cerredz Paradigm Routing Reasoning Trace
Goal
Use Paradigm Routing to answer the user's question through explicit paradigm selection and within-paradigm execution, building a reasoning trace that records both the routing decision and the reasoning it enables. The trace should evaluate the problem against all six paradigms ΓÇö Direct, Chain-of-Thought, ReAct, Plan-Execute, Reflection, ReCode ΓÇö select the best-fit paradigm with justification, then produce numbered reasoning items structured by that paradigm's interaction pattern. The goal is to create a public scratchpad that a reviewer can audit without exposing hidden private chain-of-thought. Center the scratchpad on paradigm assessment, subquestions, assumptions, evidence, contrasts, tests, intermediate conclusions, and implications, because those artifacts make paradigm routing useful rather than decorative. Preserve the user's domain, constraints, definitions, and evidence standards so the trace stays tied to the actual task. Keep uncertainty visible by naming weak assumptions, missing evidence, rival interpretations, and confidence changes as they arise. Write the result to root memory/{question_name}.md so the reasoning trace becomes a durable project artifact.
Intent
We run this trace because no single reasoning paradigm is optimal for all problems—ReAct dominates on interactive tasks, Chain-of-Thought excels on multi-step logical deduction, Plan-Execute shines on decomposable problems, and Reflection improves verifiable outputs—yet most reasoning systems blindly apply a single default paradigm to every problem they encounter. Paradigm routing addresses this by explicitly evaluating the problem against all six candidate paradigms before any reasoning begins, selecting the paradigm whose interaction pattern best matches the problem's structural characteristics, and then reasoning within that paradigm. The trace captures both the routing decision (with its justification against each paradigm's criteria) and the within-paradigm reasoning it enables, so a reviewer can independently assess whether the routing was correct and whether the selected paradigm produced reasoning that a different paradigm would not have.
This trace is designed to make paradigm selection transparent and contestable. The paradigm assessment section evaluates the problem against explicit, named criteria for each of the six paradigms—revealing not just which paradigm won but why, and by what margin over the runner-up. This matters because paradigm routing is a meta-decision that shapes everything downstream: if the router selects ReAct for a problem that needed CoT, the reasoning within ReAct will be structurally mismatched to the task regardless of how well it is executed. By making the routing rationale inspectable, the trace gives the consumer the ability to second-guess the router—to see the scores, the criteria, and the justification, and to conclude "I would have routed differently" before reading a word of the within-paradigm reasoning. This is the trace's core value proposition: it does not merely route and reason; it exposes the routing as a contestable decision that the reviewer can independently evaluate.
Instructions
Derive {question_name} from the user's actual question by lowercasing it, replacing non-alphanumeric runs with hyphens, trimming extra hyphens, and using reasoning-trace if no safe name remains. Create the root memory directory when needed, then write or replace memory/{question_name}.md with this trace. Start the file with the question, selected strategy, scale note, source constraints, and paradigm routing assessment. Build the scratchpad by repeatedly applying the paradigm routing move: evaluate the problem's characteristics against each paradigm's decision criteria, select the best-fit paradigm, then reason through the problem using that paradigm's interaction pattern. Use a medium-depth trace by default, usually around 100 numbered lines, unless the user asks for a different depth. Prefer concise public reasoning artifacts over hidden deliberation, and make every numbered item contribute a question, observation, test, comparison, inference, or synthesis. End the file with a synthesis and final answer that follow from the trace, including any important uncertainty that remains.
Background Information About The Reasoning Strategy
Paradigm Routing is a meta-reasoning strategy that operationalizes the Select-then-Solve framework from Zhou et al (2026). Before producing any reasoning, the strategy evaluates the problem against six candidate inference-time paradigms and routes to the one best suited to the problem's characteristics. The six paradigms are: Direct (single-pass generation, no intermediate reasoning ΓÇö best for factual retrieval and simple transformations), Chain-of-Thought (stepwise natural language reasoning ΓÇö best for multi-step logical and mathematical problems where intermediate steps improve correctness), ReAct (interleaved reasoning and tool-use or action ΓÇö best for problems requiring external information, environment interaction, or interleaved reasoning-action cycles), Plan-Execute (explicit planning phase followed by execution ΓÇö best for complex multi-phase problems that benefit from decomposition before execution), Reflection (generate, self-critique, and regenerate ΓÇö best for problems with verifiable solutions where self-critique improves quality), and ReCode (sandboxed code-based reasoning with execution feedback ΓÇö best for computational, symbolic, or code-execution problems).
The empirical foundation comes from testing across 4 frontier LLMs and 10 benchmarks (~18,000 runs). The central finding is that no single paradigm dominates ΓÇö ReAct outperforms alternatives on GAIA by 44 percentage points but underperforms on HumanEval by 15 points relative to CoT ΓÇö and that paradigm routing before reasoning improves aggregate accuracy from 47.6% to 53.1%, recovering 37% of the gap between the best fixed-paradigm strategy (50.3%) and the oracle upper bound (55.4%). The router uses lightweight embedding-based selection trained offline on these runs to predict which paradigm will maximize accuracy for a given problem. Zero-shot self-routing ΓÇö where the model itself chooses the paradigm ΓÇö works reliably only for GPT-5, meaning that for most models explicit routing criteria are necessary.
Algorithm for the Output Document
Restate the user's question, constraints, and evidence standard, then derive the subquestions or working items that paradigm routing needs to inspect.
Evaluate the problem against all six paradigms using routing criteria. Score each paradigm on: whether the problem requires multi-step reasoning (favors CoT, ReAct, Reflection over Direct), whether it requires external information or interaction (favors ReAct over pure reasoning paradigms), whether it benefits from explicit decomposition (favors Plan-Execute), whether the solution is verifiable (favors Reflection), and whether it involves computation or code (favors ReCode).
Select the best-fit paradigm with explicit justification naming the problem characteristics that drove the decision and why the chosen paradigm addresses them better than the runner-up.
Within the chosen paradigm, apply the paradigm's interaction pattern to reason through the problem. For CoT: produce stepwise reasoning. For ReAct: interleave reasoning with any tool-use. For Plan-Execute: decompose first, then execute. For Reflection: generate, critique, regenerate. For ReCode: write code, execute, observe output, revise. For Direct: produce the answer.
For each subquestion, case, option, hypothesis, or criterion the paradigm creates, write numbered public reasoning items that answer it before advancing.
Record assumptions, missing evidence, disconfirming signals, and confidence changes where they affect the analysis.
Synthesize the completed trace into the final answer, showing how the conclusion follows from the paradigm routing decision and the within-paradigm reasoning.
Assess the routing quality retrospectively. After executing within the chosen paradigm, evaluate whether the routing decision was correct by asking: did the selected paradigm's interaction pattern genuinely support the reasoning that was needed, or did the reasoning strain against the paradigm's constraints (e.g., ReAct forced tool-use where pure reasoning would have sufficed)? Did the runner-up paradigm's interaction pattern appear, in retrospect, better suited to the problem's structure? If the routing appears suboptimal, flag it as a routing failure, name which paradigm would have been better suited, and contribute this observation to the router's continuous improvement. A trace that cannot honestly assess its own routing decision hides the most consequential meta-decision from audit.
Implementation Details
Paradigm Routing operationalizes the Select-then-Solve framework introduced by Zhou et al. (2026) in "Paradigm Routing: Select-then-Solve for Inference-Time Compute Allocation." The core premise is that different problems demand different reasoning interaction patterns, and that allocating inference-time compute to paradigm selection before committing to any single reasoning approach yields better aggregate performance than any fixed-paradigm strategy. The six candidate paradigms—Direct, Chain-of-Thought, ReAct, Plan-Execute, Reflection, and ReCode—span the space of known inference-time reasoning strategies, each optimized for a different class of problem characteristics. The router evaluates each problem against all six paradigms using lightweight criteria derived from the empirical performance of each paradigm across the benchmark suite, selecting the paradigm predicted to maximize accuracy for the given problem's feature profile.
The empirical foundation is built on approximately 18,000 evaluation runs across 4 frontier LLMs and 10 benchmarks spanning reasoning, knowledge, interactive, and code-generation domains. The central finding is that no single paradigm dominates across all benchmarks: ReAct outperforms alternatives on GAIA (a benchmark requiring web navigation and tool interaction) by 44 percentage points, but underperforms on HumanEval (a code generation benchmark) by 15 points relative to Chain-of-Thought. This extreme performance variance across paradigms—where the best paradigm for one benchmark is the worst for another—is the empirical justification for routing: a fixed-paradigm strategy cannot be optimal across diverse problem distributions. The routing approach improved aggregate accuracy from 47.6% (uniform random paradigm selection) to 53.1%, recovering 37% of the gap between the best fixed-paradigm accuracy (50.3%) and the oracle upper bound of 55.4% (where the best paradigm is chosen with perfect foresight for each problem). This 37% gap recovery is substantial given that the router operates on lightweight embedding-based features rather than full task execution, meaning the routing decision adds negligible latency compared to the within-paradigm reasoning it enables.
The router itself uses an embedding-based classifier trained offline on the benchmark runs: each problem is encoded into a dense vector representation, and the classifier learns to predict which paradigm maximizes accuracy from the problem embedding alone. This is a critical design choice because it means paradigm selection happens before any reasoning tokens are generated, so the router's cost is a single embedding computation rather than a full forward pass through the LLM. The alternative approach—zero-shot self-routing, where the LLM itself selects the paradigm through natural language prompting—was tested across model scales and found to work reliably only for GPT-5-scale models, with smaller models (including GPT-4) exhibiting inconsistent paradigm selection that often defaulted to the model's training-distribution preference rather than the problem's actual needs. For the Vidbyte strategy execution pipeline, explicit routing criteria derived from the benchmark findings are used rather than relying on model self-routing, ensuring that the paradigm selection is grounded in empirical performance data rather than model-specific biases.
The routing criteria operate on observable problem characteristics: whether the problem requires multi-step reasoning (favors CoT, ReAct, Reflection over Direct, which performs best on single-step factual retrieval), whether it requires external information or environment interaction (favors ReAct, the only paradigm with an action-taking mechanism), whether it benefits from explicit task decomposition (favors Plan-Execute, which front-loads a planning phase), whether the solution is verifiable through execution or self-consistency checks (favors Reflection, which generates, critiques, and regenerates), and whether the problem involves computation or code (favors ReCode, which provides sandboxed execution feedback). These criteria are not equally weighted—the weights are derived from the learned classifier's feature importance scores—and problems that score high on multiple criteria are routed to the paradigm with the highest aggregate weighted score. The routing decision is always presented with the scores for all six paradigms so that a reviewer can see the margin between the selected paradigm and the runner-up, and can independently judge whether the margin was large enough to have confidence in the routing choice or narrow enough to warrant considering the runner-up's output as well.
Output Information
Write the scratchpad as Markdown in root memory/{question_name}.md before responding to the user. Include this scale note near the top of the file: "Scale: default - aim for around 100 numbered lines, or roughly 2,000 to 3,500 tokens, while adapting to the real complexity of the question." Use numbered scratchpad items for scanability, but treat the number target as approximate and subordinate to usefulness. Keep the scratchpad public, inspectable, and concise enough per line that the structure remains easy to review. After writing the file, respond with the path, the selected paradigm with justification, the scale note, and the final answer summary.
Specify Files And Length And Structure Of Output
Write the artifact to memory/{question_name}.md at the repository root, using the literal memory directory name. Structure the file with the sections Question, Paradigm, Strategy, Scale, Paradigm Assessment, Scratchpad, Synthesis, and Final Answer. The Paradigm Assessment section should contain the evaluation of the problem against all six paradigms with the selected paradigm and justification. The Paradigm section names the chosen paradigm. The Strategy section should list the slash command. The Scratchpad section should target around 100 numbered lines or roughly 2,000 to 3,500 tokens of public scratchpad detail, adjusted reasonably for very small or unusually broad questions. Use subsections inside Scratchpad when the trace becomes long enough that phases, branches, hypotheses, cases, or criteria would improve readability. If the user supplies a domain format, evidence source, or output constraint, preserve it inside this structure while keeping the default scale approximate.
Things Not to Do
- Do not route without evaluating all six paradigms. Skipping paradigm candidates because "this is obviously a Direct problem" or "ReCode can't possibly apply here" defeats the purpose of routing—the router's value comes from surfacing paradigm fits that human intuition might miss, and some of the strongest empirical gains came from problems where the winning paradigm was not the obvious choice.
- Do not present the routing decision as beyond contestation. The trace must include the scores and justification for all six paradigms, not just the winner, so that a reviewer can see the margin and second-guess the router if the margin is thin or the justification is weak. A routing decision that readers cannot evaluate is a routing decision they should not trust.
- Do not mix paradigms within a single trace unless the problem genuinely requires it (e.g., a problem with both a reasoning component and a code-execution component might route to Plan-Execute with ReCode for the execution phase). If mixing paradigms, document the sub-problem decomposition and the paradigm selected for each sub-problem separately, with separate routing justifications.
- Do not force the selected paradigm to fit after the fact. If the paradigm's interaction pattern is producing reasoning that feels unnatural, inefficient, or misaligned with the problem's structure, flag this in the trace rather than contorting the reasoning to fit the paradigm. A trace that acknowledges "ReAct forced tool-use on a problem that needed pure reasoning" is more valuable than one that invents unnecessary tool interactions to satisfy the paradigm's structure.
- Do not omit the retrospective routing quality assessment (step 8 of the algorithm). The router's continuous improvement depends on collecting data about routing failures—cases where the selected paradigm proved suboptimal—and a trace that never flags routing failures provides no signal for improvement. If every trace reports "routing was correct," the router is either perfect (empirically false) or the retrospective assessment is dishonest.
- Do not treat paradigm routing as a replacement for strategy selection. Paradigm routing selects the interaction pattern (how the model reasons) but does not select the domain-specific reasoning strategy (what the model reasons about). The Vidbyte strategy catalog provides the domain strategy; paradigm routing provides the inference-time mechanism for executing that strategy.
- Do not allow the paradigm assessment to become a perfunctory checklist. Each paradigm's evaluation must be substantive—citing specific problem characteristics that match or conflict with the paradigm's strengths—not a series of scores assigned without justification. A paradigm assessment that reads "CoT: 8/10, ReAct: 3/10, Plan-Execute: 7/10" with no explanation is not an assessment; it is a data-free ranking.
- Do not present the final answer without showing how the paradigm choice shaped it. The trace must demonstrate that the answer was produced through the selected paradigm's interaction pattern, and that this pattern contributed something a generic single-pass generation would not have. If the answer could have been produced identically by Direct (single-pass generation), the paradigm routing added cost without adding value.