RESEARCH REFERENCE / LITERATURE REVIEWED 11 NOV 2026
REASONINGAI / FIELD NOTES 001
A research-grounded guide to the major families of reasoning methods used in artificial intelligence and large language models — from formal logic and search to tool-augmented agents, process supervision, reinforcement learning, and latent computation.
No finite catalogue can list every reasoning algorithm, variant, or implementation. This is a structured map of widely used and research-significant families, not a claim of completeness, a leaderboard, or an endorsement. Methods overlap and are often composed into larger systems.
01 / REASONING FOUNDATIONS
These are long-standing forms of inference and computation. An LLM may approximate them in text, call a tool that implements them, or combine them with learned representations.
Derive a conclusion that must follow if the premises and inference rules are valid. Includes propositional and first-order logic, rule chaining, and proof construction.
Useful for consistency checks, rule-based domains, theorem proving, and explanations whose steps can be formally checked.
Infer a general pattern or rule from observed examples. The conclusion is supported by evidence, not logically guaranteed by the samples.
Useful for pattern discovery, classification, scientific generalisation, and learning rules from demonstrations.
Generate plausible explanations for an observation, then compare them against evidence. Abduction proposes hypotheses; it does not prove that the selected explanation is true.
Useful for diagnosis, debugging, root-cause analysis, and scientific hypothesis generation.
Map relations from a known case to a new situation, retrieve similar cases, and adapt the solution while checking whether the structural correspondence holds.
Useful for few-shot transfer, design, legal or scientific precedent analysis, and solving new problems from worked examples.
Distinguish correlation from cause using causal graphs, interventions, counterfactuals, and assumptions about confounding. Prediction alone is not causal identification.
Useful for intervention questions, policy analysis, fault diagnosis, and distinguishing “what happens” from “what would happen if”.
Represent uncertainty with probability distributions and update beliefs as evidence arrives. Bayesian inference combines a prior, a likelihood, and observed data.
Useful for uncertain evidence, diagnosis, sensor fusion, calibration, and reasoning where multiple explanations remain plausible.
Translate a problem into formal constraints or logical statements and use a solver or proof system to establish satisfiability, validity, or a counterexample.
Useful for exact constraints, program verification, scheduling, mathematical proof checks, and safety properties.
Search for an assignment or executable program satisfying a specification. Methods include constraint programming, synthesis from examples, and search guided by partial specifications.
Useful for code generation with tests, planning, configuration, puzzles, and tasks with machine-checkable success criteria.
02 / LANGUAGE-MODEL PROMPTING
Prompting methods change how a model decomposes, expresses, samples, or organises a solution at inference time. Gains are task- and model-dependent; a longer rationale is not automatically more correct.
Ask the model to produce intermediate steps before its final answer. Few-shot CoT supplies worked examples; zero-shot CoT uses an instruction to reason step by step.
Useful for arithmetic, symbolic, and multi-step tasks where intermediate structure helps. Wei et al. (2022); Kojima et al. (2022).
Sample diverse candidate reasoning paths and aggregate their answers, commonly by majority vote. It replaces one greedy trace with a distribution over attempts.
Useful for problems with multiple valid solution paths and a reliably comparable final answer. It increases inference cost and can amplify shared errors. Wang et al. (2022).
Break a difficult task into ordered subproblems, solve simpler pieces first, and use earlier answers to solve later ones. It explicitly targets generalisation beyond the complexity of examples.
Useful for compositional tasks and problems where a long dependency chain can be decomposed. Zhou et al. (2022).
First construct a high-level plan, then execute the substeps. Variants such as PS+ add instructions intended to reduce missing steps and calculation errors.
Useful for zero-shot multi-step tasks when the model needs an explicit task outline. A plan still needs execution checks. Wang et al. (2023).
Generate intermediate questions, answer them individually, and combine the evidence into a final response. Often paired with search or retrieval for questions requiring multiple facts.
Useful for multi-hop question answering and explicit dependency tracking. Subanswers can still be wrong or unsupported. Press et al. (2022).
Express the mathematical or logical computation as executable code rather than relying on natural-language arithmetic alone. A runtime evaluates the generated program.
Useful for numerical reasoning, repeated calculations, and tasks with deterministic execution. Program correctness and runtime safety remain separate concerns. Chen et al. (2022).
The model interprets a natural-language problem and writes a program; a Python interpreter performs the exact calculation. The model handles translation and decomposition while execution handles arithmetic.
Useful for math and symbolic tasks with executable representations. Generated code must be sandboxed and validated. Gao et al. (2022).
Use tables, typed variables, state records, equations, schemas, or explicit intermediate objects to reduce ambiguity between steps. This is a family of representations, not one standard algorithm.
Useful for long tasks, state tracking, structured extraction, and workflows where intermediate outputs can be validated independently.
03 / SEARCH & TEST-TIME COMPUTE
Search-based methods spend additional inference-time computation to explore alternatives, evaluate partial states, backtrack, or refine. Search quality depends heavily on the quality of the candidate generator and evaluator.
Generate coherent partial “thoughts” as branching states, score or evaluate them, and select branches using search strategies such as breadth-first or depth-first exploration with backtracking.
Useful for puzzles, planning, and tasks where an early decision can make later completion impossible. Yao et al. (2023).
Represent partial reasoning states as a graph rather than a single chain or tree. Nodes can be combined, transformed, refined, or reused through dependencies and feedback edges.
Useful for tasks with reusable sub-results, merging candidate ideas, or iterative refinement. Graph orchestration adds evaluator and bookkeeping complexity. Besta et al. (AAAI 2024).
Keep a limited set of promising partial candidates, expand them, and prioritise states using a score or heuristic. Beam search trades exhaustive exploration for a bounded compute budget.
Useful for structured generation and plan search when a workable heuristic can rank partial states. A poor heuristic can prune the correct path.
Allocate simulations across a decision tree using a balance of exploration and exploitation, then select actions from accumulated values. LLMs can propose actions and evaluate states.
Useful for sequential decisions, game-like search, and tasks with a simulator or rollout evaluator. Results depend on rollout fidelity and reward design.
Allocate more compute to harder inputs through additional samples, longer deliberation, iterative refinement, branching search, or verifier-guided selection. It is a resource-allocation strategy spanning many methods.
Useful for variable-difficulty workloads when extra inference budget is available. Report accuracy alongside tokens, latency, and cost.
04 / AGENTS, TOOLS & GROUNDED REASONING
Agentic methods connect language-model decisions to external environments, tools, databases, code, or feedback. They can ground a response in new information, but introduce tool reliability and execution risks.
Interleave reasoning-oriented steps with actions and environment observations. The model reasons about what to do, calls a tool, observes the result, and updates its next step.
Useful for interactive tasks and multi-hop questions requiring external evidence. Tool observations need provenance and validation. Yao et al. (2022/2023).
Retrieve relevant documents or records and condition the answer on those sources. Multi-hop variants retrieve iteratively as intermediate findings reveal new queries.
Useful for knowledge that changes, private corpora, citation-backed answers, and literature workflows. Retrieval is a grounding component, not proof that the synthesis is correct.
Select and invoke external functions, APIs, calculators, databases, compilers, or search systems. Toolformer is an influential approach to learning when and how to call tools.
Useful for exact calculations, current information, and actions that require external state. Validate arguments, permissions, outputs, and side effects.
Translate parts of a task into code, SQL, symbolic constraints, or theorem-prover inputs and execute them with specialised engines. PAL and PoT are examples of this broader pattern.
Useful for exact arithmetic, data analysis, formal constraints, and computational tasks where execution provides a checkable result.
05 / CRITIQUE, VERIFICATION & FEEDBACK
Verification can happen at the final-answer level, at each intermediate step, or through external execution. Feedback loops help only when the signal is informative and the process resists reward hacking.
Generate an initial answer, produce feedback about it, and refine the answer over multiple rounds without changing model weights.
Useful for drafting, structured tasks, and cases with criteria that the model can meaningfully assess. Self-feedback can confidently reinforce an error. Madaan et al. (2023).
Turn feedback from a previous attempt into a short verbal reflection stored in episodic memory and use it to guide the next attempt. It is a trial-and-error agent framework rather than ordinary answer rewriting.
Useful for repeated environment interactions, coding agents, and tasks where later attempts can learn from earlier failures. Shinn et al. (2023).
Have multiple model instances propose answers, critique competing arguments, and revise across rounds. Diversity can expose errors but consensus is not a correctness guarantee.
Useful for adversarial review, alternative hypotheses, and complex answers where disagreement is informative. Du et al. (2023/2024).
Train or use a verifier to score a completed solution or final answer. Best-of-N selection can sample candidates and choose the one with the highest predicted outcome quality.
Useful for reranking candidate solutions when final outcomes can be labelled. A verifier can be exploited by candidates that score well but are wrong.
Score intermediate steps for correctness or quality, rather than judging only the final result. Step-level feedback can identify where a multi-step solution first goes off track.
Useful for mathematical derivations and workflows where local steps can be evaluated. Annotation cost, label ambiguity, and reward hacking remain concerns. Lightman et al. (2023).
Validate outputs with unit tests, property-based tests, type checkers, database constraints, symbolic solvers, or proof checkers. These checks are independent signals, not a substitute for correct specifications.
Useful for code, equations, data pipelines, and any task with a deterministic or formally stated invariant.
“The model reviewed its own answer” is not equivalent to independent verification. Stronger evidence comes from an external checker, a trusted environment, a formal proof, a test suite, or an expert-reviewed source record.
06 / TRAINING PARADIGMS & ARCHITECTURES
Some approaches shape reasoning through training signals or architectural computation. They are distinct from prompting methods and should be evaluated with matched data, compute, and baselines.
Fine-tune a model on worked examples, structured solution traces, or trajectories distilled from a stronger teacher. Quality, diversity, and trace faithfulness matter more than raw volume alone.
Useful for transferring task-specific procedures and consistent formats into smaller or specialised models.
Train using labels or preferences attached to individual reasoning steps. PRM800K is a released dataset of roughly 800,000 step-level correctness labels for mathematical solutions.
Useful for training process verifiers and improving candidate selection on multi-step work. Annotation policy and inter-rater quality must be documented. PRM800K.
Optimise a model using rewards that can be checked by a rule, such as an exact answer, unit tests, or a compiler. Implementations include policy-gradient methods such as PPO and GRPO; the optimiser is not itself a reasoning algorithm.
Useful for maths, code, and other domains with scalable verifiers. Reward hacking, distribution shift, and compute cost must be measured.
Use human preferences, reward models, or outcome-level feedback to improve response policies. These objectives can improve helpfulness or task success but do not guarantee faithful intermediate reasoning.
Useful for aligning response behaviour and selecting solutions under defined quality criteria.
Maintain an internal reasoning state in hidden representations rather than decoding every intermediate step into ordinary text. Coconut feeds hidden states back as continuous “thoughts”.
Useful for research into alternative representations and potentially more compact computation. Interpretability and generalisation remain open questions. Hao et al. (2024).
Use repeated computation, recurrence, or multiple timescales to refine internal states. HRM is a 27M-parameter example with coupled high- and low-level recurrent modules; looped transformers explore related depth-through-recurrence ideas.
Useful for studying adaptive depth and structured computation. Small benchmark wins do not establish general reasoning. HRM (2025).
Combine text with image, video, audio, spatial tokens, or embodied observations. Methods include visual decomposition, grounding, spatial representations, tool-assisted perception, and action-conditioned planning.
Useful for diagrams, scientific figures, robotics, navigation, and tasks where the state is not text-only. Perception errors propagate into later reasoning.
07 / REPORTED EMPIRICAL FINDINGS
The numbers below are reported in the cited studies on their specified tasks and model setups. They are not results produced by ReasoningAI, and they should not be compared as if they came from one shared benchmark or compute budget.
Wang et al. report a 17.9-point gain on GSM8K for self-consistency over greedy CoT in their evaluated setup. The paper also reports gains on SVAMP, AQuA, StrategyQA, and ARC-Challenge. Sampling adds compute and can preserve correlated errors.
Read the paper ↗In the Game of 24 experiment, GPT-4 with CoT solved 4% of tasks, while the ToT framework reached 74%. This is a striking task-specific result, not a general 70-point improvement across reasoning benchmarks.
Read the NeurIPS paper ↗With code-davinci-002 and 14 exemplars, least-to-most prompting reached at least 99% on the reported SCAN splits, compared with 16% for CoT. The result is tied to this compositional benchmark and historical model setup.
Read the paper ↗PAL using Codex exceeded PaLM-540B with CoT by 15 absolute top-1 accuracy points on GSM8K in the paper's reported comparison. This highlights the value of delegating exact execution to a runtime.
Read the paper ↗The ReAct paper reports absolute success-rate improvements of 34 points on ALFWorld and 10 points on WebShop over the compared baselines, using one or two in-context examples. The environments, baselines, and prompting setup matter.
Read the paper ↗Lightman et al. report that their process-supervised model solved 78% of a representative subset of the MATH test set and release PRM800K, with 800,000 step-level human feedback labels. This is a result for their model, data, and evaluation protocol.
Read the paper ↗ · Dataset ↗The DeepSeek-R1 paper reports DeepSeek-R1-Zero AIME 2024 pass@1 rising from 15.6% initially to 77.9% during RL training; self-consistency decoding raised the reported result to 86.7%. This is a specific training run and benchmark, not a universal RLVR guarantee.
Read the Nature paper ↗ARC Prize reported 32% on its ARC-AGI-1 semi-private set and 2% on ARC-AGI-2 for HRM. Its ablations found minimal impact from the hierarchical architecture versus a similar-sized transformer, while an outer-loop refinement process drove substantial performance. The authors' benchmark result and independent analysis should both be read.
Read the verification analysis ↗ · Original paper ↗Each figure is a paper-reported result, not an independent reproduction by ReasoningAI. Differences in model version, prompt, decoding, sampling budget, dataset split, evaluator, and contamination controls can materially change outcomes. Always read the full methods and appendices before reusing a number.
08 / BENCHMARKS & EVALUATION
There is no single benchmark for “reasoning”. Use a portfolio of tasks that isolate different skills, and include fresh or perturbed test sets wherever possible.
| Benchmark / dataset | Primary signal | What it is good for | Watch-outs |
|---|---|---|---|
| GSM8K8.5K grade-school problems | Multi-step arithmetic word problems | Numerical decomposition and calculation workflows | Template familiarity, sensitivity to rewording, and train/test leakage |
| GSM-Symbolic | Controlled variations of GSM8K templates | Robustness to changed numbers and added clauses | Do not treat one static set as a complete measure of mathematical ability |
| MATH | Competition-style mathematics | Multi-domain mathematical problem solving | Difficulty, grading, model/tool access, and split protocol affect comparability |
| GPQA448 expert-written questions | Graduate-level biology, physics, and chemistry | Domain-intensive scientific question answering | Single-choice accuracy is not equivalent to original scientific discovery |
| BIG-Bench Hard23 challenging tasks | Logical, symbolic, and other difficult language tasks | Task diversity and prompt-method comparisons | Composite averages can obscure very uneven per-task performance |
| ARC-AGI | Abstract visual pattern induction | Rule induction and out-of-distribution generalisation | Report public and verified hidden sets separately; task augmentation and cost matter |
| HumanEval / MBPP | Code generation with executable tests | Program synthesis and functional correctness | Pass@k, sandbox, test coverage, and benchmark contamination must be reported |
| HotpotQA / FEVER | Multi-hop evidence retrieval and fact verification | Grounded reasoning with source acquisition | Evidence retrieval, answer synthesis, and citation correctness are separate metrics |
| ALFWorld / WebShop | Interactive environment success | Agent planning, tool use, and sequential decisions | Environment version, action budget, success criteria, and resets matter |
09 / FAILURE MODES & RESEARCH CAUTIONS
A credible reasoning system must explain where its method breaks, what was measured, and which checks are independent of the model.
Studies by Turpin et al. and Lanham et al. show that CoT explanations can omit influential information or vary in how faithfully they reflect the model's answer process. In 2026, FaithCoT-Bench added an expert-annotated collection of 1,000+ trajectories, including 300+ unfaithful instances, to evaluate faithfulness at the individual-trajectory level. Two EMNLP 2026 contributions extended this area: a survey organised internal alignment and external consistency, while an information-theoretic study found its proposed token-capacity bound did not explain unfaithfulness on the evaluated benchmarks, pointing to training as a possible source. Treat generated traces as inspectable outputs, not automatically as causal access to internal computation.
Turpin et al. ↗ · Lanham et al. ↗ · FaithCoT-Bench (ICLR 2026) ↗Huang et al. found that LLMs struggle to correct reasoning using only their own internal signals and sometimes degrade after an attempted correction. Prefer grounded feedback from tests, tools, evidence, or calibrated verifiers.
Huang et al. (ICLR 2024) ↗GSM-Symbolic reports performance variance across instantiations and substantial degradation when extra clauses are introduced. Evaluate counterfactuals, paraphrases, changed quantities, distractors, and longer dependency chains.
Mirzadeh et al. (ICLR 2025) ↗Benchmark saturation, public test exposure, contamination, evaluator weakness, and domain mismatch can make reported gains look broader than they are. Use hidden or fresh tests and report per-task distributions.
Evaluation resources ↗Sampling, debate, tree search, and long reasoning traces increase latency and token use. Compare methods under both matched compute and realistic deployment budgets, and publish cost-quality curves.
Reward models and automatic graders can be fooled. Validate the verifier on adversarial cases and independently audit the selected outputs, not only its scores.
DeepSeek-R1 reward design discussion ↗10 / A REPRODUCIBLE EVALUATION PROTOCOL
This protocol is consistent with the evidence-first principles in ReasoningAI Studio. It is a practical checklist, not a formal standard or a substitute for domain-specific review.
11 / SELECTED PRIMARY SOURCES
This is a curated starting bibliography rather than a systematic review. Links point to original papers, conference proceedings, official dataset repositories, or independent benchmark analyses. Read the full paper, appendices, and data cards before relying on a reported result.
Literature snapshot: 11 November 2026. This page is educational research synthesis, not a systematic review, product benchmark, or claim that ReasoningAI has independently reproduced the cited experiments. Citation links and benchmark availability may change; verify source versions and licenses before reuse.
REASONINGAI / RESEARCH WITH A PAPER TRAIL
Reasoning methods become useful research when assumptions, evidence, costs, and failure modes remain visible.