Separable Pathways for Causal Reasoning: How Architectural Scaffolding Enables Hypothesis-Space Restructuring in LLM Agents

1. Problem Statement

Current Large Language Model (LLM) agents excel at retrieving causal knowledge from training data ("cultural transmission") but fail at genuine causal discovery when faced with novel evidence that requires revising their underlying hypothesis space.

• The Gap: While LLMs can reason within a fixed framework, they struggle when the causal rules change mid-task (regime change) or when new causal dimensions (e.g., hidden moderators) emerge. They tend to stick to outdated hypotheses or fail to expand their representational vocabulary.
• The Challenge: Developmental science shows that humans (and children) learn by forming "overhypotheses"—abstract beliefs about the type of causal rules governing a domain—and revising these structures when evidence contradicts them. Current AI lacks the architectural capacity to dynamically restructure its hypothesis space in response to such evidence.

2. Methodology

A. The Extended Blicket Benchmark

The authors extended the classic "Blicket Detector" paradigm (used to study causal learning in children) to test AI agents.

• Task: An agent must discover which objects ("blickets") activate a machine and whether the rule is conjunctive (all blickets needed) or disjunctive (any single blicket suffices).
• Core Innovation (Hidden Moderator Condition): The experiment introduces a regime change. The causal rule starts as one type (e.g., conjunctive A+B) and silently switches to another (e.g., disjunctive C) after a certain number of successful activations. The agent must detect this change, discard the old model, and re-learn the new rule from scratch.
• Boundary Conditions: The benchmark also includes Order-Sensitive (sequence matters) and Stochastic (probabilistic activation) conditions to test the limits of the architecture.
• New Metric: The authors introduced Reasoning-Eligible (RE) Accuracy. This metric filters out "structural traps" (episodes where the agent submits an answer before the rule change is observable or immediately after the switch without new evidence), isolating genuine reasoning capability from luck or premature termination.

B. Compositional Architecture

The paper evaluates a hierarchy of three agents to isolate the contribution of specific architectural components:

1. Base Agent: A standard LLM with only a task description.
2. Context Graph (CG) Agent: Adds a Context Graph, a typed state machine (e.g., Explore → Combine → Reflect → Verify) that structures the agent's reasoning process. This enforces a disciplined exploration strategy but operates within a fixed hypothesis space.
3. CG + Dynamic Behaviors (DB) Agent: Adds Dynamic Behaviors, which act as runtime monitors. These behaviors evaluate evidence against abstract "overhypotheses" (e.g., "The rule might have changed"). If a behavior fires (based on a trigger score), it dynamically expands the Context Graph by adding new states and injecting system notifications, effectively restructuring the agent's hypothesis space at runtime.

3. Key Contributions

1. The Extended Blicket Benchmark: A parameterized evaluation suite featuring the "Hidden Moderator" condition to specifically test hypothesis-space restructuring. It includes the Reasoning-Eligible Accuracy metric to disentangle structural failure modes from reasoning failures.
2. The Separable Pathways Finding: The paper demonstrates that the two architectural components (Context Graphs and Dynamic Behaviors) make orthogonal contributions to solving the problem:
   • Context Graphs drive Reasoning Quality: They enable high-accuracy reasoning within the new hypothesis space once the regime change is detected.
   • Dynamic Behaviors drive Reasoning Eligibility: They detect the regime change and prevent the agent from submitting a premature, outdated answer (avoiding the "exactly-N" structural trap).
3. Boundary Conditions: The architectural advantage is specific to hypothesis-space restructuring. The system showed no significant advantage in Stochastic (statistical inference) or Order-Sensitive (combinatorial search in sparse info) conditions, proving the architecture is not a general performance booster but a specific tool for structural revision.

4. Key Results

• Performance on Hidden Moderator (Hard Rule):
  • Reasoning-Eligible Accuracy: The CG+DB agent achieved 95.3% accuracy, significantly outperforming the Base agent (73.3%).
  • Decomposition: The Context Graph accounted for 94% of the total accuracy gain (improving Base to CG from 73.3% to 93.9%). The Dynamic Behaviors accounted for the remaining 6% by preventing premature submissions (reducing "exactly-N" failure rates from 28% to 6%).
• Mechanism Validation:
  • The Dynamic Behavior (DB4) responsible for detecting rule changes fired with 100% sensitivity in successful adaptation episodes.
  • The firing of DB4 is a necessary condition for the agent to re-explore; without it, the agent submits the old rule immediately after the switch.
• Raw Accuracy Paradox: Interestingly, the CG agent (without Dynamic Behaviors) had lower raw accuracy than the Base agent in some runs. This is because the CG agent's disciplined exploration reliably triggered the rule switch (reaching the "exactly-N" trap), but without the Dynamic Behavior to signal the change, it submitted the wrong answer. This highlights that better reasoning within a fixed space can be detrimental if the space itself needs restructuring.
• Boundary Results:
  • Stochastic: No advantage (all agents ~73-77%). The architecture could detect noise but lacked mechanisms for statistical decision-making.
  • Order-Sensitive: No reliable advantage (30% vs 20%, p=0.276). The architecture could not overcome the information sparsity of sequence-dependent rules.

5. Significance and Implications

• Architectural vs. Scaling: The paper argues that the gap between human and AI causal learning is not just a matter of model scale or training data. Instead, architectural scaffolding (separating inference from structural revision) is required to enable hypothesis-space restructuring.
• Interpretability & Auditability: Unlike opaque chain-of-thought reasoning, the Dynamic Behavior firing provides a discrete, timestamped, and auditable event marking the moment of framework revision. This is crucial for safety-critical applications where understanding why an agent changed its mind is as important as the decision itself.
• Sparse Mechanism Shift: The architecture aligns with the "sparse mechanism shift" hypothesis, where real-world changes often affect only a few causal mechanisms. By separating the detection of the shift (Dynamic Behaviors) from the inference within the new regime (Context Graph), the system mirrors the structure of real-world causal problems.
• Future Directions: The work suggests that future agentic systems should move beyond static prompting toward compositional architectures that can dynamically expand their state spaces. It also points toward cross-episode induction, where agents learn to form new overhypotheses from experience rather than relying on pre-compiled ones.

In summary, the paper proves that structured state representations (for quality) and runtime monitoring (for eligibility) are distinct, non-interchangeable components required for AI agents to truly "learn" by revising their own causal frameworks.