{ "name": "Act-onomy Taxonomy", "version": "v1.0", "derived_from": "data/corpus/act-onomy_codebook.csv (V4.2)", "tier_names": [ "Action", "Subaction", "Instance" ], "summary": { "tier_1": 10, "tier_2": 46, "tier_3": 120 }, "taxonomy": [ { "code": "T1.1", "name": "Retrieval", "subactions": [ { "code": "T1.1.1", "name": "Retrieve from skill library", "instances": [ { "code": "T1.1.1.1", "name": "Retrieve from skill library", "definition": "Grab pre-built skill snippet (Minecraft “chop tree”)" } ] }, { "code": "T1.1.2", "name": "Retrieve from local corpus", "instances": [ { "code": "T1.1.2.1", "name": "Retrieve from local corpus", "definition": "Diverse filetype reading; read between 1 to 15 documents and/or tables”" } ] }, { "code": "T1.1.3", "name": "Retrieve from external knowledge base", "instances": [ { "code": "T1.1.3.1", "name": "Retrieve from external knowledge base", "definition": "Recall three relevant distinct problems (P28): “Recall three (03) relevant and distinct problems (different from the user task)”" } ] }, { "code": "T1.1.4", "name": "Retrieve from open web", "instances": [ { "code": "T1.1.4.1", "name": "Retrieve from open web", "definition": "High reliance on web search to offload knowledge retrieval to external sources\"" } ] }, { "code": "T1.1.5", "name": "Retrieve relevant context", "instances": [ { "code": "T1.1.5.1", "name": "Retrieve relevant context", "definition": "LLM leverages retrieved HLS-related context to transform input C algorithms or natural language descriptions”" } ] } ] }, { "code": "T1.2", "name": "Memory", "subactions": [ { "code": "T1.2.1", "name": "Store Information", "instances": [ { "code": "T1.2.1.1", "name": "Store information in working memory", "definition": "Scratchpad for current inputs and intermediate results." }, { "code": "T1.2.1.2", "name": "Store episodic trajectories", "definition": "Record full action sequences for later training." }, { "code": "T1.2.1.3", "name": "Store knowledge in semantic memory", "definition": "Save general world facts." }, { "code": "T1.2.1.4", "name": "Store experiences in episodic memory", "definition": "Save events as episodes." }, { "code": "T1.2.1.5", "name": "Store information in long-term memory", "definition": "Store information in external storage." }, { "code": "T1.2.1.6", "name": "Maintain curriculum library", "definition": "Keep a syllabus of mastered and upcoming skills, arranged by difficulty, so the agent learns in a sensible order." } ] }, { "code": "T1.2.2", "name": "Update Information", "instances": [ { "code": "T1.2.2.1", "name": "Update memory", "definition": "Update experiences." } ] }, { "code": "T1.2.3", "name": "Discard information", "instances": [ { "code": "T1.2.3.1", "name": "Discard information from working memory", "definition": "Discard local memory after subtask completion." }, { "code": "T1.2.3.2", "name": "Discard redundant game state, keep summaries", "definition": "Automatic context compaction to manage thousands of reasoning steps, preserving only LLM responses and action summaries while discarding redundant game state”" } ] }, { "code": "T1.2.4", "name": "Consolidate memory", "instances": [ { "code": "T1.2.4.1", "name": "Consolidate working memory into long-term memory", "definition": "Retain only successful analysis path to global memory." }, { "code": "T1.2.4.2", "name": "Compact context window", "definition": "Automatic context compaction to manage the thousands of reasoning steps required, preserving only LLM responses and action summaries while discarding redundant game state\"" } ] }, { "code": "T1.2.5", "name": "Read memory", "instances": [ { "code": "T1.2.5.1", "name": "Read from working memory", "definition": "Grab intermediate results from scratchpad." }, { "code": "T1.2.5.2", "name": "Read from long-term memory", "definition": "Importance-weighted retrieval of discoveries." } ] } ] }, { "code": "T2.1", "name": "Planning", "subactions": [ { "code": "T2.1.1", "name": "Decompose task", "instances": [ { "code": "T2.1.1.1", "name": "Decompose into subtasks", "definition": "Todo_write tool (15 invocations by GPT-5.2)" }, { "code": "T2.1.1.2", "name": "Decompose into subgoals with success conditions", "definition": "LLM decomposes the task into a sequence of subgoals, each paired with an executable success-condition function”" }, { "code": "T2.1.1.3", "name": "Decompose by role specialization", "definition": "Splitting the instruction-following challenge into distinct reasoning and grounding roles handled by a reasoner agent and an actor agent\"" } ] }, { "code": "T2.1.2", "name": "Formulate a workflow or plan", "instances": [ { "code": "T2.1.2.1", "name": "Formulate a high-level plan", "definition": "METRO starts with the identification of the strategic action a_i for every expert utterance u_i in transcript D\"" }, { "code": "T2.1.2.2", "name": "Formulate an analysis workflow", "definition": "The Planner accurately interprets user intent and formulates a comprehensive analysis workflow”" }, { "code": "T2.1.2.3", "name": "Plan navigation through environment", "definition": "Planning to navigate the web; tasks require navigation through an average of 4.2 web pages”" }, { "code": "T2.1.2.4", "name": "Plan function or tool use", "definition": "I can use the pandas method corr() to calculate the Pearson correlation coefficient\"" }, { "code": "T2.1.2.5", "name": "Plan Code or artifact structure", "definition": "Planning: HTML Structure, CSS Styling, JavaScript Functionality\"" }, { "code": "T2.1.2.6", "name": "Formulate plan from template", "definition": "HLSTuner formulates a detailed plan that specifies: (1) the combination of HLS directives, (2) target code segments, and (3) the insertion actions\"" } ] }, { "code": "T2.1.3", "name": "Select Strategy", "instances": [ { "code": "T2.1.3.1", "name": "Select among candidate strategies", "definition": "Selects effective HLS directive combination strategies and inserts directives within the specific structure (e.g., loops and arrays)”" }, { "code": "T2.1.3.2", "name": "Switch to fallback strategy", "definition": "Incorporating a dummy score prediction as a fallback mechanism\"" } ] }, { "code": "T2.1.4", "name": "Modify Plan", "instances": [ { "code": "T2.1.4.1", "name": "Replan dynamically based on feedback", "definition": "RequireReplan provides the LLM-based executor with the capability to dynamically adjust the plan\"" }, { "code": "T2.1.4.2", "name": "Refine Requirements", "definition": "Refine requirement based on validated test cases to ensure alignment”" } ] } ] }, { "code": "T2.2", "name": "Reasoning", "subactions": [ { "code": "T2.2.1", "name": "Generating", "instances": [ { "code": "T2.2.1.1", "name": "Generate candidate options", "definition": "Brainstorm possible next moves." }, { "code": "T2.2.1.2", "name": "Generate structured artifacts", "definition": "Each agent then emits a set of initial codes; cluster semantically similar codes and generate preliminary themes\"" }, { "code": "T2.2.1.3", "name": "Generate evaluations", "definition": "LLM Group for multifaceted evaluation generates diverse debugging instructions”" } ] }, { "code": "T2.2.2", "name": "Analysing", "instances": [ { "code": "T2.2.2.1", "name": "Analyse artifact structure and behavior", "definition": "Analyze frontend framework and interactive components”" }, { "code": "T2.2.2.2", "name": "Detect patterns or trends in data", "definition": "Identifying patterns, directions, or changes in data over time or across contexts”" }, { "code": "T2.2.2.3", "name": "Interpret meaning of artifacts", "definition": "Logical reasoning to interpret papers and code; mathematical reasoning to modify and run code; causal reasoning to infer scientific insights from results\"" }, { "code": "T2.2.2.4", "name": "Classify inputs into categories", "definition": "Delegator can determine which of the k tasks the input string belongs to by observing only a constant number of metadata tokens." } ] }, { "code": "T2.2.3", "name": "Explaining", "instances": [ { "code": "T2.2.3.1", "name": "Explain reasoning or outcomes", "definition": "Explain the failure from a user requirement perspective\"" } ] }, { "code": "T2.2.4", "name": "Summarizing/Distilling", "instances": [ { "code": "T2.2.4.1", "name": "Summarize recent observations and trajectories", "definition": "Condense recent observations and trajectories into key takeaways." } ] }, { "code": "T2.2.5", "name": "Inferring", "instances": [ { "code": "T2.2.5.1", "name": "Infer hidden state from observable evidence", "definition": "Damage calculation reveals stat distributions." }, { "code": "T2.2.5.2", "name": "Infer causal relationship", "definition": "Determine plausible causal factor." }, { "code": "T2.2.5.3", "name": "Infer structure from indirect evidence", "definition": "Agents must infer the directory layout by inspecting package structures and README files\"" }, { "code": "T2.2.5.4", "name": "Infer errors", "definition": "M_q has the capacity to identify exactly one bug at a time." } ] }, { "code": "T2.2.6", "name": "Comparing & Ranking", "instances": [ { "code": "T2.2.6.1", "name": "Compare values across sources", "definition": "Quantify and compare across sources." }, { "code": "T2.2.6.2", "name": "Rank items by criteria", "definition": "Order items by importance." } ] }, { "code": "T2.2.7", "name": "Contextualizing", "instances": [ { "code": "T2.2.7.1", "name": "Package prior reasoning as context for subsequent calls", "definition": "Feed reasoning so far as context to next LLM call." }, { "code": "T2.2.7.2", "name": "Construct structured context object", "definition": "Global interactive multi-behavior overlapping network is constructed based on the simulated behaviors of reposting and reposting with a comment, as well as following”" }, { "code": "T2.2.7.3", "name": "Configure agent persona or role-conditioning", "definition": "I want you to play as {role}, imitating {role}’s personality and values”" }, { "code": "T2.2.7.4", "name": "Assign roles in a multi-agent team", "definition": "Specifying a scope, or role, for an agent allows it to be treated as a computational object. map each one to an employee in a human organization." } ] }, { "code": "T2.2.8", "name": "Combining & Synthesis", "instances": [ { "code": "T2.2.8.1", "name": "Combine information from multiple sources", "definition": "Combine information from multiple sources... to produce a coherent solution”" }, { "code": "T2.2.8.2", "name": "Aggregate observations into a structured representation", "definition": "Supplementary semantic map M't, which aggregates the information from Mt over successive time steps. The intuition resembles a form of majority voting\"" } ] }, { "code": "T2.2.9", "name": "Filtering", "instances": [ { "code": "T2.2.9.1", "name": "Filter information by threshold", "definition": "Selecting relevant information based on criteria, quality, or thresholds”" } ] } ] }, { "code": "T2.3", "name": "Evaluate", "subactions": [ { "code": "T2.3.1", "name": "Evaluating with gold", "instances": [ { "code": "T2.3.1.1", "name": "Compare against gold reference", "definition": "We prompt LLM to review the correct HLS-C code, pairing buggy code segments with corresponding error messages to construct debugging CoT”" }, { "code": "T2.3.1.2", "name": "Score on gold criteria", "definition": "The Gold Checker evaluates each annotation on three binary criteria: equivalence, completeness, and correctness, where a score of 1 indicates the criterion is met”" } ] }, { "code": "T2.3.2", "name": "Evaluating with goals/requirements/constraints", "instances": [ { "code": "T2.3.2.1", "name": "goal-completion check", "definition": "Independently checks whether objectives are truly complete, preventing the orchestrator from advancing when the main agent incorrectly believes a task is finished”" }, { "code": "T2.3.2.2", "name": "Requirement-satisfaction check", "definition": "The final results from all subtasks are synthesized to meet the user's requirements”" }, { "code": "T2.3.2.3", "name": "Constraint / budget check", "definition": "HLSTuner enables QoR-aware reasoning to align optimization goals with hardware constraints”" }, { "code": "T2.3.2.4", "name": "Domain-rule / best-practice check", "definition": "Codified best-practices, such as the standard order of operations (e.g., quality control must precede normalization)\"" } ] }, { "code": "T2.3.3", "name": "Evaluating without ground truth", "instances": [ { "code": "T2.3.3.1", "name": "Score on quality dimensions", "definition": "We use GPT-4o as the default evaluator, each dimension scored on a scale of 0 to 2”" }, { "code": "T2.3.3.2", "name": "Check rubric compliance", "definition": "Verify the implementation of every item listed in {Rubrics}\"" }, { "code": "T2.3.3.3", "name": "Evaluate visual/behavioral correctness", "definition": "Evaluates dynamic behavior and visual correctness based on screenshots\"" }, { "code": "T2.3.3.4", "name": "Evaluate internal consistency", "definition": "Task an evaluator agent with determining whether each theme is consistent with its supporting quotes”" }, { "code": "T2.3.3.5", "name": "Evaluate intermediate results", "definition": "Agents with domain expertise verify intermediate results and reduce errors”" }, { "code": "T2.3.3.6", "name": "Provide qualitative judgement", "definition": "Upon successful generation of a result, the Evaluator agent, ALLMe assesses the outcome and provides natural language feedback if it deems revisions are necessary.”" }, { "code": "T2.3.3.7", "name": "Simulate counterfactual outcomes", "definition": "Mentally simulate an alternative scenario — either before acting (imagine 'if I do X, what happens?') or post-hoc (imagine 'what would have happened if action X had been different?'). Includes counterfactual trajectory generation for causal analysis." }, { "code": "T2.3.3.8", "name": "Validate predicted issues", "definition": "The agent assesses the contextual applicability of potential bugs, reducing the probability that the LLM forcibly generates trivial results”" } ] } ] }, { "code": "T2.4", "name": "Deciding", "subactions": [ { "code": "T2.4.1", "name": "Make a decision", "instances": [ { "code": "T2.4.1.1", "name": "Make a decision according to memory", "definition": "The agent conditions its decisions on I, the current observation st, and a history buffer of n past state-action pairs." } ] }, { "code": "T2.4.2", "name": "Pick scores", "instances": [ { "code": "T2.4.2.1", "name": "Select action by score", "definition": "Argmax = highest score; softmax = probabilistic sample; majority vote across evaluators." } ] }, { "code": "T2.4.3", "name": "Decide accept or not", "instances": [ { "code": "T2.4.3.1", "name": "Decline out-of-scope queries", "definition": "Appropriately reject queries that exceed their knowledge boundaries or conflict with their role settings”" } ] }, { "code": "T2.4.4", "name": "Decide under uncertainty", "instances": [ { "code": "T2.4.4.1", "name": "Fork trajectory at uncertainty", "definition": "Fork the current generative state and generate a new parallel trajectory from this point of high uncertainty”" } ] } ] }, { "code": "T3.1", "name": "Grounding", "subactions": [ { "code": "T3.1.1", "name": "Interact with users", "instances": [ { "code": "T3.1.1.1", "name": "Accept instructions from humans", "definition": "Receive task or command from human user." }, { "code": "T3.1.1.2", "name": "Ask for clarification from people", "definition": "Proactively ask human when unclear." }, { "code": "T3.1.1.3", "name": "Communicate task outcome natural language", "definition": "Agent informs the user: ‘Buy a nice rich navy bathing dress’." }, { "code": "T3.1.1.4", "name": "Communicate via visualization", "definition": "CellAgent generated differential expression and marker gene visualizations, enabling intuitive interpretation of cluster identities”" }, { "code": "T3.1.1.5", "name": "Communicate via structured format", "definition": "Generate outputs in a JSON format (or lists of JSON objects)”" }, { "code": "T3.1.1.6", "name": "Communicate refusal/inability", "definition": "Recognize and refuse queries that conflict with their role knowledge\"" }, { "code": "T3.1.1.7", "name": "Express tone to user", "definition": "Models exhibit a strong default tendency to excessively affirm, apologize, and express gratitude”" }, { "code": "T3.1.1.8", "name": "Disclose self-information to user", "definition": "Identity disclosure, S2S systems often proactively mention that they are intelligent assistants." } ] }, { "code": "T3.1.2", "name": "Interact with physical environments", "instances": [ { "code": "T3.1.2.1", "name": "Perceive physical environment", "definition": "Convert images to text via VLMs." }, { "code": "T3.1.2.2", "name": "Affect physical environments", "definition": "Send language commands to robot arm." } ] }, { "code": "T3.1.3", "name": "Interact with digital environments", "instances": [ { "code": "T3.1.3.1", "name": "Navigate digital interfaces", "definition": "Browse websites, scroll pages, navigate menus, move within apps." }, { "code": "T3.1.3.2", "name": "Modify digital objects", "definition": "Annotate UI Components." }, { "code": "T3.1.3.3", "name": "Issue operational commands", "definition": "Invoke structured API or data endpoint." } ] }, { "code": "T3.1.4", "name": "Interact with other agents", "instances": [ { "code": "T3.1.4.1", "name": "Monitor peer agent's state or decision", "definition": "Observe another agent's chosen action before making one's own decision — a prerequisite for oversight, override, or trust-based deference in multi-agent systems." }, { "code": "T3.1.4.2", "name": "Receive feedback from peer agent", "definition": "Evaluator provides natural-language feedback." }, { "code": "T3.1.4.3", "name": "Send message to peer agent", "definition": "Agent engages in multi-turn interactions...to generate deliverables\"" }, { "code": "T3.1.4.4", "name": "Recommend action to peer agent", "definition": "Communicate a proposed action to a peer agent for their consideration, knowing the peer may accept or override it — agent-to-agent advisory communication." }, { "code": "T3.1.4.5", "name": "Override peer agent's decision", "definition": "Replace another agent's selected action with one's own decision — the active side of human-in-the-loop oversight or hierarchical agent control." }, { "code": "T3.1.4.6", "name": "Dispatch task to sub-agent", "definition": "A central orchestrator maintains a high-level route plan while dynamically dispatching sub-agents based on game context”" }, { "code": "T3.1.4.7", "name": "Argue or debate with peer agent", "definition": "Multi-agent argumentation toward better answer." } ] }, { "code": "T3.1.5", "name": "Augment with external computation", "instances": [ { "code": "T3.1.5.1", "name": "Execute code", "definition": "Test Runner detects logic inconsistencies." }, { "code": "T3.1.5.2", "name": "Invoke specialized computation tool", "definition": "Use calculator / heavy compute tools." }, { "code": "T3.1.5.3", "name": "Invoke visual inspection tool", "definition": "Zoom-in tool for region-of-interest." } ] } ] }, { "code": "T3.2", "name": "Executing", "subactions": [ { "code": "T3.2.1", "name": "Executing plan", "instances": [ { "code": "T3.2.1.1", "name": "Execute strategy", "definition": "An insertion agent executes this plan for HLS-C optimization”" } ] }, { "code": "T3.2.2", "name": "Executing debug", "instances": [ { "code": "T3.2.2.1", "name": "Adopt debugging instructions", "definition": "Operating under strict instruction adherence, this agent adopts the instructions to implement debugging”" }, { "code": "T3.2.2.2", "name": "Rewrite code after bug fix", "definition": "Debugging specialist rewrites the entire code every time it fixes a bug." } ] }, { "code": "T3.2.3", "name": "Terminating", "instances": [ { "code": "T3.2.3.1", "name": "Provide final answer", "definition": "Final Answer: The average number of lines per scene in the ep_7.csv dataset is approximately 6.02\" (p32)" }, { "code": "T3.2.3.2", "name": "Generate refusal", "definition": "Providing clear refusal responses with appropriate explanations”" } ] } ] }, { "code": "T4.1", "name": "Learning", "subactions": [ { "code": "T4.1.1", "name": "Learning reasoning", "instances": [ { "code": "T4.1.1.1", "name": "Update reasoning via prompt update", "definition": "Rewrite own prompt template." } ] }, { "code": "T4.1.2", "name": "Learning grounding", "instances": [ { "code": "T4.1.2.1", "name": "Update grounding via code-based skills", "definition": "Improve web-navigation code snippets." }, { "code": "T4.1.2.2", "name": "Update retrieval procedures", "definition": "Improve how the agent searches for information." } ] }, { "code": "T4.1.3", "name": "Learning knowledge", "instances": [ { "code": "T4.1.3.1", "name": "Update source code as procedural memory", "definition": "Self-patch agent's own source code." }, { "code": "T4.1.3.2", "name": "Update semantic memory with knowledge", "definition": "Expand Error Repository with new mnemonic." }, { "code": "T4.1.3.3", "name": "Update memory from new experiences", "definition": "Refinement of memories (represented in text) as the system encounters new scenarios." }, { "code": "T4.1.3.4", "name": "Update hypothesis with new evidence", "definition": "Allowing the model to update its textual understanding, refine its hypothesis, or even trigger further visual exploration”" } ] }, { "code": "T4.1.4", "name": "Learning LLM parameters", "instances": [ { "code": "T4.1.4.1", "name": "Update parametric policy", "definition": "Change the model's internal weights through training." }, { "code": "T4.1.4.2", "name": "Update LLM parameters via SL/RL/RLHF", "definition": "SL/RL/RLHF weight adjustment." }, { "code": "T4.1.4.3", "name": "Update action parameters based on feedback", "definition": "Auto-adjust params from exception info." } ] }, { "code": "T4.1.5", "name": "Learning instructions", "instances": [ { "code": "T4.1.5.1", "name": "Learning instructions", "definition": "Extract rule from examples for future use." } ] } ] }, { "code": "T4.2", "name": "Reflection", "subactions": [ { "code": "T4.2.1", "name": "Reflect on errors and failures", "instances": [ { "code": "T4.2.1.1", "name": "Diagnose failure against ground truth", "definition": "Analyzes stuck states by comparing current situation against ground truth sources (porymap data, knowledge base) to diagnose navigation failures”" }, { "code": "T4.2.1.2", "name": "Inspect Error Pattern", "definition": "An inspection agent examines the erroneous code and error messages parsed from the HLS tool test results”" }, { "code": "T4.2.1.3", "name": "Analyze log to formulate fix instructions", "definition": "Reasoning-to-instruction on HLS log." }, { "code": "T4.2.1.4", "name": "Reflect on failed episodes for prior knowledge", "definition": "RequireReplan...dynamically adjust the plan, improving the robustness to exceptions\"" }, { "code": "T4.2.1.5", "name": "Self-Correct Step Implementation", "definition": "Modify step implementation based on error message”" } ] }, { "code": "T4.2.2", "name": "Reflect on self-outcomes", "instances": [ { "code": "T4.2.2.1", "name": "Self-reflect", "definition": "The results of the execution are fed back, enabling GPT-4 to refine its responses and propose further iterations\"" }, { "code": "T4.2.2.2", "name": "Reflect on proposed fix", "definition": "Reflects on the code after assuming the fix to ensure the modification is reasonable”" }, { "code": "T4.2.2.3", "name": "Pre-action self-check", "definition": "These steps are labeled as reflective, indicating the model is engaging in self-monitoring or explicit reasoning before issuing commands”" }, { "code": "T4.2.2.4", "name": "Refine Strategy Iteratively", "definition": "Iterative refinement using current QoR." } ] }, { "code": "T4.2.3", "name": "Reflect on external feedback", "instances": [ { "code": "T4.2.3.1", "name": "Receive and integrate external feedback", "definition": "The Evaluator drives a self-reflective optimization mechanism, which leverages automated evaluation methods for various analysis tasks to iteratively refine outcomes, replacing subjective manual assessments." } ] } ] } ] }