Are traditional cognitive theories missing interaction effects between mechanisms?
This explores whether models of cognition that study mechanisms one at a time (causality, reasoning, memory, knowledge retrieval) are blind to what happens when those mechanisms collide, reinforce, or interfere with each other.
This explores whether traditional cognitive theories — which tend to isolate one mechanism at a time — miss the effects that only appear when mechanisms interact. The corpus suggests this is a recurring blind spot, and several notes converge on it from very different directions.
The most direct evidence is that single-mechanism theories openly admit their own gaps. Causal belief networks model causal reasoning well but cannot represent associative links, analogical mappings, or emotion-driven belief shifts — the framework itself treats causality as a tractable starting point rather than a full account of how people reason Can causal models alone capture how humans actually reason?. The same one-lens-isn't-enough pattern shows up in how we study the machines: representational analysis alone finds correlations without causes, and causal analysis alone shows effects without explaining them — only pairing the two produces a complete mechanistic claim Can we understand LLM mechanisms with only representational analysis?. In both cases, the interesting behavior lives in the seam between methods, not inside any one of them.
Where the corpus gets sharper is on compounding — interaction effects that aren't just additive but multiplicative. Three cognitive traps in human-AI interaction (mistaking the map for the territory, conflating intuition with reasoning, and confirmation bias) don't simply stack; they multiply each other's distorting power when they co-occur, producing epistemic drift that none would cause alone Why do people trust AI outputs they shouldn't?. That's exactly the kind of effect a mechanism-by-mechanism theory would never see.
Interactions also turn out to be destructive, not just amplifying. Knowledge retrieval sits in lower network layers and reasoning adjustment in higher ones, so training that boosts reasoning can quietly degrade knowledge-heavy domains like medicine — an interference effect invisible if you study reasoning in isolation Why does reasoning training help math but hurt medical tasks?. Planning and execution interfere similarly: pulling the decomposer apart from the solver improves accuracy precisely because it stops the two from contaminating each other Does separating planning from execution improve reasoning accuracy?, and isolating reasoning operations into modular tools elicits capability that monolithic prompting can't reach Can modular cognitive tools unlock reasoning without training?. The lesson cuts both ways — sometimes the missing variable is a harmful interaction you have to engineer apart.
What you might not expect is that the same mechanism can flip sign depending on what it interacts with. Extended 'thinking' is counterproductive in a vanilla model — it breeds self-doubt — but RL training transforms that identical mechanism into productive gap analysis, meaning reasoning quality is mediated by training, not intrinsic to the mechanism Does extended thinking help or hurt model reasoning?. And memory-amortized inference reframes cognition itself as the interaction between memory reuse and inference rather than either alone Can cognition work by reusing memory instead of recomputing?. The throughline: across human and machine cognition, the corpus keeps finding that the action is in the interactions — and theories built around single, separable mechanisms are structurally positioned to miss it.
Sources 8 notes
Causal belief networks excel at modeling causal reasoning but cannot represent associative links, analogical mappings, or emotion-driven belief shifts. The GenMinds framework itself acknowledges this as a tractable starting point rather than a complete theory.
Representational analysis alone identifies correlations without causation; causal analysis alone shows behavioral effects without explaining them. Only paired methods—locating candidate features representationally, then verifying causally—produce complete mechanistic claims.
Rose-Frame identifies map-territory confusion, intuition-reason conflation, and confirmation-bias reinforcement as traps that multiply their distorting effects when they co-occur. Evidence from cross-linguistic overreliance and architectural transformer biases confirms the compounding mechanism operates universally.
Two-phase inference model shows knowledge retrieval operates in lower network layers while reasoning adjustment happens in higher layers. This separation explains why reasoning training improves math but can degrade knowledge-intensive domains like medicine.
Modular architectures with separate decomposer and solver models outperform monolithic LLMs, with decomposition ability transferring across domains while solving ability does not. The separation prevents planning-execution interference and produces more generalizable skills.
Four cognitive tools implemented as sandboxed LLM calls improved GPT-4.1 on AIME2024 from 26.7% to 43.3% without any RL training. Modularity enforces operation isolation that pure prompting cannot guarantee, eliciting pre-existing reasoning capability.
Vanilla models use thinking mode counterproductively, inducing self-doubt that degrades performance. RL training reverses this, transforming the same mechanism into beneficial gap analysis. Training mediates reasoning quality, not just quantity.
Memory-Amortized Inference proposes intelligence arises from structured reuse of prior inference paths over topological memory, inverting RL's reward-forward logic into cause-backward reconstruction. This duality explains energy efficiency and suggests memory trajectories form the substrate of adaptive thought.