on the grade, uses them, discards them in favor of web search, or combines both before generation.", "description": "It targets the failure where poor retrieval silently produces a confident wrong answer. Retrieved documents are also decomposed into smaller strips and filtered so that only relevant parts reach the generator. The method plugs into existing RAG pipelines without retraining the generator.", "example": null, "broader": ["retrieval-augmented-generation"], "narrower": [], "related": ["self-rag", "agentic-rag", "web-search-tool"], "introduced": 2024, "sources": [{"title": "Corrective Retrieval Augmented Generation", "authors": "Yan et al.", "year": 2024, "url": "https://arxiv.org/abs/2401.15884"}], "url": "https://protologue.com/t/corrective-rag/", "citation": "Protologue. (2026). Corrective RAG. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0135). https://protologue.com/t/corrective-rag/"} | |
| {"id": "PTL-0136", "slug": "active-retrieval", "term": "Active Retrieval", "aliases": ["FLARE", "forward-looking active retrieval"], "category": "Retrieval & Tool Use", "definition": "Active retrieval augmented generation decides during generation when and what to retrieve; in FLARE, the model drafts the next sentence and, if it contains low-confidence tokens, uses that draft as a search query and regenerates the sentence with the retrieved evidence.", "description": "Retrieving once before generation suits short answers but not long-form text whose information needs change as it is written. Triggering retrieval only on uncertainty limits extra calls to the points where the model is likely to be wrong.", "example": null, "broader": ["retrieval-augmented-generation"], "narrower": [], "related": ["log-probabilities", "self-rag", "agentic-rag"], "introduced": 2023, "sources": [{"title": "Active Retrieval Augmented Generation", "authors": "Jiang et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.06983"}], "url": "https://protologue.com/t/active-retrieval/", "citation": "Protologue. (2026). Active Retrieval. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0136). https://protologue.com/t/active-retrieval/"} | |
| {"id": "PTL-0137", "slug": "web-search-tool", "term": "Web Search Tool", "aliases": ["web browsing", "search grounding", "grounding with search", "browsing tool"], "category": "Retrieval & Tool Use", "definition": "A web search tool lets a language model issue search queries, read the results, and ground its answer in current web content, usually returning citations to the pages it used.", "description": "Search is the main remedy for a model's knowledge cutoff and for questions about prices, news, and other changing facts. Early work trained a model to browse with a text-based browser and cite its sources. Today providers offer search as a built-in tool. Search results are untrusted input, so they are a vector for indirect prompt injection.", "example": null, "broader": ["function-calling"], "narrower": [], "related": ["knowledge-cutoff", "citation-generation", "retrieval-augmented-generation", "indirect-prompt-injection"], "introduced": 2021, "sources": [{"title": "WebGPT: Browser-assisted question-answering with human feedback", "authors": "Nakano et al.", "year": 2021, "url": "https://arxiv.org/abs/2112.09332"}, {"title": "Web search tool", "authors": "Anthropic", "year": 2025, "url": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool"}], "url": "https://protologue.com/t/web-search-tool/", "citation": "Protologue. (2026). Web Search Tool. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0137). https://protologue.com/t/web-search-tool/"} | |
| {"id": "PTL-0138", "slug": "code-execution-tool", "term": "Code Execution Tool", "aliases": ["code interpreter", "code sandbox", "Python tool"], "category": "Retrieval & Tool Use", "definition": "A code execution tool gives a language model a sandboxed environment in which it can write and run code, such as Python or shell commands, and read the output, so that it can compute, analyze data, and check its own work.", "description": "Running code replaces error-prone mental arithmetic and lets the model iterate on errors, which is the same idea behind program-aided language models. Sandboxes typically have no network access and resource limits, because the code is model-written and may be influenced by untrusted input.", "example": null, "broader": ["function-calling"], "narrower": [], "related": ["program-aided-language-models", "program-of-thoughts", "critic", "coding-agent"], "introduced": null, "sources": [{"title": "Code execution tool", "authors": "Anthropic", "year": 2025, "url": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/code-execution-tool"}, {"title": "PAL: Program-aided Language Models", "authors": "Gao et al.", "year": 2022, "url": "https://arxiv.org/abs/2211.10435"}], "url": "https://protologue.com/t/code-execution-tool/", "citation": "Protologue. (2026). Code Execution Tool. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0138). https://protologue.com/t/code-execution-tool/"} | |
| {"id": "PTL-0139", "slug": "citation-generation", "term": "Citation Generation", "aliases": ["attributed generation", "source attribution", "grounded citations"], "category": "Retrieval & Tool Use", "definition": "Citation generation is having a model attach citations to the sources that support each claim in its answer, so readers can verify the answer and unsupported claims are easier to spot.", "description": "Citations are usually drawn from retrieved documents or search results, and some APIs return them as structured spans pointing into the source. Evaluations have found that models often cite sources that do not fully support the sentence they are attached to, so citation quality, both recall and precision, needs to be measured rather than assumed.", "example": null, "broader": [], "narrower": [], "related": ["retrieval-augmented-generation", "web-search-tool", "hallucination"], "introduced": 2023, "sources": [{"title": "Enabling Large Language Models to Generate Text with Citations", "authors": "Gao et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.14627"}], "url": "https://protologue.com/t/citation-generation/", "citation": "Protologue. (2026). Citation Generation. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0139). https://protologue.com/t/citation-generation/"} | |
| {"id": "PTL-0140", "slug": "multi-agent-system", "term": "Multi-Agent System", "aliases": ["multi-agent framework", "MAS", "agent team"], "category": "Agents & Orchestration", "definition": "A multi-agent system is an application in which several LLM-based agents, each with its own role, instructions, tools, or context, interact by messaging, delegating, or debating to accomplish a task.", "description": "Splitting work across agents can parallelize research, keep each agent's context focused, and let specialized prompts handle different steps. It also multiplies token cost and adds coordination failures, such as agents duplicating work, losing information at handoffs, or reinforcing each other's mistakes. Common structures include orchestrator-workers, handoffs between specialists, and debate.", "example": null, "broader": [], "narrower": [], "related": ["ai-agent", "orchestrator-workers", "handoff", "multi-agent-debate", "agent2agent-protocol"], "introduced": null, "sources": [{"title": "Large Language Model based Multi-Agents: A Survey of Progress and Challenges", "authors": "Guo et al.", "year": 2024, "url": "https://arxiv.org/abs/2402.01680"}], "url": "https://protologue.com/t/multi-agent-system/", "citation": "Protologue. (2026). Multi-Agent System. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0140). https://protologue.com/t/multi-agent-system/"} | |
| {"id": "PTL-0141", "slug": "plan-and-execute", "term": "Plan-and-Execute", "aliases": ["planner-executor", "plan-then-execute", "ReWOO", "LLMCompiler"], "category": "Agents & Orchestration", "definition": "Plan-and-execute is an agent design in which a planner model first writes a complete multi-step plan and executor calls then carry out the steps, rather than deciding one action at a time after every observation as in ReAct.", "description": "Planning up front reduces repeated reasoning over the growing history, cuts tokens and latency, and lets independent steps run in parallel. ReWOO plans with placeholders for tool results, and LLMCompiler schedules tool calls as a dependency graph. The trade-off is adaptability, so designs usually add replanning when a step fails or returns something unexpected.", "example": null, "broader": ["agentic-workflow"], "narrower": [], "related": ["react", "plan-and-solve-prompting", "orchestrator-workers", "parallelization"], "introduced": 2023, "sources": [{"title": "ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models", "authors": "Xu et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.18323"}, {"title": "An LLM Compiler for Parallel Function Calling", "authors": "Kim et al.", "year": 2023, "url": "https://arxiv.org/abs/2312.04511"}], "url": "https://protologue.com/t/plan-and-execute/", "citation": "Protologue. (2026). Plan-and-Execute. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0141). https://protologue.com/t/plan-and-execute/"} | |
| {"id": "PTL-0142", "slug": "handoff", "term": "Handoff", "aliases": ["agent handoff", "agent transfer", "delegation"], "category": "Agents & Orchestration", "definition": "A handoff is the transfer of an ongoing task or conversation from one agent to another that specializes in it, typically implemented as a tool the first agent can call to delegate control.", "description": "Handoffs let a triage agent route users to, for example, a billing or technical support agent, each with its own instructions and tools. Unlike orchestrator-workers, where control returns to the orchestrator, a handoff passes control along. What context is passed with it decides whether the receiving agent can continue without asking again.", "example": null, "broader": [], "narrower": [], "related": ["routing", "orchestrator-workers", "multi-agent-system"], "introduced": null, "sources": [{"title": "Handoffs", "authors": "OpenAI", "year": 2025, "url": "https://openai.github.io/openai-agents-python/handoffs/"}], "url": "https://protologue.com/t/handoff/", "citation": "Protologue. (2026). Handoff. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0142). https://protologue.com/t/handoff/"} | |
| {"id": "PTL-0143", "slug": "human-in-the-loop", "term": "Human-in-the-Loop", "aliases": ["HITL", "human approval", "human oversight", "human checkpoint"], "category": "Agents & Orchestration", "definition": "Human-in-the-loop is a design in which an AI system pauses at defined checkpoints for a person to approve, correct, or supply information before it continues, especially before consequential or irreversible actions.", "description": "Checkpoints trade autonomy for safety and accuracy. They are commonly placed before sending messages, spending money, deleting data, or deploying code, and when the agent is uncertain. Requiring confirmation for consequential actions is a key mitigation for prompt injection and excessive agency, because an attacker who steers the agent still cannot complete the action unseen.", "example": null, "broader": [], "narrower": [], "related": ["ai-agent", "excessive-agency", "computer-use-agent", "evaluator-optimizer"], "introduced": null, "sources": [{"title": "Building effective agents", "authors": "Anthropic", "year": 2024, "url": "https://www.anthropic.com/research/building-effective-agents"}, {"title": "LLM06:2025 Excessive Agency", "authors": "OWASP Gen AI Security Project", "year": 2025, "url": "https://genai.owasp.org/llmrisk/llm062025-excessive-agency/"}], "url": "https://protologue.com/t/human-in-the-loop/", "citation": "Protologue. (2026). Human-in-the-Loop. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0143). https://protologue.com/t/human-in-the-loop/"} | |
| {"id": "PTL-0144", "slug": "computer-use-agent", "term": "Computer-Use Agent", "aliases": ["computer use", "GUI agent", "browser agent", "browser use"], "category": "Agents & Orchestration", "definition": "A computer-use agent operates a graphical computer interface the way a person does, by looking at screenshots and issuing mouse clicks, keystrokes, and scrolls, so it can work with software that has no API.", "description": "The agent loop alternates between a screenshot and an action until the task is done. Benchmarks such as OSWorld measure performance on real desktop tasks. Because these agents read arbitrary on-screen content, they are exposed to prompt injection from web pages and documents, so they are run in isolated environments with human confirmation for sensitive actions.", "example": null, "broader": ["ai-agent"], "narrower": [], "related": ["multimodal-prompting", "human-in-the-loop", "indirect-prompt-injection"], "introduced": 2024, "sources": [{"title": "OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments", "authors": "Xie et al.", "year": 2024, "url": "https://arxiv.org/abs/2404.07972"}, {"title": "Computer use tool", "authors": "Anthropic", "year": 2024, "url": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool"}], "url": "https://protologue.com/t/computer-use-agent/", "citation": "Protologue. (2026). Computer-Use Agent. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0144). https://protologue.com/t/computer-use-agent/"} | |
| {"id": "PTL-0145", "slug": "coding-agent", "term": "Coding Agent", "aliases": ["software engineering agent", "agentic coding", "AI coding assistant"], "category": "Agents & Orchestration", "definition": "A coding agent is an AI agent that works in a software repository by reading files, editing code, running commands and tests, and iterating on the results to complete programming tasks.", "description": "Research found that the interface matters as much as the model. Tools designed for a model, with concise file viewers, search, and editing commands that report errors, improved results over a raw shell. Coding agents are commonly evaluated on SWE-bench, which asks them to resolve real GitHub issues so that hidden tests pass.", "example": null, "broader": ["ai-agent"], "narrower": [], "related": ["code-execution-tool", "vibe-coding", "function-calling"], "introduced": 2024, "sources": [{"title": "SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering", "authors": "Yang et al.", "year": 2024, "url": "https://arxiv.org/abs/2405.15793"}, {"title": "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?", "authors": "Jimenez et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.06770"}], "url": "https://protologue.com/t/coding-agent/", "citation": "Protologue. (2026). Coding Agent. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0145). https://protologue.com/t/coding-agent/"} | |
| {"id": "PTL-0146", "slug": "vibe-coding", "term": "Vibe Coding", "aliases": [], "category": "Agents & Orchestration", "definition": "Vibe coding is building software by describing what you want to an AI model in natural language and accepting the code it generates, largely without reading or understanding that code.", "description": "Andrej Karpathy coined the term in February 2025 for throwaway projects where speed matters more than correctness. Collins English Dictionary named it its 2025 word of the year. The term is often used loosely for any AI-assisted programming, but its original sense, not reviewing the code, carries real risks of security flaws and unmaintainable software when applied to production systems.", "example": null, "broader": [], "narrower": [], "related": ["coding-agent", "prompt-engineering"], "introduced": 2025, "sources": [{"title": "Post introducing \"vibe coding\"", "authors": "Karpathy", "year": 2025, "url": "https://x.com/karpathy/status/1886192184808149383"}], "url": "https://protologue.com/t/vibe-coding/", "citation": "Protologue. (2026). Vibe Coding. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0146). https://protologue.com/t/vibe-coding/"} | |
| {"id": "PTL-0147", "slug": "agent2agent-protocol", "term": "Agent2Agent Protocol", "aliases": ["A2A", "A2A protocol"], "category": "Agents & Orchestration", "definition": "The Agent2Agent (A2A) protocol is an open specification that lets AI agents built by different vendors and frameworks discover each other's capabilities, delegate tasks, and exchange results.", "description": "Google announced A2A in April 2025 and later contributed it to the Linux Foundation. It is designed to complement the Model Context Protocol. MCP connects an agent to tools and data, while A2A connects agents to other agents. Agents advertise what they can do in a machine-readable agent card.", "example": null, "broader": [], "narrower": [], "related": ["model-context-protocol", "multi-agent-system", "handoff"], "introduced": 2025, "sources": [{"title": "Announcing the Agent2Agent Protocol (A2A)", "authors": "Google", "year": 2025, "url": "https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/"}, {"title": "A2A Protocol", "authors": "Linux Foundation A2A Project", "year": 2025, "url": "https://a2a-protocol.org/latest/"}], "url": "https://protologue.com/t/agent2agent-protocol/", "citation": "Protologue. (2026). Agent2Agent Protocol. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0147). https://protologue.com/t/agent2agent-protocol/"} | |
| {"id": "PTL-0148", "slug": "agent-skills", "term": "Agent Skills", "aliases": ["skills", "SKILL.md", "progressive disclosure"], "category": "Agents & Orchestration", "definition": "Agent skills are packaged folders of instructions, scripts, and reference files, each anchored by a SKILL.md file with a name and description, that an agent loads only when a task calls for them.", "description": "Only each skill's short description sits in the agent's context at startup. The full instructions are read when relevant, and supporting files are opened only if needed, a pattern called progressive disclosure. This lets an agent carry many specialized procedures without filling its context window, and lets organizations package their own workflows for reuse.", "example": null, "broader": [], "narrower": [], "related": ["context-engineering", "ai-agent", "model-context-protocol", "system-prompt"], "introduced": 2025, "sources": [{"title": "Equipping agents for the real world with Agent Skills", "authors": "Anthropic", "year": 2025, "url": "https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills"}], "url": "https://protologue.com/t/agent-skills/", "citation": "Protologue. (2026). Agent Skills. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0148). https://protologue.com/t/agent-skills/"} | |
| {"id": "PTL-0149", "slug": "context-compaction", "term": "Context Compaction", "aliases": ["compaction", "conversation summarization", "context summarization"], "category": "Agents & Orchestration", "definition": "Context compaction is summarizing an agent's or conversation's history when it nears the context limit and continuing with the summary in place of the full history, so that long tasks can run beyond one context window.", "description": "Good compaction keeps decisions, open problems, and key details while dropping redundant tool output. Overly aggressive summaries lose information the agent later needs. It is one of several long-horizon strategies in context engineering, alongside structured note-taking outside the context and delegating work to sub-agents with fresh contexts.", "example": null, "broader": [], "narrower": [], "related": ["context-engineering", "context-window", "agent-memory", "context-rot"], "introduced": 2025, "sources": [{"title": "Effective context engineering for AI agents", "authors": "Anthropic", "year": 2025, "url": "https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents"}], "url": "https://protologue.com/t/context-compaction/", "citation": "Protologue. (2026). Context Compaction. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0149). https://protologue.com/t/context-compaction/"} | |
| {"id": "PTL-0150", "slug": "prompt-optimization", "term": "Prompt Optimization", "aliases": ["automatic prompt optimization", "automated prompt engineering"], "category": "Prompt Optimization", "definition": "Prompt optimization is the automatic search for a better prompt, in which a language model proposes candidate instructions or examples and each candidate is scored on an evaluation set until performance stops improving.", "description": "Methods differ in how they propose changes. Some resample paraphrases, as in APE, some show the optimizer a history of scored prompts, as in OPRO, some critique errors in natural language and edit accordingly, as in ProTeGi and TextGrad, and some use evolutionary search, as in Promptbreeder and EvoPrompt. All of them depend on a representative evaluation set and metric, and can overfit to it.", "example": null, "broader": [], "narrower": ["protegi", "textgrad", "promptbreeder", "evoprompt", "agentic-context-engineering"], "related": ["automatic-prompt-engineer", "opro", "dspy", "evals"], "introduced": null, "sources": [{"title": "The Prompt Report: A Systematic Survey of Prompt Engineering Techniques", "authors": "Schulhoff et al.", "year": 2024, "url": "https://arxiv.org/abs/2406.06608"}, {"title": "Automatic Prompt Optimization with \"Gradient Descent\" and Beam Search", "authors": "Pryzant et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.03495"}], "url": "https://protologue.com/t/prompt-optimization/", "citation": "Protologue. (2026). Prompt Optimization. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0150). https://protologue.com/t/prompt-optimization/"} | |
| {"id": "PTL-0151", "slug": "protegi", "term": "ProTeGi", "aliases": ["Prompt Optimization with Textual Gradients", "textual gradients"], "category": "Prompt Optimization", "definition": "ProTeGi (Prompt Optimization with Textual Gradients) improves a prompt by having a language model describe, in natural language, what is wrong with it based on the errors it produces, then edit the prompt to fix those problems, keeping the best candidates with beam search.", "description": "The natural-language critique plays the role of a gradient, and editing the prompt plays the role of a step in the opposite direction. A bandit-style selection procedure limits how many evaluations are needed to choose among candidates.", "example": null, "broader": ["prompt-optimization"], "narrower": [], "related": ["textgrad", "automatic-prompt-engineer", "opro"], "introduced": 2023, "sources": [{"title": "Automatic Prompt Optimization with \"Gradient Descent\" and Beam Search", "authors": "Pryzant et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.03495"}], "url": "https://protologue.com/t/protegi/", "citation": "Protologue. (2026). ProTeGi. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0151). https://protologue.com/t/protegi/"} | |
| {"id": "PTL-0152", "slug": "textgrad", "term": "TextGrad", "aliases": ["textual backpropagation"], "category": "Prompt Optimization", "definition": "TextGrad is a framework that optimizes systems built from LLM calls by propagating natural-language feedback backward through the computation, so that each component, such as a prompt or a piece of code, is revised according to criticism of the final output.", "description": "It applies the idea of backpropagation to text, with an API modeled on PyTorch. Its authors used it to optimize prompts, code solutions, and answers to scientific questions. Like other feedback-driven optimizers, it is only as good as the evaluation signal it receives.", "example": null, "broader": ["prompt-optimization"], "narrower": [], "related": ["protegi", "dspy", "self-refine"], "introduced": 2024, "sources": [{"title": "TextGrad: Automatic \"Differentiation\" via Text", "authors": "Yuksekgonul et al.", "year": 2024, "url": "https://arxiv.org/abs/2406.07496"}], "url": "https://protologue.com/t/textgrad/", "citation": "Protologue. (2026). TextGrad. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0152). https://protologue.com/t/textgrad/"} | |
| {"id": "PTL-0153", "slug": "promptbreeder", "term": "Promptbreeder", "aliases": [], "category": "Prompt Optimization", "definition": "Promptbreeder is an evolutionary prompt-optimization method in which a population of task prompts is mutated and selected by fitness on training examples, while the mutation prompts that generate the changes are themselves evolved.", "description": "Because the instructions for how to mutate prompts also improve over time, the method is described as self-referential. It discovered domain-specific prompts that outperformed hand-written chain-of-thought prompts on arithmetic and commonsense benchmarks, at the cost of many model calls.", "example": null, "broader": ["prompt-optimization"], "narrower": [], "related": ["evoprompt", "opro", "automatic-prompt-engineer"], "introduced": 2023, "sources": [{"title": "Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution", "authors": "Fernando et al.", "year": 2023, "url": "https://arxiv.org/abs/2309.16797"}], "url": "https://protologue.com/t/promptbreeder/", "citation": "Protologue. (2026). Promptbreeder. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0153). https://protologue.com/t/promptbreeder/"} | |
| {"id": "PTL-0154", "slug": "evoprompt", "term": "EvoPrompt", "aliases": [], "category": "Prompt Optimization", "definition": "EvoPrompt optimizes prompts with evolutionary algorithms, such as genetic algorithms and differential evolution, in which a language model performs the crossover and mutation steps on a population of prompts.", "description": "Using the model as the evolutionary operator keeps every candidate a readable, coherent prompt while borrowing the exploration properties of classical evolutionary search. It needs only API access to the model, not its gradients.", "example": null, "broader": ["prompt-optimization"], "narrower": [], "related": ["promptbreeder", "automatic-prompt-engineer"], "introduced": 2023, "sources": [{"title": "EvoPrompt: Connecting LLMs with Evolutionary Algorithms Yields Powerful Prompt Optimizers", "authors": "Guo et al.", "year": 2023, "url": "https://arxiv.org/abs/2309.08532"}], "url": "https://protologue.com/t/evoprompt/", "citation": "Protologue. (2026). EvoPrompt. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0154). https://protologue.com/t/evoprompt/"} | |
| {"id": "PTL-0155", "slug": "model-cascade", "term": "Model Cascade", "aliases": ["LLM cascade", "FrugalGPT", "cascading"], "category": "Prompt Optimization", "definition": "A model cascade sends each query to a cheaper model first and escalates it to progressively more capable and expensive models only when a scoring function judges the cheaper model's answer unreliable.", "description": "Because many queries are easy, most are answered cheaply, and only hard ones pay for the strongest model. FrugalGPT reported matching the best single model's performance at a large reduction in cost. The design depends on a reliable way to judge when an answer is good enough. Routing, by contrast, chooses a model before any answer is generated.", "example": null, "broader": [], "narrower": [], "related": ["routing", "confidence-calibration", "llm-as-a-judge"], "introduced": 2023, "sources": [{"title": "FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance", "authors": "Chen et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.05176"}], "url": "https://protologue.com/t/model-cascade/", "citation": "Protologue. (2026). Model Cascade. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0155). https://protologue.com/t/model-cascade/"} | |
| {"id": "PTL-0156", "slug": "rlvr", "term": "Reinforcement Learning with Verifiable Rewards", "aliases": ["RLVR", "RL from verifiable rewards", "outcome-based reinforcement learning"], "category": "Reasoning Models & Test-Time Compute", "definition": "Reinforcement learning with verifiable rewards (RLVR) trains a language model with rewards computed by an automatic check of correctness, such as whether a math answer matches the reference or code passes its tests, instead of by a learned reward model.", "description": "Verifiable rewards are hard to game in the ways a learned reward model can be, and large-scale RL with them is how reasoning models learn to produce long, self-correcting chains of thought. The approach is limited to tasks with checkable answers, so open-ended writing and judgment still rely on preference-based rewards.", "example": null, "broader": [], "narrower": [], "related": ["reasoning-model", "rlhf", "reward-model", "reward-hacking", "self-taught-reasoner"], "introduced": 2024, "sources": [{"title": "Tulu 3: Pushing Frontiers in Open Language Model Post-Training", "authors": "Lambert et al.", "year": 2024, "url": "https://arxiv.org/abs/2411.15124"}, {"title": "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning", "authors": "DeepSeek-AI et al.", "year": 2025, "url": "https://arxiv.org/abs/2501.12948"}], "url": "https://protologue.com/t/rlvr/", "citation": "Protologue. (2026). Reinforcement Learning with Verifiable Rewards. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0156). https://protologue.com/t/rlvr/"} | |
| {"id": "PTL-0157", "slug": "budget-forcing", "term": "Budget Forcing", "aliases": ["s1", "wait token", "thinking budget control"], "category": "Reasoning Models & Test-Time Compute", "definition": "Budget forcing controls how long a reasoning model thinks at inference time, by forcibly ending its reasoning when a token budget is reached or by appending a word such as \"Wait\" when it tries to stop early, which makes it keep reasoning.", "description": "It was introduced with s1, a model fine-tuned on only about a thousand reasoning examples, where extending thinking with budget forcing improved accuracy on competition math. It is a simple way to trade compute for accuracy, though the gains flatten beyond a point.", "example": null, "broader": ["test-time-compute-scaling"], "narrower": [], "related": ["reasoning-model", "extended-thinking", "overthinking"], "introduced": 2025, "sources": [{"title": "s1: Simple test-time scaling", "authors": "Muennighoff et al.", "year": 2025, "url": "https://arxiv.org/abs/2501.19393"}], "url": "https://protologue.com/t/budget-forcing/", "citation": "Protologue. (2026). Budget Forcing. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0157). https://protologue.com/t/budget-forcing/"} | |
| {"id": "PTL-0158", "slug": "chain-of-thought-monitoring", "term": "Chain-of-Thought Monitoring", "aliases": ["CoT monitoring", "CoT monitorability", "reasoning trace monitoring"], "category": "Reasoning Models & Test-Time Compute", "definition": "Chain-of-thought monitoring is reading a model's reasoning trace, with automated monitors or people, to detect intentions to misbehave, such as plans to cheat on a task, before or alongside its actions.", "description": "Because reasoning models often state their intentions in natural language while solving hard tasks, their reasoning offers an unusual view into their decision making. Researchers from several labs have argued that this monitorability is valuable but fragile. Training pressure on the reasoning text, or architectures that reason without words, could remove it, and reasoning traces are not always faithful to the model's actual computation.", "example": null, "broader": [], "narrower": [], "related": ["reasoning-model", "unfaithful-chain-of-thought", "reward-hacking", "extended-thinking"], "introduced": 2025, "sources": [{"title": "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety", "authors": "Korbak et al.", "year": 2025, "url": "https://arxiv.org/abs/2507.11473"}], "url": "https://protologue.com/t/chain-of-thought-monitoring/", "citation": "Protologue. (2026). Chain-of-Thought Monitoring. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0158). https://protologue.com/t/chain-of-thought-monitoring/"} | |
| {"id": "PTL-0159", "slug": "red-teaming", "term": "Red Teaming", "aliases": ["adversarial testing", "AI red teaming", "automated red teaming"], "category": "Security & Adversarial Prompting", "definition": "Red teaming is adversarial testing of an AI model or system, by people or by other models, to find inputs that make it produce harmful, insecure, or otherwise unwanted behavior before attackers or users do.", "description": "Human red teamers bring creativity and domain expertise, while automated red teaming uses a language model to generate and refine attacks at scale. Findings feed training data, guardrails, and policy, and red teaming is repeated as models and defenses change.", "example": null, "broader": [], "narrower": [], "related": ["jailbreak", "prompt-injection", "guardrails", "pair-attack", "constitutional-classifiers"], "introduced": 2022, "sources": [{"title": "Red Teaming Language Models with Language Models", "authors": "Perez et al.", "year": 2022, "url": "https://arxiv.org/abs/2202.03286"}], "url": "https://protologue.com/t/red-teaming/", "citation": "Protologue. (2026). Red Teaming. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0159). https://protologue.com/t/red-teaming/"} | |
| {"id": "PTL-0160", "slug": "guardrails", "term": "Guardrails", "aliases": ["safety filters", "input/output filtering", "safeguard model", "moderation", "content filter"], "category": "Security & Adversarial Prompting", "definition": "Guardrails are checks placed around a model call that screen inputs and outputs, using classifiers, rules, or a second model, to block, redact, or redirect content that violates a policy before it reaches the model or the user.", "description": "Typical guardrails detect jailbreak attempts and prompt injection, unsafe content, personal data, and off-topic requests. Safeguard models such as Llama Guard classify messages against a taxonomy of risks, and toolkits such as NeMo Guardrails define programmable rails for dialogue. Guardrails add latency and can over-block, and none is complete, so they are layered with model training and least-privilege design.", "example": null, "broader": [], "narrower": ["constitutional-classifiers"], "related": ["jailbreak", "prompt-injection", "over-refusal", "red-teaming"], "introduced": 2023, "sources": [{"title": "Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations", "authors": "Inan et al.", "year": 2023, "url": "https://arxiv.org/abs/2312.06674"}, {"title": "NeMo Guardrails: A Toolkit for Controllable and Safe LLM Applications with Programmable Rails", "authors": "Rebedea et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.10501"}], "url": "https://protologue.com/t/guardrails/", "citation": "Protologue. (2026). Guardrails. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0160). https://protologue.com/t/guardrails/"} | |
| {"id": "PTL-0161", "slug": "dual-llm-pattern", "term": "Dual LLM Pattern", "aliases": ["privileged and quarantined LLM", "CaMeL"], "category": "Security & Adversarial Prompting", "definition": "The dual LLM pattern defends against prompt injection by splitting work between a privileged model, which sees only trusted input and can use tools, and a quarantined model, which processes untrusted content but has no tool access, with ordinary code passing references between them.", "description": "The privileged model never reads untrusted text directly, so instructions hidden in that text cannot trigger its tools. CaMeL extends the idea by having the privileged model write a program whose data flows are tracked and checked against security policies. These designs give security guarantees by construction rather than by training, at the cost of flexibility.", "example": null, "broader": [], "narrower": [], "related": ["prompt-injection", "indirect-prompt-injection", "lethal-trifecta", "spotlighting", "instruction-hierarchy"], "introduced": 2023, "sources": [{"title": "The Dual LLM pattern for building AI assistants that can resist prompt injection", "authors": "Willison", "year": 2023, "url": "https://simonwillison.net/2023/Apr/25/dual-llm-pattern/"}, {"title": "Defeating Prompt Injections by Design", "authors": "Debenedetti et al.", "year": 2025, "url": "https://arxiv.org/abs/2503.18813"}, {"title": "Design Patterns for Securing LLM Agents against Prompt Injections", "authors": "Beurer-Kellner et al.", "year": 2025, "url": "https://arxiv.org/abs/2506.08837"}], "url": "https://protologue.com/t/dual-llm-pattern/", "citation": "Protologue. (2026). Dual LLM Pattern. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0161). https://protologue.com/t/dual-llm-pattern/"} | |
| {"id": "PTL-0162", "slug": "tool-poisoning", "term": "Tool Poisoning", "aliases": ["tool poisoning attack", "MCP tool poisoning"], "category": "Security & Adversarial Prompting", "definition": "Tool poisoning is an attack in which malicious instructions are hidden in a tool's description or metadata, such as an MCP server's tool definitions, so that the agent reading the description is steered into harmful actions that the user does not see.", "description": "Agents read tool descriptions as trusted context, but users rarely see them in full. A poisoned tool can instruct the agent to read sensitive files or to send data through another, legitimate tool. Mitigations include showing full tool descriptions, pinning tool versions so descriptions cannot change silently, and installing servers only from trusted sources.", "example": null, "broader": ["indirect-prompt-injection"], "narrower": [], "related": ["model-context-protocol", "function-calling", "lethal-trifecta"], "introduced": 2025, "sources": [{"title": "MCP Security Notification: Tool Poisoning Attacks", "authors": "Invariant Labs", "year": 2025, "url": "https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks"}], "url": "https://protologue.com/t/tool-poisoning/", "citation": "Protologue. (2026). Tool Poisoning. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0162). https://protologue.com/t/tool-poisoning/"} | |
| {"id": "PTL-0163", "slug": "lethal-trifecta", "term": "Lethal Trifecta", "aliases": [], "category": "Security & Adversarial Prompting", "definition": "The lethal trifecta is the combination of three capabilities that makes an AI agent vulnerable to data theft through prompt injection, namely access to private data, exposure to untrusted content, and the ability to communicate externally.", "description": "Any one or two are manageable, but with all three an attacker who plants instructions in content the agent reads can make it send private data out. Because no prompt-injection defense is fully reliable, the practical advice is to remove at least one of the three from any agent configuration.", "example": null, "broader": [], "narrower": [], "related": ["prompt-injection", "indirect-prompt-injection", "excessive-agency", "dual-llm-pattern", "tool-poisoning"], "introduced": 2025, "sources": [{"title": "The lethal trifecta for AI agents: private data, untrusted content, and external communication", "authors": "Willison", "year": 2025, "url": "https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/"}], "url": "https://protologue.com/t/lethal-trifecta/", "citation": "Protologue. (2026). Lethal Trifecta. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0163). https://protologue.com/t/lethal-trifecta/"} | |
| {"id": "PTL-0164", "slug": "crescendo-attack", "term": "Crescendo Attack", "aliases": ["Crescendo", "multi-turn jailbreak"], "category": "Security & Adversarial Prompting", "definition": "The Crescendo attack is a multi-turn jailbreak that starts with harmless questions on a topic and gradually escalates over several turns, building on the model's own earlier answers until it produces content it would refuse if asked directly.", "description": "Each individual message looks benign, so defenses that check single prompts miss it. The attack exploits the model's tendency to stay consistent with the conversation so far. Defenses need to evaluate the trajectory of a conversation, not only its latest turn.", "example": null, "broader": ["jailbreak"], "narrower": [], "related": ["many-shot-jailbreaking", "multi-turn-degradation", "guardrails"], "introduced": 2024, "sources": [{"title": "Great, Now Write an Article About That: The Crescendo Multi-Turn LLM Jailbreak Attack", "authors": "Russinovich et al.", "year": 2024, "url": "https://arxiv.org/abs/2404.01833"}], "url": "https://protologue.com/t/crescendo-attack/", "citation": "Protologue. (2026). Crescendo Attack. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0164). https://protologue.com/t/crescendo-attack/"} | |
| {"id": "PTL-0165", "slug": "pair-attack", "term": "PAIR", "aliases": ["Prompt Automatic Iterative Refinement", "automated jailbreaking"], "category": "Security & Adversarial Prompting", "definition": "PAIR (Prompt Automatic Iterative Refinement) is an automated jailbreak method in which an attacker language model repeatedly proposes jailbreak prompts against a target model and refines them based on the target's responses, needing only black-box access.", "description": "The original work reported successful jailbreaks often within about twenty queries, and the resulting prompts are readable, unlike adversarial suffixes. It is widely used as a baseline for automated red teaming.", "example": null, "broader": ["jailbreak"], "narrower": [], "related": ["red-teaming", "adversarial-suffix"], "introduced": 2023, "sources": [{"title": "Jailbreaking Black Box Large Language Models in Twenty Queries", "authors": "Chao et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.08419"}], "url": "https://protologue.com/t/pair-attack/", "citation": "Protologue. (2026). PAIR. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0165). https://protologue.com/t/pair-attack/"} | |
| {"id": "PTL-0166", "slug": "excessive-agency", "term": "Excessive Agency", "aliases": [], "category": "Security & Adversarial Prompting", "definition": "Excessive agency is the vulnerability of giving an LLM-based system more functionality, permissions, or autonomy than its task requires, so that a manipulated or mistaken model can take damaging actions.", "description": "It is listed as LLM06 in the OWASP Top 10 for LLM Applications (2025), which names three root causes, excessive functionality, excessive permissions, and excessive autonomy. Mitigations are least-privilege tools and credentials, narrowly scoped functions instead of open-ended ones, and human approval for high-impact actions.", "example": null, "broader": [], "narrower": [], "related": ["human-in-the-loop", "lethal-trifecta", "prompt-injection", "function-calling"], "introduced": null, "sources": [{"title": "LLM06:2025 Excessive Agency", "authors": "OWASP Gen AI Security Project", "year": 2025, "url": "https://genai.owasp.org/llmrisk/llm062025-excessive-agency/"}], "url": "https://protologue.com/t/excessive-agency/", "citation": "Protologue. (2026). Excessive Agency. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0166). https://protologue.com/t/excessive-agency/"} | |
| {"id": "PTL-0167", "slug": "deliberative-alignment", "term": "Deliberative Alignment", "aliases": [], "category": "Security & Adversarial Prompting", "definition": "Deliberative alignment is a safety-training method in which a reasoning model is taught the text of a safety specification and trained to recall and reason about the relevant policies in its chain of thought before answering.", "description": "Instead of learning safe behavior only from labeled examples, the model learns to consult explicit rules at inference time. Its authors reported better resistance to jailbreaks together with fewer unnecessary refusals. It is related to constitutional AI, which also trains from written principles.", "example": null, "broader": [], "narrower": [], "related": ["constitutional-ai", "reasoning-model", "over-refusal", "jailbreak"], "introduced": 2024, "sources": [{"title": "Deliberative Alignment: Reasoning Enables Safer Language Models", "authors": "Guan et al.", "year": 2024, "url": "https://arxiv.org/abs/2412.16339"}], "url": "https://protologue.com/t/deliberative-alignment/", "citation": "Protologue. (2026). Deliberative Alignment. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0167). https://protologue.com/t/deliberative-alignment/"} | |
| {"id": "PTL-0168", "slug": "constitutional-classifiers", "term": "Constitutional Classifiers", "aliases": [], "category": "Security & Adversarial Prompting", "definition": "Constitutional classifiers are input and output safeguard classifiers trained on synthetic data generated from a written constitution that defines permitted and restricted content, designed to block universal jailbreaks.", "description": "Because the training data is generated from natural-language rules, the classifiers can be updated as threats change by editing the constitution. Their authors reported that thousands of hours of red teaming found no universal jailbreak against the protected system, at a modest cost in refusals and compute.", "example": null, "broader": ["guardrails"], "narrower": [], "related": ["constitutional-ai", "jailbreak", "red-teaming"], "introduced": 2025, "sources": [{"title": "Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming", "authors": "Sharma et al.", "year": 2025, "url": "https://arxiv.org/abs/2501.18837"}], "url": "https://protologue.com/t/constitutional-classifiers/", "citation": "Protologue. (2026). Constitutional Classifiers. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0168). https://protologue.com/t/constitutional-classifiers/"} | |
| {"id": "PTL-0169", "slug": "context-rot", "term": "Context Rot", "aliases": ["long-context degradation", "context degradation"], "category": "Failure Modes & Evaluation", "definition": "Context rot is the decline in a model's reliability as the amount of input in its context grows, even on simple tasks and well within its advertised context window.", "description": "A 2025 study of 18 models found performance degrading with input length in ways that depended on the task and on the content around the relevant information, such as similar-looking distractors. It is a practical reason to keep contexts lean, retrieve only what is needed, and compact long histories rather than relying on maximum context length.", "example": null, "broader": [], "narrower": [], "related": ["lost-in-the-middle", "needle-in-a-haystack", "context-window", "context-engineering", "context-compaction"], "introduced": 2025, "sources": [{"title": "Context Rot: How Increasing Input Tokens Impacts LLM Performance", "authors": "Hong et al.", "year": 2025, "url": "https://www.trychroma.com/research/context-rot"}], "url": "https://protologue.com/t/context-rot/", "citation": "Protologue. (2026). Context Rot. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0169). https://protologue.com/t/context-rot/"} | |
| {"id": "PTL-0170", "slug": "overthinking", "term": "Overthinking", "aliases": ["overthinking problem", "excessive reasoning"], "category": "Failure Modes & Evaluation", "definition": "Overthinking is the tendency of reasoning models to spend far more reasoning tokens than a problem requires, for example writing long chains of thought for trivial arithmetic, which adds cost and latency with little or no accuracy gain.", "description": "Analysis of o1-style models found that later reasoning rounds on simple problems rarely contribute new solutions. Remedies include adjustable reasoning effort or budgets, routing easy queries to non-reasoning models, and training that rewards efficient reasoning.", "example": null, "broader": [], "narrower": [], "related": ["reasoning-model", "extended-thinking", "budget-forcing", "chain-of-draft", "test-time-compute-scaling"], "introduced": 2024, "sources": [{"title": "Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs", "authors": "Chen et al.", "year": 2024, "url": "https://arxiv.org/abs/2412.21187"}], "url": "https://protologue.com/t/overthinking/", "citation": "Protologue. (2026). Overthinking. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0170). https://protologue.com/t/overthinking/"} | |
| {"id": "PTL-0171", "slug": "reversal-curse", "term": "Reversal Curse", "aliases": [], "category": "Failure Modes & Evaluation", "definition": "The reversal curse is the failure of language models trained on a fact stated as \"A is B\" to answer the reversed question \"what is B?\", for example knowing who a celebrity's mother is but not who that woman's famous son is.", "description": "It shows that models do not automatically store facts in both directions. In practice, providing the relevant fact in the prompt fixes it, since in-context reasoning handles the reversal, which is one more argument for retrieval over relying on memorized knowledge.", "example": null, "broader": [], "narrower": [], "related": ["hallucination", "knowledge-cutoff", "retrieval-augmented-generation"], "introduced": 2023, "sources": [{"title": "The Reversal Curse: LLMs trained on \"A is B\" fail to learn \"B is A\"", "authors": "Berglund et al.", "year": 2023, "url": "https://arxiv.org/abs/2309.12288"}], "url": "https://protologue.com/t/reversal-curse/", "citation": "Protologue. (2026). Reversal Curse. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0171). https://protologue.com/t/reversal-curse/"} | |
| {"id": "PTL-0172", "slug": "reward-hacking", "term": "Reward Hacking", "aliases": ["specification gaming", "reward overoptimization", "reward gaming"], "category": "Failure Modes & Evaluation", "definition": "Reward hacking is when a model trained or optimized against a reward finds ways to score highly that do not achieve the intended goal, exploiting flaws in the reward model, grader, or tests.", "description": "In language models it shows up as answers padded to look thorough, flattery that raters prefer, and code that special-cases tests instead of solving the problem. Verifiable rewards reduce some forms, while careful graders, held-out tests, and monitoring of reasoning traces help detect others.", "example": null, "broader": [], "narrower": [], "related": ["reward-model", "rlhf", "rlvr", "sycophancy", "chain-of-thought-monitoring"], "introduced": null, "sources": [{"title": "Concrete Problems in AI Safety", "authors": "Amodei et al.", "year": 2016, "url": "https://arxiv.org/abs/1606.06565"}], "url": "https://protologue.com/t/reward-hacking/", "citation": "Protologue. (2026). Reward Hacking. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0172). https://protologue.com/t/reward-hacking/"} | |
| {"id": "PTL-0173", "slug": "irrelevant-context-distraction", "term": "Irrelevant Context Distraction", "aliases": ["distractibility", "GSM-IC"], "category": "Failure Modes & Evaluation", "definition": "Irrelevant context distraction is the drop in a model's accuracy when a prompt contains information that is unrelated to the question, such as an extra sentence in a math word problem, which the model wrongly incorporates into its reasoning.", "description": "A benchmark built by adding irrelevant sentences to grade-school math problems showed large accuracy drops across prompting methods. Instructing the model to ignore irrelevant information, using self-consistency, and filtering context first, as in System 2 Attention, all reduce the effect. It matters for RAG, where retrieved passages often include off-topic text.", "example": null, "broader": [], "narrower": [], "related": ["system-2-attention", "retrieval-augmented-generation", "context-rot", "prompt-sensitivity"], "introduced": 2023, "sources": [{"title": "Large Language Models Can Be Easily Distracted by Irrelevant Context", "authors": "Shi et al.", "year": 2023, "url": "https://arxiv.org/abs/2302.00093"}], "url": "https://protologue.com/t/irrelevant-context-distraction/", "citation": "Protologue. (2026). Irrelevant Context Distraction. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0173). https://protologue.com/t/irrelevant-context-distraction/"} | |
| {"id": "PTL-0174", "slug": "multi-turn-degradation", "term": "Multi-Turn Degradation", "aliases": ["lost in conversation", "multi-turn performance drop"], "category": "Failure Modes & Evaluation", "definition": "Multi-turn degradation is the drop in model performance when a task's requirements are revealed across several conversation turns instead of all at once, because models commit to early assumptions and fail to recover.", "description": "A large 2025 study found that models that performed well on fully specified single-turn instructions did markedly worse in multi-turn settings, mainly through increased unreliability rather than a loss of capability. The practical advice is to consolidate requirements into a fresh prompt when a conversation goes off track.", "example": null, "broader": [], "narrower": [], "related": ["instruction-following", "context-compaction", "crescendo-attack", "context-rot"], "introduced": 2025, "sources": [{"title": "LLMs Get Lost In Multi-Turn Conversation", "authors": "Laban et al.", "year": 2025, "url": "https://arxiv.org/abs/2505.06120"}], "url": "https://protologue.com/t/multi-turn-degradation/", "citation": "Protologue. (2026). Multi-Turn Degradation. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0174). https://protologue.com/t/multi-turn-degradation/"} | |
| {"id": "PTL-0175", "slug": "text-degeneration", "term": "Text Degeneration", "aliases": ["neural text degeneration", "repetition", "repetition loops"], "category": "Failure Modes & Evaluation", "definition": "Text degeneration is the tendency of language models, especially under greedy or beam-search decoding, to produce bland, repetitive, or looping text, such as the same phrase repeated over and over.", "description": "Maximizing likelihood at each step favors safe, high-probability continuations that reinforce themselves once repetition starts. Sampling methods such as top-p and top-k, repetition or frequency penalties, and modern training largely mitigate it, but loops still appear in long generations and agent traces.", "example": null, "broader": [], "narrower": [], "related": ["greedy-decoding", "top-p-sampling", "top-k-sampling", "temperature"], "introduced": 2019, "sources": [{"title": "The Curious Case of Neural Text Degeneration", "authors": "Holtzman et al.", "year": 2019, "url": "https://arxiv.org/abs/1904.09751"}], "url": "https://protologue.com/t/text-degeneration/", "citation": "Protologue. (2026). Text Degeneration. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0175). https://protologue.com/t/text-degeneration/"} | |
| {"id": "PTL-0176", "slug": "over-refusal", "term": "Over-Refusal", "aliases": ["exaggerated safety", "false refusal", "overrefusal"], "category": "Failure Modes & Evaluation", "definition": "Over-refusal is a model declining safe requests because they superficially resemble harmful ones, such as refusing to explain how to kill a Python process.", "description": "It is the cost side of safety training, since overly cautious models frustrate users and withhold legitimate help. Test suites such as XSTest pair safe prompts with unsafe look-alikes to measure the balance, and methods such as deliberative alignment aim to reduce refusals while improving robustness.", "example": null, "broader": [], "narrower": [], "related": ["guardrails", "rlhf", "deliberative-alignment", "constitutional-classifiers"], "introduced": 2023, "sources": [{"title": "XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models", "authors": "Röttger et al.", "year": 2023, "url": "https://arxiv.org/abs/2308.01263"}], "url": "https://protologue.com/t/over-refusal/", "citation": "Protologue. (2026). Over-Refusal. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0176). https://protologue.com/t/over-refusal/"} | |
| {"id": "PTL-0177", "slug": "llm-judge-bias", "term": "LLM Judge Bias", "aliases": ["position bias", "verbosity bias", "self-enhancement bias", "self-preference bias"], "category": "Failure Modes & Evaluation", "definition": "LLM judge bias is the set of systematic errors that language models make when grading outputs, notably preferring the answer shown first (position bias), longer answers (verbosity bias), and their own outputs (self-enhancement bias).", "description": "These biases can make automated evaluations favor the wrong prompt or model. Common mitigations are evaluating each pair in both orders and counting only consistent verdicts, instructing the judge to ignore length, using clear rubrics and reference answers, and checking judge agreement against a sample of human ratings.", "example": null, "broader": [], "narrower": [], "related": ["llm-as-a-judge", "evals", "sycophancy"], "introduced": 2023, "sources": [{"title": "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena", "authors": "Zheng et al.", "year": 2023, "url": "https://arxiv.org/abs/2306.05685"}], "url": "https://protologue.com/t/llm-judge-bias/", "citation": "Protologue. (2026). LLM Judge Bias. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0177). https://protologue.com/t/llm-judge-bias/"} | |
| {"id": "PTL-0178", "slug": "data-contamination", "term": "Data Contamination", "aliases": ["benchmark contamination", "test set contamination", "training data leakage"], "category": "Failure Modes & Evaluation", "definition": "Data contamination is the presence of benchmark or test examples in a model's training data, which inflates its measured performance because it may have memorized the answers.", "description": "Since models are trained on large web crawls, public benchmarks often leak into training sets. Contamination can be detected, for example, by prompting a model to complete the beginning of test instances and checking for verbatim matches. For application evaluations, the safeguard is a private, regularly refreshed test set.", "example": null, "broader": [], "narrower": [], "related": ["evals", "knowledge-cutoff"], "introduced": 2023, "sources": [{"title": "Time Travel in LLMs: Tracing Data Contamination in Large Language Models", "authors": "Golchin and Surdeanu", "year": 2023, "url": "https://arxiv.org/abs/2308.08493"}], "url": "https://protologue.com/t/data-contamination/", "citation": "Protologue. (2026). Data Contamination. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0178). https://protologue.com/t/data-contamination/"} | |
| {"id": "PTL-0179", "slug": "evals", "term": "Evals", "aliases": ["evaluations", "LLM evaluation", "eval suite", "test set"], "category": "Failure Modes & Evaluation", "definition": "Evals are systematic tests of a model or LLM application, consisting of a set of representative inputs and a grading method, used to measure performance against defined success criteria and to compare prompts, models, and changes.", "description": "Grading can be exact-match or code-based, done by an LLM judge with a rubric, or done by people. Code-based grading is fastest and most reliable where it applies. A good eval set mirrors real usage, including edge cases, and turns prompt engineering from guesswork into measurement. Without one, changes that fix one case can silently break others.", "example": null, "broader": [], "narrower": ["pass-at-k", "rag-evaluation"], "related": ["llm-as-a-judge", "prompt-optimization", "data-contamination"], "introduced": null, "sources": [{"title": "Define success criteria and build evaluations", "authors": "Anthropic", "year": 2024, "url": "https://platform.claude.com/docs/en/test-and-evaluate/develop-tests"}, {"title": "Holistic Evaluation of Language Models", "authors": "Liang et al.", "year": 2022, "url": "https://arxiv.org/abs/2211.09110"}], "url": "https://protologue.com/t/evals/", "citation": "Protologue. (2026). Evals. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0179). https://protologue.com/t/evals/"} | |
| {"id": "PTL-0180", "slug": "pass-at-k", "term": "pass@k", "aliases": ["pass@1", "pass at k"], "category": "Failure Modes & Evaluation", "definition": "pass@k is an evaluation metric for code generation, and other tasks with automatic checks, defined as the probability that at least one of k sampled solutions to a problem passes its tests.", "description": "It was popularized with an unbiased estimator computed from n ≥ k samples per problem. pass@1 measures single-attempt reliability, while larger k measures whether the model can solve the problem at all, which is relevant when best-of-n sampling or verification is available.", "example": null, "broader": ["evals"], "narrower": [], "related": ["best-of-n-sampling", "coding-agent", "rlvr"], "introduced": 2021, "sources": [{"title": "Evaluating Large Language Models Trained on Code", "authors": "Chen et al.", "year": 2021, "url": "https://arxiv.org/abs/2107.03374"}], "url": "https://protologue.com/t/pass-at-k/", "citation": "Protologue. (2026). pass@k. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0180). https://protologue.com/t/pass-at-k/"} | |
| {"id": "PTL-0181", "slug": "rag-evaluation", "term": "RAG Evaluation", "aliases": ["Ragas", "RAG metrics", "faithfulness", "context relevance"], "category": "Failure Modes & Evaluation", "definition": "RAG evaluation measures a retrieval-augmented system's retrieval and generation separately, typically checking whether the retrieved context is relevant, whether the answer is faithful to that context, and whether the answer addresses the question.", "description": "Separating the stages shows whether a bad answer came from retrieval or from generation. Frameworks such as Ragas compute these metrics automatically with an LLM, without reference answers, which makes it practical to compare chunking, retrieval, and prompt choices.", "example": null, "broader": ["evals"], "narrower": [], "related": ["retrieval-augmented-generation", "llm-as-a-judge", "hallucination", "chunking"], "introduced": 2023, "sources": [{"title": "Ragas: Automated Evaluation of Retrieval Augmented Generation", "authors": "Es et al.", "year": 2023, "url": "https://arxiv.org/abs/2309.15217"}], "url": "https://protologue.com/t/rag-evaluation/", "citation": "Protologue. (2026). RAG Evaluation. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0181). https://protologue.com/t/rag-evaluation/"} | |
| {"id": "PTL-0182", "slug": "kv-cache", "term": "KV Cache", "aliases": ["key-value cache", "KV caching"], "category": "Foundations", "definition": "The KV cache stores the attention keys and values a Transformer has already computed for earlier tokens, so that generating each new token requires computing only that token's keys and values instead of reprocessing the whole sequence.", "description": "Caching makes generation fast, but its memory grows with context length and with the number of concurrent requests, which is a main constraint on long contexts and serving costs. Serving systems such as vLLM manage it in pages to reduce waste. Prompt caching extends the idea across requests by keeping the cache for a shared prompt prefix, which is why cached prompts must begin with identical content to get the discount.", "example": null, "broader": [], "narrower": [], "related": ["prompt-caching", "context-window", "token"], "introduced": null, "sources": [{"title": "Efficient Memory Management for Large Language Model Serving with PagedAttention", "authors": "Kwon et al.", "year": 2023, "url": "https://arxiv.org/abs/2309.06180"}, {"title": "Prompt caching", "authors": "Anthropic", "year": 2024, "url": "https://platform.claude.com/docs/en/build-with-claude/prompt-caching"}], "url": "https://protologue.com/t/kv-cache/", "citation": "Protologue. (2026). KV Cache. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0182). https://protologue.com/t/kv-cache/"} | |
| {"id": "PTL-0183", "slug": "knowledge-distillation", "term": "Knowledge Distillation", "aliases": ["distillation", "model distillation", "teacher-student training"], "category": "Foundations", "definition": "Knowledge distillation is training a smaller student model to imitate a larger teacher model, using the teacher's output probabilities or generated responses as training targets, so that the student keeps much of the teacher's ability at lower cost.", "description": "The original method trains on the teacher's full probability distribution, whose soft targets carry more information than single correct labels. With LLMs, distillation is commonly done by fine-tuning a small model on responses, including reasoning traces, generated by a large one, which is how several small reasoning models have been built. Some model providers' terms restrict using their outputs to train competing models.", "example": null, "broader": ["fine-tuning"], "narrower": [], "related": ["reasoning-model", "model-cascade"], "introduced": 2015, "sources": [{"title": "Distilling the Knowledge in a Neural Network", "authors": "Hinton et al.", "year": 2015, "url": "https://arxiv.org/abs/1503.02531"}, {"title": "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning", "authors": "DeepSeek-AI et al.", "year": 2025, "url": "https://arxiv.org/abs/2501.12948"}], "url": "https://protologue.com/t/knowledge-distillation/", "citation": "Protologue. (2026). Knowledge Distillation. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0183). https://protologue.com/t/knowledge-distillation/"} | |
| {"id": "PTL-0184", "slug": "emergent-abilities", "term": "Emergent Abilities", "aliases": ["emergence", "emergent capabilities"], "category": "Foundations", "definition": "Emergent abilities are capabilities that appear in large language models above a certain scale but are near chance in smaller ones, such as multi-step arithmetic or benefiting from chain-of-thought prompting.", "description": "The idea explained why some prompting techniques only work with large models. It is disputed. Later analysis argued that many apparent jumps are artifacts of all-or-nothing metrics such as exact match, and that smoother metrics show gradual improvement with scale. The practical lesson stands either way, which is that a technique that fails on a small model may work on a larger one, and the reverse.", "example": null, "broader": [], "narrower": [], "related": ["large-language-model", "chain-of-thought", "in-context-learning", "benchmark"], "introduced": 2022, "sources": [{"title": "Emergent Abilities of Large Language Models", "authors": "Wei et al.", "year": 2022, "url": "https://arxiv.org/abs/2206.07682"}, {"title": "Are Emergent Abilities of Large Language Models a Mirage?", "authors": "Schaeffer et al.", "year": 2023, "url": "https://arxiv.org/abs/2304.15004"}], "url": "https://protologue.com/t/emergent-abilities/", "citation": "Protologue. (2026). Emergent Abilities. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0184). https://protologue.com/t/emergent-abilities/"} | |
| {"id": "PTL-0185", "slug": "benchmark", "term": "Benchmark", "aliases": ["LLM benchmark", "evaluation benchmark"], "category": "Benchmarks", "definition": "A benchmark is a fixed, public set of test tasks with a standard scoring method, used to measure and compare the capabilities of language models.", "description": "Reported scores depend on the prompting setup, such as the number of examples, whether chain-of-thought or extended thinking is used, and how many attempts are sampled, so scores are only comparable under the same conditions. Public benchmarks also risk data contamination and eventually saturate. They are useful for choosing models, but an application's own evals are the better guide to how a model will perform on its tasks.", "example": null, "broader": [], "narrower": ["benchmark-saturation", "mmlu", "mmlu-pro", "gpqa", "humanitys-last-exam", "big-bench-hard", "arc-agi", "simpleqa", "truthfulqa", "hellaswag", "mmmu", "gsm8k", "math-benchmark", "aime", "humaneval", "livecodebench", "swe-bench", "terminal-bench", "tau-bench", "osworld", "gaia", "browsecomp", "ifeval", "mt-bench", "chatbot-arena", "ruler"], "related": ["evals", "data-contamination", "pass-at-k"], "introduced": null, "sources": [{"title": "Holistic Evaluation of Language Models", "authors": "Liang et al.", "year": 2022, "url": "https://arxiv.org/abs/2211.09110"}, {"title": "Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models", "authors": "Srivastava et al.", "year": 2022, "url": "https://arxiv.org/abs/2206.04615"}], "url": "https://protologue.com/t/benchmark/", "citation": "Protologue. (2026). Benchmark. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0185). https://protologue.com/t/benchmark/"} | |
| {"id": "PTL-0186", "slug": "benchmark-saturation", "term": "Benchmark Saturation", "aliases": ["saturated benchmark"], "category": "Benchmarks", "definition": "Benchmark saturation is when the best models score near a benchmark's maximum, so that it no longer distinguishes between them and its remaining errors mostly reflect flawed or ambiguous questions.", "description": "Saturation has followed every widely used benchmark, often within a few years, and is a main reason new, harder benchmarks keep appearing. MMLU-Pro and Humanity's Last Exam were both created because earlier benchmarks had saturated. Contamination can make a benchmark look saturated before models have truly mastered it.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["data-contamination", "mmlu", "humanitys-last-exam"], "introduced": null, "sources": [{"title": "Humanity's Last Exam", "authors": "Phan et al.", "year": 2025, "url": "https://arxiv.org/abs/2501.14249"}, {"title": "MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark", "authors": "Wang et al.", "year": 2024, "url": "https://arxiv.org/abs/2406.01574"}], "url": "https://protologue.com/t/benchmark-saturation/", "citation": "Protologue. (2026). Benchmark Saturation. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0186). https://protologue.com/t/benchmark-saturation/"} | |
| {"id": "PTL-0187", "slug": "mmlu", "term": "MMLU", "aliases": ["Massive Multitask Language Understanding"], "category": "Benchmarks", "definition": "MMLU (Massive Multitask Language Understanding) is a benchmark of four-option multiple-choice questions across 57 subjects, from elementary mathematics to law and medicine, that measures a model's breadth of knowledge and problem solving.", "description": "It was the standard headline score for general knowledge for several years and is traditionally reported with five examples in the prompt (5-shot). Frontier models now score near the top, so it is considered saturated, and it has largely been succeeded by harder variants such as MMLU-Pro.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["mmlu-pro", "gpqa", "benchmark-saturation", "few-shot-prompting"], "introduced": 2020, "sources": [{"title": "Measuring Massive Multitask Language Understanding", "authors": "Hendrycks et al.", "year": 2020, "url": "https://arxiv.org/abs/2009.03300"}], "url": "https://protologue.com/t/mmlu/", "citation": "Protologue. (2026). MMLU. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0187). https://protologue.com/t/mmlu/"} | |
| {"id": "PTL-0188", "slug": "mmlu-pro", "term": "MMLU-Pro", "aliases": [], "category": "Benchmarks", "definition": "MMLU-Pro is a harder, reasoning-focused successor to MMLU that offers ten answer options instead of four and replaces trivial or noisy questions with more challenging ones.", "description": "More options lower the score obtainable by guessing, and the harder questions make chain-of-thought prompting matter more than it does on MMLU. Its creators also reported that scores vary less with the wording of the prompt than on the original benchmark.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["mmlu", "gpqa", "chain-of-thought", "prompt-sensitivity"], "introduced": 2024, "sources": [{"title": "MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark", "authors": "Wang et al.", "year": 2024, "url": "https://arxiv.org/abs/2406.01574"}], "url": "https://protologue.com/t/mmlu-pro/", "citation": "Protologue. (2026). MMLU-Pro. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0188). https://protologue.com/t/mmlu-pro/"} | |
| {"id": "PTL-0189", "slug": "gpqa", "term": "GPQA", "aliases": ["GPQA Diamond", "Graduate-Level Google-Proof Q&A"], "category": "Benchmarks", "definition": "GPQA (Graduate-Level Google-Proof Q&A) is a benchmark of difficult multiple-choice questions in biology, physics, and chemistry, written by domain experts so that skilled non-experts cannot answer them reliably even with unrestricted web access.", "description": "Its high-quality \"Diamond\" subset is the version most often reported in model announcements. Because the questions resist search, it measures expert-level reasoning more than retrieval. It is also used to study scalable oversight, the question of how people can supervise models that know more than they do.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["mmlu-pro", "humanitys-last-exam", "reasoning-model"], "introduced": 2023, "sources": [{"title": "GPQA: A Graduate-Level Google-Proof Q&A Benchmark", "authors": "Rein et al.", "year": 2023, "url": "https://arxiv.org/abs/2311.12022"}], "url": "https://protologue.com/t/gpqa/", "citation": "Protologue. (2026). GPQA. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0189). https://protologue.com/t/gpqa/"} | |
| {"id": "PTL-0190", "slug": "humanitys-last-exam", "term": "Humanity's Last Exam", "aliases": ["HLE"], "category": "Benchmarks", "definition": "Humanity's Last Exam (HLE) is a benchmark of 2,500 expert-written questions across dozens of academic subjects, some with images, designed to remain hard for frontier models after earlier academic benchmarks saturated.", "description": "Questions were contributed by subject experts and kept only if leading models failed them at the time, and answers are short and automatically checkable. Its creators also measure calibration, how well a model's stated confidence matches its accuracy, because models often answer the questions they get wrong with high confidence.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["benchmark-saturation", "gpqa", "confidence-calibration"], "introduced": 2025, "sources": [{"title": "Humanity's Last Exam", "authors": "Phan et al.", "year": 2025, "url": "https://arxiv.org/abs/2501.14249"}], "url": "https://protologue.com/t/humanitys-last-exam/", "citation": "Protologue. (2026). Humanity's Last Exam. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0190). https://protologue.com/t/humanitys-last-exam/"} | |
| {"id": "PTL-0191", "slug": "big-bench-hard", "term": "BIG-Bench Hard", "aliases": ["BBH", "BIG-bench"], "category": "Benchmarks", "definition": "BIG-Bench Hard (BBH) is a set of 23 challenging tasks from the BIG-bench collection on which earlier language models performed below average human raters, used to measure multi-step reasoning.", "description": "Its creators showed that chain-of-thought prompting greatly improved results on these tasks, which made BBH a standard test of reasoning-elicitation techniques. BIG-bench itself is a collaborative collection of over 200 tasks.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["chain-of-thought", "few-shot-prompting"], "introduced": 2022, "sources": [{"title": "Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them", "authors": "Suzgun et al.", "year": 2022, "url": "https://arxiv.org/abs/2210.09261"}, {"title": "Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models", "authors": "Srivastava et al.", "year": 2022, "url": "https://arxiv.org/abs/2206.04615"}], "url": "https://protologue.com/t/big-bench-hard/", "citation": "Protologue. (2026). BIG-Bench Hard. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0191). https://protologue.com/t/big-bench-hard/"} | |
| {"id": "PTL-0192", "slug": "arc-agi", "term": "ARC-AGI", "aliases": ["ARC", "Abstraction and Reasoning Corpus", "ARC-AGI-2", "ARC Prize"], "category": "Benchmarks", "definition": "ARC-AGI (the Abstraction and Reasoning Corpus) is a benchmark of visual grid puzzles in which a model must infer a transformation rule from a few examples and apply it to a new grid, designed to measure how efficiently a system acquires new skills.", "description": "The puzzles are easy for most people but require reasoning that cannot be memorized, because each task is novel. François Chollet introduced it in 2019 alongside a definition of intelligence as skill-acquisition efficiency. Successive versions, such as ARC-AGI-2, were made harder as systems improved, and many results are reported with explicit compute costs.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["test-time-compute-scaling", "reasoning-model"], "introduced": 2019, "sources": [{"title": "On the Measure of Intelligence", "authors": "Chollet", "year": 2019, "url": "https://arxiv.org/abs/1911.01547"}, {"title": "ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems", "authors": "Chollet et al.", "year": 2025, "url": "https://arxiv.org/abs/2505.11831"}], "url": "https://protologue.com/t/arc-agi/", "citation": "Protologue. (2026). ARC-AGI. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0192). https://protologue.com/t/arc-agi/"} | |
| {"id": "PTL-0193", "slug": "simpleqa", "term": "SimpleQA", "aliases": [], "category": "Benchmarks", "definition": "SimpleQA is a benchmark of short, fact-seeking questions with a single indisputable answer, collected adversarially against strong models, that measures whether a model answers factual questions correctly and declines when it does not know.", "description": "Answers are graded as correct, incorrect, or not attempted, so it rewards knowing what you don't know as well as knowing facts. It is used to measure hallucination in closed-book settings and the effect of adding search.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["hallucination", "confidence-calibration", "web-search-tool", "truthfulqa"], "introduced": 2024, "sources": [{"title": "Measuring short-form factuality in large language models", "authors": "Wei et al.", "year": 2024, "url": "https://arxiv.org/abs/2411.04368"}], "url": "https://protologue.com/t/simpleqa/", "citation": "Protologue. (2026). SimpleQA. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0193). https://protologue.com/t/simpleqa/"} | |
| {"id": "PTL-0194", "slug": "truthfulqa", "term": "TruthfulQA", "aliases": [], "category": "Benchmarks", "definition": "TruthfulQA is a benchmark of questions crafted so that some people would answer them falsely because of common misconceptions, used to measure whether a model repeats popular falsehoods instead of answering truthfully.", "description": "Its creators found that larger models of the time were often less truthful, because they imitated human misconceptions in their training data more faithfully. It remains a common reference for truthfulness, though, like other older public benchmarks, it is exposed to data contamination.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["hallucination", "sycophancy", "simpleqa"], "introduced": 2021, "sources": [{"title": "TruthfulQA: Measuring How Models Mimic Human Falsehoods", "authors": "Lin et al.", "year": 2021, "url": "https://arxiv.org/abs/2109.07958"}], "url": "https://protologue.com/t/truthfulqa/", "citation": "Protologue. (2026). TruthfulQA. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0194). https://protologue.com/t/truthfulqa/"} | |
| {"id": "PTL-0195", "slug": "hellaswag", "term": "HellaSwag", "aliases": [], "category": "Benchmarks", "definition": "HellaSwag is a commonsense reasoning benchmark in which a model must pick the most plausible continuation of a description of an everyday situation from four options, with wrong options generated and filtered to fool models.", "description": "When it was introduced, the questions were trivial for people but hard for the best models. Modern models score close to human level, so it is a standard example of a saturated benchmark, though it still appears in model reports.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["benchmark-saturation"], "introduced": 2019, "sources": [{"title": "HellaSwag: Can a Machine Really Finish Your Sentence?", "authors": "Zellers et al.", "year": 2019, "url": "https://arxiv.org/abs/1905.07830"}], "url": "https://protologue.com/t/hellaswag/", "citation": "Protologue. (2026). HellaSwag. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0195). https://protologue.com/t/hellaswag/"} | |
| {"id": "PTL-0196", "slug": "mmmu", "term": "MMMU", "aliases": [], "category": "Benchmarks", "definition": "MMMU (Massive Multi-discipline Multimodal Understanding) is a benchmark of college-level questions that require reasoning over images, such as charts, diagrams, tables, and chemical structures, together with text, across 30 subjects in six disciplines.", "description": "It tests whether multimodal models can combine visual perception with domain knowledge, not only describe images. It is a standard headline score for vision-language models.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["multimodal-prompting", "mmlu"], "introduced": 2023, "sources": [{"title": "MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI", "authors": "Yue et al.", "year": 2023, "url": "https://arxiv.org/abs/2311.16502"}], "url": "https://protologue.com/t/mmmu/", "citation": "Protologue. (2026). MMMU. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0196). https://protologue.com/t/mmmu/"} | |
| {"id": "PTL-0197", "slug": "gsm8k", "term": "GSM8K", "aliases": ["Grade School Math 8K"], "category": "Benchmarks", "definition": "GSM8K is a benchmark of about 8,500 grade-school math word problems that each take several steps of basic arithmetic to solve, used to measure multi-step mathematical reasoning.", "description": "It was the main testbed for chain-of-thought, self-consistency, and many other reasoning techniques, and was introduced together with the idea of training verifiers to check solutions. Frontier models now solve nearly all of it, so it is saturated, but variants that add irrelevant information or change numbers still reveal brittleness.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["chain-of-thought", "self-consistency", "irrelevant-context-distraction", "math-benchmark"], "introduced": 2021, "sources": [{"title": "Training Verifiers to Solve Math Word Problems", "authors": "Cobbe et al.", "year": 2021, "url": "https://arxiv.org/abs/2110.14168"}], "url": "https://protologue.com/t/gsm8k/", "citation": "Protologue. (2026). GSM8K. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0197). https://protologue.com/t/gsm8k/"} | |
| {"id": "PTL-0198", "slug": "math-benchmark", "term": "MATH", "aliases": ["MATH dataset", "MATH-500"], "category": "Benchmarks", "definition": "MATH is a benchmark of 12,500 competition mathematics problems, spanning seven subjects and five difficulty levels, each with a full step-by-step solution.", "description": "It is much harder than grade-school arithmetic and was long a key measure of mathematical reasoning. A 500-problem subset, MATH-500, created for research on process supervision, is the version most often reported today.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["gsm8k", "aime", "process-reward-model"], "introduced": 2021, "sources": [{"title": "Measuring Mathematical Problem Solving With the MATH Dataset", "authors": "Hendrycks et al.", "year": 2021, "url": "https://arxiv.org/abs/2103.03874"}, {"title": "Let's Verify Step by Step", "authors": "Lightman et al.", "year": 2023, "url": "https://arxiv.org/abs/2305.20050"}], "url": "https://protologue.com/t/math-benchmark/", "citation": "Protologue. (2026). MATH. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0198). https://protologue.com/t/math-benchmark/"} | |
| {"id": "PTL-0199", "slug": "aime", "term": "AIME", "aliases": ["AIME 2024", "AIME 2025", "American Invitational Mathematics Examination"], "category": "Benchmarks", "definition": "AIME, the American Invitational Mathematics Examination, is a US high-school mathematics competition whose problems, each with an integer answer, are widely used as a benchmark for reasoning models, usually identified by year, as in AIME 2024 or AIME 2025.", "description": "Each exam has only 15 problems, so scores are typically averaged over many sampled attempts and vary noticeably between runs. A new exam each year provides problems written after many models' training data was collected, which limits contamination.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["math-benchmark", "reasoning-model", "test-time-compute-scaling", "self-consistency"], "introduced": null, "sources": [{"title": "Learning to reason with LLMs", "authors": "OpenAI", "year": 2024, "url": "https://openai.com/index/learning-to-reason-with-llms/"}, {"title": "s1: Simple test-time scaling", "authors": "Muennighoff et al.", "year": 2025, "url": "https://arxiv.org/abs/2501.19393"}], "url": "https://protologue.com/t/aime/", "citation": "Protologue. (2026). AIME. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0199). https://protologue.com/t/aime/"} | |
| {"id": "PTL-0200", "slug": "humaneval", "term": "HumanEval", "aliases": [], "category": "Benchmarks", "definition": "HumanEval is a benchmark of 164 hand-written Python programming problems, each with a function signature, a description, and hidden unit tests, used to measure code generation with the pass@k metric.", "description": "It established functional correctness, whether generated code passes tests, as the standard for evaluating code models, rather than similarity to a reference solution. It is now saturated and small, so harder and contamination-resistant benchmarks such as LiveCodeBench and SWE-bench have taken its place.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["pass-at-k", "livecodebench", "swe-bench"], "introduced": 2021, "sources": [{"title": "Evaluating Large Language Models Trained on Code", "authors": "Chen et al.", "year": 2021, "url": "https://arxiv.org/abs/2107.03374"}], "url": "https://protologue.com/t/humaneval/", "citation": "Protologue. (2026). HumanEval. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0200). https://protologue.com/t/humaneval/"} | |
| {"id": "PTL-0201", "slug": "livecodebench", "term": "LiveCodeBench", "aliases": [], "category": "Benchmarks", "definition": "LiveCodeBench is a coding benchmark that continuously collects new problems from programming contests, such as LeetCode, AtCoder, and Codeforces, so that models can be evaluated only on problems published after their training data was collected.", "description": "Dating each problem lets evaluators detect and avoid data contamination. Besides code generation, it covers self-repair, predicting the output of code, and predicting test outputs.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["data-contamination", "humaneval", "pass-at-k"], "introduced": 2024, "sources": [{"title": "LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code", "authors": "Jain et al.", "year": 2024, "url": "https://arxiv.org/abs/2403.07974"}], "url": "https://protologue.com/t/livecodebench/", "citation": "Protologue. (2026). LiveCodeBench. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0201). https://protologue.com/t/livecodebench/"} | |
| {"id": "PTL-0202", "slug": "swe-bench", "term": "SWE-bench", "aliases": ["SWE-bench Verified"], "category": "Benchmarks", "definition": "SWE-bench is a benchmark of real GitHub issues from popular Python repositories, in which a model or agent must edit the codebase so that tests written for the actual fix pass.", "description": "It measures practical software engineering in large, unfamiliar codebases rather than isolated functions, and is the standard benchmark for coding agents. Because some original tasks were underspecified or had faulty tests, a 500-task subset screened by human developers, SWE-bench Verified, is the version most often reported.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["coding-agent", "humaneval", "terminal-bench"], "introduced": 2023, "sources": [{"title": "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?", "authors": "Jimenez et al.", "year": 2023, "url": "https://arxiv.org/abs/2310.06770"}, {"title": "Introducing SWE-bench Verified", "authors": "OpenAI", "year": 2024, "url": "https://openai.com/index/introducing-swe-bench-verified/"}], "url": "https://protologue.com/t/swe-bench/", "citation": "Protologue. (2026). SWE-bench. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0202). https://protologue.com/t/swe-bench/"} | |
| {"id": "PTL-0203", "slug": "terminal-bench", "term": "Terminal-Bench", "aliases": [], "category": "Benchmarks", "definition": "Terminal-Bench is a benchmark of hard, realistic tasks that an AI agent must complete by running commands in a containerized Linux terminal, graded by automated tests of the final state of the environment.", "description": "Grading checks outcomes, not the commands used, so agents can solve tasks any way that works. It measures the command-line skills that coding and DevOps agents rely on, such as building software, configuring systems, and debugging.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["coding-agent", "swe-bench", "code-execution-tool"], "introduced": null, "sources": [{"title": "Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces", "authors": "Merrill et al.", "year": 2026, "url": "https://arxiv.org/abs/2601.11868"}], "url": "https://protologue.com/t/terminal-bench/", "citation": "Protologue. (2026). Terminal-Bench. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0203). https://protologue.com/t/terminal-bench/"} | |
| {"id": "PTL-0204", "slug": "tau-bench", "term": "τ-bench", "aliases": ["tau-bench", "TAU-bench", "pass^k"], "category": "Benchmarks", "definition": "τ-bench is a benchmark in which an agent with API tools and a written domain policy, for example for airline or retail customer service, converses with a simulated user and is judged by whether the final database state is correct.", "description": "It tests tool use, policy following, and conversation together, as in real deployments. It introduced pass^k, the probability that an agent succeeds on all of k repeated trials of the same task, which exposes how inconsistent agents can be even when they often succeed once.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["function-calling", "ai-agent", "pass-at-k"], "introduced": 2024, "sources": [{"title": "τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains", "authors": "Yao et al.", "year": 2024, "url": "https://arxiv.org/abs/2406.12045"}], "url": "https://protologue.com/t/tau-bench/", "citation": "Protologue. (2026). τ-bench. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0204). https://protologue.com/t/tau-bench/"} | |
| {"id": "PTL-0205", "slug": "osworld", "term": "OSWorld", "aliases": [], "category": "Benchmarks", "definition": "OSWorld is a benchmark of open-ended tasks that computer-use agents must complete in real operating systems and applications, graded by scripts that check the resulting state of the computer.", "description": "Tasks span office software, web browsers, file management, and multi-application workflows. It is the standard benchmark for computer-use agents and showed a large gap between early agents and people.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["computer-use-agent", "multimodal-prompting"], "introduced": 2024, "sources": [{"title": "OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments", "authors": "Xie et al.", "year": 2024, "url": "https://arxiv.org/abs/2404.07972"}], "url": "https://protologue.com/t/osworld/", "citation": "Protologue. (2026). OSWorld. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0205). https://protologue.com/t/osworld/"} | |
| {"id": "PTL-0206", "slug": "gaia", "term": "GAIA", "aliases": [], "category": "Benchmarks", "definition": "GAIA is a benchmark for general AI assistants made of questions that are conceptually simple for people but require several steps of web browsing, file reading, multimodal understanding, and tool use to answer.", "description": "Each question has a single short answer, which makes grading exact despite the open-ended process. It contrasts with benchmarks of expert knowledge by testing whether assistants can reliably carry out ordinary multi-step tasks.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["ai-agent", "web-search-tool", "browsecomp"], "introduced": 2023, "sources": [{"title": "GAIA: a benchmark for General AI Assistants", "authors": "Mialon et al.", "year": 2023, "url": "https://arxiv.org/abs/2311.12983"}], "url": "https://protologue.com/t/gaia/", "citation": "Protologue. (2026). GAIA. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0206). https://protologue.com/t/gaia/"} | |
| {"id": "PTL-0207", "slug": "browsecomp", "term": "BrowseComp", "aliases": [], "category": "Benchmarks", "definition": "BrowseComp is a benchmark of hard-to-find but easy-to-verify questions that require an agent to search the web persistently and creatively, often across many pages, to locate an entangled piece of information.", "description": "Answers are short so grading is simple, while finding them can take a person hours. It measures the persistence and search strategy of browsing and deep-research agents rather than knowledge the model already has.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["web-search-tool", "agentic-rag", "gaia"], "introduced": 2025, "sources": [{"title": "BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents", "authors": "Wei et al.", "year": 2025, "url": "https://arxiv.org/abs/2504.12516"}], "url": "https://protologue.com/t/browsecomp/", "citation": "Protologue. (2026). BrowseComp. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0207). https://protologue.com/t/browsecomp/"} | |
| {"id": "PTL-0208", "slug": "ifeval", "term": "IFEval", "aliases": ["Instruction-Following Eval"], "category": "Benchmarks", "definition": "IFEval (Instruction-Following Eval) is a benchmark of prompts containing instructions whose compliance can be checked automatically, such as \"write at least 300 words\" or \"do not use any commas\", used to measure instruction following without a human or model judge.", "description": "Verifiable instructions make scoring objective and cheap, which is why IFEval is a common headline score for instruction-tuned models. It measures adherence to format and content constraints, not answer quality.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["instruction-following", "mt-bench"], "introduced": 2023, "sources": [{"title": "Instruction-Following Evaluation for Large Language Models", "authors": "Zhou et al.", "year": 2023, "url": "https://arxiv.org/abs/2311.07911"}], "url": "https://protologue.com/t/ifeval/", "citation": "Protologue. (2026). IFEval. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0208). https://protologue.com/t/ifeval/"} | |
| {"id": "PTL-0209", "slug": "mt-bench", "term": "MT-Bench", "aliases": [], "category": "Benchmarks", "definition": "MT-Bench is a benchmark of 80 two-turn questions across categories such as writing, reasoning, math, and coding, in which a strong LLM acts as the judge and scores each response.", "description": "It was introduced with the systematic study of LLM-as-a-judge, which found that strong model judges agreed with human preferences about as well as people agreed with each other, and also documented judge biases. Its small size and reliance on a model judge limit how finely it can separate today's models.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["llm-as-a-judge", "llm-judge-bias", "chatbot-arena"], "introduced": 2023, "sources": [{"title": "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena", "authors": "Zheng et al.", "year": 2023, "url": "https://arxiv.org/abs/2306.05685"}], "url": "https://protologue.com/t/mt-bench/", "citation": "Protologue. (2026). MT-Bench. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0209). https://protologue.com/t/mt-bench/"} | |
| {"id": "PTL-0210", "slug": "chatbot-arena", "term": "Chatbot Arena", "aliases": ["LMArena", "Arena", "LMSYS Chatbot Arena"], "category": "Benchmarks", "definition": "Chatbot Arena is a public platform where people chat with two anonymous models side by side and vote for the better response, with the votes aggregated into a leaderboard of model ratings.", "description": "Crowdsourced preferences capture open-ended quality that fixed test sets miss, and new questions from real users resist contamination. Ratings can also reflect style, such as length and formatting, which is why style-controlled rankings are offered. The platform was later renamed, first LMArena and then Arena.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["mt-bench", "llm-judge-bias", "llm-as-a-judge"], "introduced": 2023, "sources": [{"title": "Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference", "authors": "Chiang et al.", "year": 2024, "url": "https://arxiv.org/abs/2403.04132"}], "url": "https://protologue.com/t/chatbot-arena/", "citation": "Protologue. (2026). Chatbot Arena. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0210). https://protologue.com/t/chatbot-arena/"} | |
| {"id": "PTL-0211", "slug": "ruler", "term": "RULER", "aliases": [], "category": "Benchmarks", "definition": "RULER is a synthetic long-context benchmark that extends needle-in-a-haystack retrieval with harder tasks, such as multiple needles, multi-hop tracing, and aggregation, at configurable context lengths, to measure how much of its context window a model can actually use.", "description": "Its creators found that many models performed well on simple needle retrieval but degraded well before their advertised context length on the harder tasks. It is a standard reference for a model's effective, as opposed to claimed, context size.", "example": null, "broader": ["benchmark"], "narrower": [], "related": ["needle-in-a-haystack", "context-window", "context-rot", "lost-in-the-middle"], "introduced": 2024, "sources": [{"title": "RULER: What's the Real Context Size of Your Long-Context Language Models?", "authors": "Hsieh et al.", "year": 2024, "url": "https://arxiv.org/abs/2404.06654"}], "url": "https://protologue.com/t/ruler/", "citation": "Protologue. (2026). RULER. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0211). https://protologue.com/t/ruler/"} | |
| {"id": "PTL-0212", "slug": "harness-engineering", "term": "Harness Engineering", "aliases": ["agent harness", "agent harness engineering"], "category": "Agents & Orchestration", "definition": "Harness engineering is the design of the software around a language model that turns it into an agent, including the agent loop, tools, context management, permissions and safety controls, and feedback mechanisms such as tests and linters.", "description": "The term spread in early 2026, after OpenAI described building software with Codex agents in which engineers' main work shifted from writing code to designing the environments, constraints and feedback loops the agents work within. It treats an agent as a model plus a harness, so two agents built on the same model can differ widely in reliability. Harness engineering contains context engineering, which concerns what enters the context window at each step, and adds the runtime, tooling and control surfaces around it. Coding agents such as Codex and Claude Code are the most studied examples.", "example": null, "broader": [], "narrower": [], "related": ["context-engineering", "coding-agent", "ai-agent", "agentic-workflow", "context-compaction", "guardrails", "human-in-the-loop"], "introduced": 2026, "sources": [{"title": "Harness engineering: leveraging Codex in an agent-first world", "authors": "Lopopolo", "year": 2026, "url": "https://openai.com/index/harness-engineering/"}, {"title": "Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents -- A Source-Code Study of Eleven Systems", "authors": "Barbaste et al.", "year": 2026, "url": "https://arxiv.org/abs/2609.00006"}], "url": "https://protologue.com/t/harness-engineering/", "citation": "Protologue. (2026). Harness Engineering. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0212). https://protologue.com/t/harness-engineering/"} | |
| {"id": "PTL-0213", "slug": "recursive-language-model", "term": "Recursive Language Model", "aliases": ["RLM", "RLMs", "recursive LM"], "category": "Reasoning Models & Test-Time Compute", "definition": "A recursive language model (RLM) is an inference strategy in which a long prompt is kept outside the model's context window as an external resource that the model inspects programmatically, splitting the task into parts and calling itself recursively on them.", "description": "RLMs address inputs far longer than a context window without first compressing or summarizing them. The model works in a code environment where the prompt is stored as data, reads only the parts it needs, and delegates sub-problems to recursive calls whose results it combines. The authors report handling inputs about two orders of magnitude beyond the context window and outperforming compaction and common agent scaffolds on long-context tasks at comparable cost. The approach treats long-context handling as a form of test-time compute rather than a property of the model.", "example": null, "broader": ["test-time-compute-scaling"], "narrower": [], "related": ["context-window", "context-compaction", "context-rot", "lost-in-the-middle", "code-execution-tool"], "introduced": 2025, "sources": [{"title": "Recursive Language Models", "authors": "Zhang et al.", "year": 2025, "url": "https://arxiv.org/abs/2512.24601"}], "url": "https://protologue.com/t/recursive-language-model/", "citation": "Protologue. (2026). Recursive Language Model. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0213). https://protologue.com/t/recursive-language-model/"} | |
| {"id": "PTL-0214", "slug": "agentic-context-engineering", "term": "Agentic Context Engineering", "aliases": ["ACE"], "category": "Prompt Optimization", "definition": "Agentic Context Engineering (ACE) is a method for improving a language model application by treating its context, such as a system prompt or agent memory, as an evolving playbook that a loop of generation, reflection, and curation updates incrementally from execution feedback.", "description": "ACE adapts the inputs to a model rather than its weights. Its authors identify two failure modes of earlier context-adaptation methods, brevity bias, where summaries drop domain detail, and context collapse, where repeated rewriting erodes accumulated knowledge, and address them by adding and refining individual entries instead of rewriting the whole context. It can run offline to build a system prompt or online as agent memory, and learns from natural execution feedback without labeled data.", "example": null, "broader": ["prompt-optimization"], "narrower": [], "related": ["context-engineering", "agent-memory", "reflexion", "self-refine"], "introduced": 2025, "sources": [{"title": "Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models", "authors": "Zhang et al.", "year": 2025, "url": "https://arxiv.org/abs/2510.04618"}], "url": "https://protologue.com/t/agentic-context-engineering/", "citation": "Protologue. (2026). Agentic Context Engineering. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0214). https://protologue.com/t/agentic-context-engineering/"} | |
| {"id": "PTL-0215", "slug": "sleep-time-compute", "term": "Sleep-time Compute", "aliases": ["sleep-time computation"], "category": "Reasoning Models & Test-Time Compute", "definition": "Sleep-time compute is computation a language model performs on a context before any query arrives, anticipating likely questions and precomputing useful inferences so that less computation, and lower latency, is needed when queries come.", "description": "It moves part of test-time reasoning to idle time between interactions, which suits settings where a stable context, such as a codebase or document, is queried repeatedly. The cost can be amortized across several related queries about the same context. Its authors report that the benefit depends on how predictable the queries are from the context.", "example": null, "broader": ["test-time-compute-scaling"], "narrower": [], "related": ["agent-memory", "reasoning-model", "prompt-caching"], "introduced": 2025, "sources": [{"title": "Sleep-time Compute: Beyond Inference Scaling at Test-time", "authors": "Lin et al.", "year": 2025, "url": "https://arxiv.org/abs/2504.13171"}], "url": "https://protologue.com/t/sleep-time-compute/", "citation": "Protologue. (2026). Sleep-time Compute. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.2.2, PTL-0215). https://protologue.com/t/sleep-time-compute/"} | |