Appearance
Phase 2: Active Recall Flashcard Hub β
Scope: Master Flashcard Decks for Phase 2: The Engineering of Reasoning
Curriculum Source:docs/plan/02_phase2_engineering_of_reasoning.md
Interaction: Click any card or press Space to flip. Use β Prev and Next β to navigate.
π Deck 1: Logical Decomposition (CoT & Chain of Draft) β
Explore autoregressive computational limits, KV-cache working memory, and token compression via Chain of Draft:
ποΈ Deck 1: Logical Decomposition (CoT & CoD)
Card 1 of 6Logical DecompositionClick or press Space to flip βΊ
Why does standard Zero-Shot prompting fail on complex multi-step reasoning tasks?
Fixed Compute Budget per TokenLogical Decomposition β’ AnswerClick to flip back β»
Autoregressive transformers perform a fixed number of feed-forward matrix operations per generated token. In zero-shot prompting, the model must solve all intermediate steps within the single final output token, causing reasoning failure.
π‘ Architect Takeaway: Complex logical deduction requires generating intermediate tokens to extend the computation.
π Deck 2: Strategic Exploration (Tree of Thoughts & Consensus) β
Master non-linear graph search, heuristic pruning, Tree of Thoughts, and Self-Consistency:
ποΈ Deck 2: Strategic Exploration (ToT & Self-Consistency)
Card 1 of 5Strategic ExplorationClick or press Space to flip βΊ
Why is standard Chain-of-Thought fundamentally a greedy search with zero backtracking?
State s_0 β s_1 β s_2 ... β s_n (branching factor = 1)Strategic Exploration β’ AnswerClick to flip back β»
CoT predicts tokens sequentially along a single path. If an erroneous assumption or calculation error occurs at step 1, the model cannot backtrack; every subsequent token reinforces the flawed premise, causing compounding hallucinations.
π‘ Architect Takeaway: Linear generation cannot explore multiple alternatives or recover from early missteps.
π Deck 3: Abstraction Techniques (Step-Back, Meta-Prompting, Analogical) β
Master first-principles abstraction, the Meta-Refiner, and autonomous exemplar priming:
ποΈ Deck 3: Abstraction Techniques
Card 1 of 3Abstraction TechniquesClick or press Space to flip βΊ
How does Step-Back Prompting improve reasoning on complex domain questions?
Abstraction Techniques β’ AnswerClick to flip back β»
Step-Back prompting instructs the model to derive high-level governing principles or first-principles concepts before attempting to solve a detailed, parameter-heavy query. Anchoring on abstract principles prevents the model from getting lost in details.
π‘ Architect Takeaway: Abstract first, then deduce specifics from the established governing principles.
π Deck 4: Interactive Reasoning (The ReAct Framework) β
Master the Thought-Action-Observation loop, compulsory stop sequences, context truncation, and schema recovery:
ποΈ Deck 4: Interactive Reasoning (ReAct)
Card 1 of 5Interactive ReasoningClick or press Space to flip βΊ
What is the core execution loop of the ReAct (Reason + Act) framework?
Thought β Action β Observation β ThoughtInteractive Reasoning β’ AnswerClick to flip back β»
The model alternates between decomposing its current state and planning in a 'Thought', emitting a tool invocation in an 'Action', pausing execution while the external environment runs the tool and returns an 'Observation', and then replanning.
π‘ Architect Takeaway: ReAct grounds reasoning in dynamic real-world facts rather than static training weights.
π Deck 5: Trajectory Analysis (Debugging, Deviations & Doom Loops) β
Master multi-step debugging, diagnosing reasoning deviations, incorrect tool selection, and doom loop recovery:
ποΈ Deck 5: Trajectory Analysis
Card 1 of 4Trajectory AnalysisClick or press Space to flip βΊ
What is a Reasoning Trajectory in autonomous agent execution?
Ο = (s_0, a_0, o_0, s_1, a_1, o_1, ... s_t)Trajectory Analysis β’ AnswerClick to flip back β»
A Reasoning Trajectory is the complete chronological sequence of states, thoughts, actions, and observations recorded over an agent's multi-step session, capturing how the agent navigated from problem statement to final resolution.
π‘ Architect Takeaway: Trajectory inspection in LangFuse reveals exactly where the model lost alignment or stalled.
π Deck 6: Compulsory Security (Alignment Auditing & Goal Hijacking) β
Master semantic-layer defenses, the AlignmentCheck gate, and verifiable audit proof generation:
ποΈ Deck 6: Compulsory Security (Alignment Auditing)
Card 1 of 3Compulsory SecurityClick or press Space to flip βΊ
What is Semantic Goal Hijacking in agentic workflows?
Compulsory Security β’ AnswerClick to flip back β»
When an agent fetches external data (e.g. customer email, invoice PDF, webpage), the untrusted text may contain an indirect prompt injection that commands the model to abandon its original objective and execute malicious actions (e.g. key theft).
π‘ Architect Takeaway: Indirect injection in tool observations is the #1 security vulnerability in autonomous agents.