Neural Mechanism Research

Neural Mechanism Research

Explain what the computation actually makes possible, how learning discovers it, and how firmly the explanation is established. The reader should be able to reconstruct a small example and predict what would change if the design changed.

This is a neural-computation variant of deep-architecture-research combined with the reporting principles of understanding-first-report. The essential guidance is self-contained here; neither skill has to be loaded again. Use their deeper references only if the task also requires a substantial source-code audit or a separate reporting review.

Start with the real question

Quote the relevant original user wording before substantive analysis. For a long prompt, preserve the central questions and mark omissions. Reconstruct the underlying puzzle, including the user's proposed explanation, rather than replacing it with a generic survey.

Map the few causal relationships that matter. For example:

architecture / parameterization
          │
          ├── forward: what functions become possible?
          ├── geometry: what changes, mixes, or becomes separable?
          └── backward: which errors change which parameters?
                         │
data / loss / optimizer ─┘
                         ▼
              learned behavior and cost
                         │
              evidence and alternatives

If the user has provided clear scope and requested research, state the interpretation and proceed. Clarify only material ambiguity. Do not introduce a repeated confirmation gate or reopen accepted scope.

Treat meaning as a testable explanation

For the main mechanism, connect these perspectives rather than filling six disconnected sections:

• Forward computation: concrete inputs, shapes, parameters, intermediate states, outputs, and what the next component can do with them.
• Geometry and information: projections, rank, basis changes, nonlinear boundaries, mixing, routing, compression, invariances, and the distinction between coordinate count and independent information.
• Backward computation: start at the actual objective, follow the chain rule to the relevant parameters, and explain what an update rewards or suppresses.
• Learning dynamics: initialization, symmetry breaking, feature learning, gradient conflicts, saturation, conditioning, optimization noise, and dependence on the parameterization.
• Physical or dynamical interpretations: use energy, attractors, diffusion, discrete dynamics, or other physical language only where the equations justify the mapping. State which assumptions make it valid and where it breaks.
• Engineering and history: distinguish a useful inductive bias from an empirically selected size, implementation convenience, resource tradeoff, or historical convention.

Use only the perspectives that change the answer, but do not omit gradients, geometry, or historical limitations when they are central to the user's puzzle.

Explain the explanation

For each decisive claim, make this chain understandable:

ordinary-language meaning → tiny example → exact mechanism
                                   │
                       evidence / counterexample / limit

An analogy must map to actual variables or operations. “More capacity,” “richer features,” “knowledge lookup,” and “better gradients” are starting questions, not completed explanations. Explain what capacity counts, how features are selected, where a stored relation is encoded, or which derivative changes.

Prefer low prerequisite explanations without sacrificing correctness. Show real shapes alongside small examples. A two-dimensional sketch is a teaching abstraction, not a literal picture of a high-dimensional learned space.

Gather evidence across time and alternatives

For substantial research, read references/evidence-and-experiments.md. It supplies the historical comparison method, gradient checks, and causal-evidence boundaries.

Use primary papers, original implementations, released model configurations, and author research reports. Pin the version or date for decisive sources. Separate the original motivation from later explanations and present-day engineering choices.

When the mechanism has a long history, follow the chain from precursors to initial experiments, later ablations, conflicting results, and successful variants. Do not pad the history to meet a paper count. A decade-scale question deserves a decade-scale comparison, not a list of recent papers.

For each material disagreement, determine whether the papers changed the task, metric, scale, compute budget, optimizer, training duration, model family, intervention, or definition. A change in conditions can reconcile apparently contradictory results; unresolved disagreement stays unresolved.

Do not infer consensus from citation counts or implementation popularity. Distinguish:

• exact algebra or a theorem under stated assumptions;
• replicated or converging empirical evidence;
• a result observed in one experimental setting;
• a plausible mechanistic hypothesis;
• a pedagogical analogy;
• an open question.