Concept

• Answer Cleansing - Deterministically parse the model’s free-form extraction into a canonical answer type — number, choice letter, yes/no, or stripped free text.
• Branch-Solve-Merge - Spawn N independent solver trajectories on the same problem, then merge their outputs with a dedicated agent into a stronger combined solution.
• Chain-of-Thought Prompting - Eliciting step-by-step natural-language reasoning before the final answer measurably improves multi-step task accuracy, because emitting the reasoning as visible text is the only way a model can build on its own intermediate insight.
• Compositional Mapping Exemplars - Teach solve-stage few-shots to name component outputs and how they concatenate or repeat, so composition reuses simpler meanings instead of flat I/O pairs.
• Compositional Translation Failure Modes - Under least-to-most on compositional commands, failures split into bad decompositions versus translation bugs — modifiers, conjunction order, copies.
• Consistency as Uncertainty Signal - Use the fraction of sampled paths that agree with the majority answer as a cheap calibration cue — low agreement means the model may not know.
• CoT Failure Modes - Categorize wrong chain-of-thought traces — logical mistakes, calculator errors, unnecessary steps, prior-knowledge gaps, or never starting to reason.
• CoT for Action Grounding - Use natural-language chain-of-thought to map instructions onto discrete action sequences when the hard part is choosing and ordering grounded steps.
• CoT Length Generalization - With sufficient scale, chain-of-thought exemplars of short chains can transfer to longer out-of-distribution step counts where standard prompting fails.
• CoT Path Correctness Gap - A correct final answer does not imply a correct reasoning path — treat CoT traces as debugging evidence, not as verified proofs of intermediate steps.
• Cross-Task Few-Shot CoT Transfer - Few-shot CoT exemplars from another domain help mainly when answer format matches; mismatched formats collapse gains — often worse than zero-shot CoT.
• Decomposition Before Composition Failures - Least-to-most fails when subquestions are nonsensical or when a sub-answer is wrong; CoT often fails earlier by picking the wrong fact or operator.
• Deliberate Search Cost Tradeoff - Pay for tree search over thoughts only when CoT already fails — knobs like beam size, votes, and mixed generate/evaluate models buy performance at 5–100× token cost.
• Domain-Specific Decomposition Prompts - Decomposition few-shots rarely transfer across domains — invest in within-domain demos because a correct split usually decides success.
• External Calculator for CoT - Post-process equations inside a chain of thought with a deterministic calculator so correct plans are not failed by arithmetic slips.
• Forced Reasoning Before Tools - Disable tool choice for a turn while still declaring tools so the model must plan in language before it is allowed to call functions.
• Least-to-Most Prompting - Decompose a hard problem into simpler subproblems and solve them in order, feeding each answer into the next prompt for easy-to-hard generalization.
• LM Search Heuristic - Use the language model itself to score or vote on frontier states so search can prune, look ahead, or backtrack without a hand-coded heuristic.
• Natural-Language Intermediate Reasoning - Extra tokens or post-answer rationales are not enough — useful CoT needs natural-language steps before the answer, not only equations or padding.
• Output-Side Reasoning Augmentation - Improve hard tasks by eliciting intermediate tokens before the final answer — orthogonal to input-side tricks like instructions, few-shots, or soft prompts.
• Pause Tokens - Fine-tune a model to emit meaningless filler tokens before answering, giving it extra timesteps to incorporate context before committing to output.
• Plan-and-Solve Prompting - Prompt the model to form an overarching plan before working the problem step by step, without any tool use or think-act-observe loop.
• Prompted IR Expansion - Expand compact intermediate expressions into flat outputs with a few-shot rewrite prompt when you lack — or prefer not to run — a dedicated executor.
• Reasoning Trigger Sensitivity - Zero-shot CoT accuracy swings widely with the exact reasoning cue — instructive triggers help; misleading or irrelevant ones collapse to baseline or worse.
• Rejection Sampling Best-of-N - Sample n candidate completions from a base policy and pick the highest-scoring one under a reward or judge model — more inference compute, no extra training.
• Scale-Dependent Chain of Thought - Useful chain-of-thought reasoning emerges only past a model-scale threshold and often hurts smaller models — do not extrapolate from tiny-model ablations.
• Self-Consistency Decoding - Sample multiple independent CoT reasoning paths and marginalize to the most consistent final answer — a self-ensemble without a trained verifier.
• Self-Critique Prompting - Ask the model to check its own output for errors before accepting it, as a lightweight reflection step inside a single generation or agent turn.
• Test-Time Compute Sampling - Generate several candidate completions instead of one and select the best, trading roughly linear extra cost for higher output quality.
• Tree of Thoughts - Deliberate inference that searches a tree of coherent intermediate thoughts with LM generation, evaluation, lookahead, and backtracking.
• Two-Stage Reasoning–Answer Extraction - First prompt for a reasoning trace, then a second self-augmented prompt that appends an answer-format trigger and parses the final answer.
• When Chain of Thought Helps - Prefer CoT for hard multi-step tasks on large models with flat baseline scaling; skip it when the task is already easy or the model is too small.
• Zero-Plus-Few-Shot CoT - Insert a zero-shot reasoning trigger into few-shot CoT exemplars so each demo still shows steps but also cues the model with the task-agnostic phrase.
• Zero-Shot Chain of Thought - Elicit multi-step reasoning with a fixed trigger such as “Let’s think step by step,” without per-task few-shot reasoning exemplars.