Patterns

UserQuery -> MCPToolPayload

Expose catalog schema registry queries as callable tools via the Model Context Protocol.

ShortPrompt -> DenseStructuredPrompt

Expand short scene descriptions into dense, structured, spatial-temporally rich prompts using an LLM prior to video/image generation.

Tensor[Tokens] -> Tensor[PositionEncodedTokens]

Encode spatial, temporal, and modal structures by applying a unified 3D rotary position embedding to mixed-modality tokens.

MultimodalTokens -> MultimodalTokens

Route reasoning tasks through causal self-attention and multimodal generation tasks through full-attention denoising within a shared transformer backbone.

Audio -> TranscriptWithSpeakerLabels

Segment audio using voice activity detection (FSMN-VAD), extract speaker embeddings via verification models (CAM++), transcribe audio with ASR, and align outputs to produce punctuated, speaker-attributed transcripts.

(Text, List<AudioSegment>) -> List<AudioToken>

Interleave tokenized text segments with discrete quantized audio segments to condition speech generation on conversational history.

FrameEmbedding -> List<QuantizedToken>

Predict high-order audio codebooks autoregressively within each frame using a small decoder conditioned on a base frame embedding.

TargetAgent -> DialogHistory

Simulate an adversarial dialogue loop between an attacker LLM agent and a target LLM agent to test resistance to multi-step manipulation.

PromptResponsePair -> EvaluationReport

Evaluate model responses against safety, privacy, or security rules using an LLM judge to produce a binary violation score and reasoning.

VulnerabilityType -> List<AdversarialPrompt>

Generate targeted prompt-injection or jailbreak payloads tailored to a specific vulnerability class using a generator LLM.

(LLMClaim, List<RawOutput>) -> VerifiedClaim

Filter claimed LLM milestones by checking if the exact substring exists within raw tool execution logs.

(Set<Fact>, Goal) -> Intent

Plan exploration steps by checking known objective facts against a target goal to yield unexecuted intents, eliminating redundant tool executions.

TaskDescription -> SourceCode

Generate an initial runnable scaffolding file tailored to a structured task description.

SourceCode, ExecutionLogs, BenchmarkMetrics -> ImprovedSourceCode

Refine task agent source code iteratively by analyzing previous execution logs, diff histories, and execution error metrics.

Prompt + Schema -> ParsedObject

Enforce a validation schema on model inference to guarantee structured JSON outputs.

GeneratorContext -> VerifiedOutput | ReproducibleFailure

Execute a candidate generation, then evaluate it using a verifier agent that is kept completely blind to the generation agent's context and history.

RepositoryMetadata -> LLMToolPrompt

Extract function metadata from dynamic third-party package registries to compile structural prompt instructions for LLM tool selection.

Embedding & NoisyTrajectory -> ActionTrajectory

Denoise a continuous robot action trajectory sequence using a flow-matching diffusion transformer conditioned on vision-language embeddings.

Prompt -> List<AdversarialPrompt>

Synthesize variations of test prompts using semantic manipulation operators to test LLM safety guardrails.

UserFeedback + BasePrompt -> AugmentedPrompt

Augment prompt payloads dynamically by aggregating preceding interaction feedback into the agent's context window.

GroupCompletions -> RelativeRewards

Compute relative policy advantages from group-level rewards across multiple outputs generated for a single prompt, eliminating the memory and compute overhead of a separate critic model.

WebpageDOM -> AccessibilityTree<ReferenceID>

Extract a webpage's accessibility tree, mapping interactive elements to short reference IDs for AI agent consumption.