PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Reasoning with Foundation Models

Jiankai Sun, Chuanyang Zheng, Enze Xie, Zhengying Liu, Ruihang Chu, Jianing Qiu, Jiaqi Xu, Mingyu Ding, Hongyang Li, Mengzhe Geng, Yue Wu, Wenhai Wang, Junsong Chen, Zhangyue Yin, Xiaozhe Ren, Jie Fu, Junxian He, Wu Yuan, Qi Liu, Xihui Liu, Yu Li, Hao Dong, Yu Cheng, Ming Zhang, Pheng Ann Heng, Jifeng Dai, Ping Luo, Jingdong Wang, Ji-Rong Wen, Xipeng Qiu, Yike Guo, Hui Xiong, Qun Liu, Zhenguo Li

arXiv:2312.11562v5cs.AIcs.CLcs.CVcs.LG

TL;DR

Foundation models have prompted growing interest in whether they can exhibit human-like reasoning, while recent multimodal and interactive reasoning advances lack a systematic, comprehensive survey. This paper surveys more than 650 papers and concludes that no consensus has emerged on how to bring foundation-model reasoning to a consistently superhuman level.

  • Problem

    Recent multimodal and interactive reasoning advances lack a systematic, comprehensive survey, amid debate over whether foundation models can exhibit human-like reasoning.

    Source: §1 · p. 5

  • Method

    The paper surveys more than 650 papers, primarily from the past two years, covering reasoning tasks, approaches, techniques, benchmarks, applications, challenges, and future directions.

    Source: §1 · p. 7

  • Results

    Across the surveyed work, consensus has yet to emerge on how to advance foundation-model reasoning to a consistently superhuman level.

    Source: §7 · p. 97

  • Takeaways & Limitations

    The survey offers a broad account of the field's current state and future possibilities, while emphasizing interpretability and security.

    Source: §1 · p. 7, §7 · p. 97

    In causal discovery, LLM performance can vary with prompt design and is not yet stable or consistent.

    Source: Causal Discovery · p. 30

Abstract

from arXiv · show

Reasoning, a crucial ability for complex problem-solving, plays a pivotal role in various real-world settings such as negotiation, medical diagnosis, and criminal investigation. It serves as a fundamental methodology in the field of Artificial General Intelligence (AGI). With the ongoing development of foundation models, e.g., Large Language Models (LLMs), there is a growing interest in exploring their abilities in reasoning tasks. In this paper, we introduce seminal foundation models proposed or adaptable for reasoning, highlighting the latest advancements in various reasoning tasks, methods, and benchmarks. We then delve into the potential future directions behind the emergence of reasoning abilities within foundation models. We also discuss the relevance of multimodal learning, autonomous agents, and super alignment in the context of reasoning. By discussing these future research directions, we hope to inspire researchers in their exploration of this field, stimulate further advancements in reasoning with foundation models, and contribute to the development of AGI.

1 Introduction

Reasoning with foundation models has drawn growing interest, but a systematic survey of recent multimodal and interactive reasoning advances is still needed. This survey connects reasoning tasks and techniques across the field and reviews more than 650 papers, primarily from the past two years.

1 Introduction

  • The introduction distinguishes formal language reasoning from natural language reasoning, which supports tasks such as question answering, retrieval, summarization, and sentiment analysis.Formal reasoning includes software and hardware verification, theorem proving, and automated reasoning.Source: §1 · p. 5, Figure 1 · p. 6
  • Earlier surveys covered specific aspects of foundation models, while recent multimodal and interactive reasoning lacked a systematic, comprehensive survey.The paper frames multimodal and interactive reasoning as closer to human reasoning styles.Source: §1 · pp. 5–6
  • It takes a broad perspective across reasoning tasks, including commonsense, mathematical, logical, causal, visual, multimodal, embodied, and defeasible reasoning.The survey aims to connect research efforts and highlight relationships across these areas.Source: §1 · pp. 6–7
  • The survey covers more than 650 papers, primarily from the past two years, and reviews tasks, approaches, techniques, benchmarks, applications, challenges, and future directions.It aims to provide a comprehensive understanding of the field's current state and future possibilities.Source: §1 · p. 7

2 Background

This section defines reasoning across several traditions, surveys its forms and mathematical representations, and introduces language, vision, and multimodal foundation models. It also covers application-driven tasks and research on foundation-model reasoning across evaluation, multilingual, mathematical, and zero-shot settings, noting the need for more sophisticated reasoning systems.

2.1 Definition of Reasoning

  • Reasoning broadly connects information to informed judgments or solutions; cognitive, logical, and natural-language definitions emphasize incomplete knowledge, premise-based inference, and integrating explicit or implicit knowledge, respectively.Cognitive reasoning models conclusions despite incomplete or inconsistent knowledge (Furbach et al., 2019); logical reasoning derives conclusions from premises (Nunes, 2012); natural-language reasoning combines knowledge to infer conclusions about realistic or hypothetical worlds (Yu et al., 2023a).Source: §2.1 · p. 8, Philosophy · p. 9, Logic · p. 9

2.1.1 Deductive, Abductive, and Inductive Reasoning

  • Deduction derives conclusive specifics from premises, induction generalizes probable patterns from observations, and abduction proposes plausible explanations; deduction is most studied, while induction and abduction remain under-explored.Inductive work includes CLUTRR for kinship classification (Sinha et al., 2019) and relation induction involving unseen entities (Yang et al., 2022c).Source: §2.1.1 · p. 10
  • Other taxonomies distinguish formal from informal, neural from symbolic or neural-symbolic, backward from forward, single-step from multi-step, and unimodal from multimodal reasoning.These classifications describe formality, computational framework, reasoning direction, number of steps, and input modality; additional categories include factual, counterfactual, defeasible, default, and abstract reasoning.Source: §2.1.1 · pp. 10–11

2.1.2 Mathematical Representation

  • Mathematical representations illustrate reasoning through propositions and modus ponens, predicates and quantifiers, set operations, shortest paths, conditional probability, and axiomatic proofs.These examples span propositional and predicate logic, set theory, graph theory, probability, and formal systems; complex reasoning problems may require more elaborate formalisms.Source: Propositional Logic · p. 12, Predicate Logic · p. 12, Set Theory · p. 12

Formal Systems

  • With foundation models, reasoning research has moved beyond strict traditional categories toward flexible, application-driven tasks such as commonsense and mathematical reasoning.The examples include understanding everyday situations and solving math word problems, reflecting a focus on addressing specific challenges across varied contexts.Source: Formal Systems · p. 13

2.2.1 Language Foundation Models and Language Prompt

  • Foundation models have advanced language tasks from coherent text generation and understanding to dialogue, translation, summarization, and question answering.Examples include GPT-3 (Brown et al., 2020), GPT-4, PaLM (Chowdhery et al., 2022), and open-source LLaMA models (Touvron et al., 2023a; Touvron et al., 2023b); PanGu-Σ (Ren et al., 2023) reports 6.3x higher training throughput through heterogeneous computing.Source: §2.2.1 · p. 14

2.2.2 Vision Foundation Models and Visual Prompt

  • Vision foundation models support classification, retrieval, detection, segmentation, and video understanding, while model-fusion systems combine specialized models for image editing, inpainting, and video tracking.SAM (Kirillov et al., 2023) produces object masks from prompts and demonstrates zero-shot segmentation; fusion approaches also use prompts to guide segmentation and downstream editing or tracking.Source: §2.2.2 · pp. 14–15, Model Fusion: Enhancing Visual Task… · p. 15

2.2.3 Multimodal Foundation Models

  • Multimodal models align or integrate text, images, and other inputs for tasks such as segmentation, captioning, visual question answering, and audio-visual understanding.Combining modalities can support reasoning and decisions across domains including robotics, healthcare, autonomous vehicles, and multimedia analysis.Source: §2.2.3 · pp. 15–17

2.2.4 Potential for Applications in Reasoning

  • Minerva (Lewkowycz et al., 2022), built on PaLM, answered nearly a third of more than 200 undergraduate problems across quantitative sciences.The model has up to 540B parameters.Source: §2.2.4 · p. 17
  • Zero-shot-CoT (Kojima et al., 2022) uses one prompt template without handcrafted few-shot examples across arithmetic, symbolic, and logical reasoning tasks.Examples include MultiArith, GSM8K, AQUA-RAT, SVAMP, Date Understanding, and Tracking Shuffled Objects.Source: §2.2.4 · pp. 17–18
  • Laban et al. (2023) identify challenges evaluating complex LLM tasks and the need for improved benchmarks; Shi et al. (2023) demonstrate multilingual models’ reasoning beyond language.Shi et al. report commonsense reasoning and semantic judgment in word-in-context settings, with language models using chain-of-thought processes as multilingual reasoners.Source: §2.2.4 · p. 17
  • STaR (Zelikman et al., 2022) iteratively generates rationales and fine-tunes on correct answers, while MWP-BERT (Liang et al., 2022b) applies BERT and RoBERTa pretraining to math word problems.The passage gives BERT’s size as 110M and RoBERTa’s as 123M.Source: §2.2.4 · p. 17
  • Intelligent systems capable of more sophisticated reasoning beyond simple pattern recognition are still needed.Source: §2.2.4 · p. 18

3 Reasoning Tasks

The section surveys a broad taxonomy of reasoning tasks for foundation models, spanning everyday inference, formal problem solving, and applications across modalities and domains.

3 Reasoning Tasks

  • The taxonomy includes commonsense, mathematical, logical, and causal reasoning tasks.It also includes multimodal and visual reasoning, embodied reasoning, and other conceptual or applied tasks.Source: §3 · p. 18
  • Multimodal, visual, and embodied reasoning cover reasoning across data modalities, interpretation of visual data, and embodied agents interacting with environments.Source: §3 · p. 18
  • Other tasks include abstract and defeasible reasoning, medical and bioinformatic applications, and long-chain reasoning.Source: §3 · p. 18

3.1 Commonsense Reasoning

Commonsense reasoning concerns inferences about everyday situations and knowledge. The survey organizes work around question answering, physical reasoning, and spatial reasoning, alongside broader approaches and benchmarks.

3.1 Commonsense Reasoning

  • Commonsense reasoning is the capacity to infer assumptions about everyday situations; work includes structured representations, model evaluations, and commonsense-specific adaptation.Madaan et al. (2022) cast structured commonsense reasoning as code generation for graph construction; evaluations span zero- and few-shot settings, benchmarks, and model sizes, while fine-tuning work targets social intelligence and commonsense generation.Source: §3.1 · pp. 18–20

3.1.1 Commonsense Question and Answering (QA)

  • Commonsense QA requires implicit everyday knowledge beyond explicit facts; CQA has about 12,000 multiple-choice questions, and CoS-E adds human explanations.CAGE fine-tunes a language model on question inputs and human-generated explanations to generate explanations.Source: §3.1.1 · pp. 20–21

3.1.2 Physical Commonsense Reasoning

  • Physical commonsense reasoning concerns object properties and interactions; ESPRIT explains key physical events, while datasets and benchmarks assess physical or audiovisual reasoning.PACS contains 13,400 question-answer pairs, 1,377 distinct questions, and 1,526 videos; PIQA tests choosing between two solutions, and NEWTON is a physical-reasoning repository, pipeline, and benchmark.Source: §3.1.2 · pp. 21–22

3.1.3 Spatial Commonsense Reasoning

  • Spatial commonsense reasoning infers object positions and relations; image-synthesis models outperform other tested models at learning coherent spatial relationships in one benchmark.The reported spatial insights from these models are also useful for natural-language tasks requiring spatial commonsense reasoning.Source: §3.1.3 · p. 22

3.2 Mathematical Reasoning

The survey treats mathematical reasoning as symbolic problem solving, covering arithmetic, geometry, theorem proving, and scientific reasoning. It reviews methods and resources while noting boundaries in the scientific-reasoning landscape.

3.2 Mathematical Reasoning

  • Mathematical reasoning uses symbolic expressions and learned rule sets; the survey also notes weak correlation between LLM performance and question difficulty.MT2Net retrieves supporting facts from financial reports and generates executable reasoning programs for MultiHiertt questions.Source: §3.2 · p. 23

3.2.1 Arithmetic Reasoning

  • Math word problems test arithmetic reasoning, which remains challenging for language models; approaches include equation templates, expression trees, prompting, and step-level verification.Chain-of-thought prompting with a 540-billion-parameter language model performs comparably to task-specific fine-tuned models across multiple tasks; PromptPG dynamically selects in-context examples.Source: §3.2.1 · pp. 23–24

3.2.2 Geometry Reasoning

  • Geometry solvers map problems to logical forms or programs, generate sequences for calculation and proof, or search for goals using theorem knowledge as rules.Source: §3.2.2 · p. 24

3.2.3 Automated Theorem Proving

  • Learnable theorem proving aims to reduce reliance on expert manpower and broaden applicability, using language models to propose proof actions that tree search assembles into valid steps.The survey also describes autoformalization, LeanDojo resources, and correction methods intended to address hallucinations and improve conjectures.Source: §3.2.3 · pp. 24–25

3.2.4 Scientific Reasoning

  • Scientific reasoning spans AI-for-science models and benchmarks, but research centers on traditional sciences while quantum modeling remains scarce amid data, standardization, and hardware constraints.Benchmarks include textbook-derived scientific problems, interactive science tasks, and multimodal curriculum questions; the survey calls for wider disciplinary coverage.Source: §3.2.4 · pp. 25–27

3.3 Logical Reasoning

Logical reasoning derives conclusions from premises using formal logic. The survey describes prompting and fine-tuning approaches, alongside work on propositional and predicate logic.

3.3 Logical Reasoning

  • LLM logical reasoning uses either prompting for stepwise natural-language inference or fine-tuning; symbolic solvers and refinement offer another route, while natural language risks unfaithful reasoning and hallucinations.Logic-LM translates natural-language problems into symbolic forms for deterministic solvers, then uses solver errors to revise those forms.Source: §3.3 · pp. 27–28, Figure 7 · p. 28

3.3.1 Propositional Logic

  • Propositional logic uses true-or-false statements combined by logical connectives; GPT-2 and GPT-3 have been fine-tuned to simulate non-recursive resolution rules.Source: §3.3.1 · pp. 28–29

3.3.2 Predicate Logic

  • Predicate logic extends propositional logic with variables, domains, objects, relations, and functions; its representations can support interpretable pattern and rule discovery.Inductive logic programming uses first-order logic to represent hypotheses and data.Source: §3.3.2 · p. 29

3.4 Causal Reasoning

The section organizes causal reasoning around causal discovery, effect inference, attribution, judgment, and counterfactual reasoning. It reviews promising but inconsistent LLM performance in causal discovery and counterfactual tasks, alongside limits on actual causal inference.

3.4 Causal Reasoning

  • Causal reasoning covers discovery, effect inference, attribution, and judgment, with causal graphs depicting variables and directed relationships.Discovery identifies causal structure; effect inference estimates intervention outcomes; attribution identifies causes, while judgment assigns reward or blame.Source: §3.4 · pp. 29–30, Figure 8 · p. 30

Causal Discovery

  • LLMs show variable causal-discovery performance, and Code-LLMs with code prompts perform better than text-only LLMs in causal reasoning.Prompt crafting affects performance, and ChatGPT has produced false negatives in neuropathic-pain diagnosis.Source: Causal Discovery · pp. 30–31

Type Causality and Actual Causality

  • Type causality concerns variable relationships, whereas actual causality focuses on the causes of individual events.Current LLMs struggle with new knowledge and high-precision high-stakes decisions, and lack physical measurements for actual inductive inference.Source: Type Causality and Actual Causality · p. 31

3.4.1 Counterfactual Reasoning

  • Counterfactual reasoning evaluates hypothetical alternatives to actual events, distinguishing factual reasoning in the default world from reasoning in a counterfactual world.The section describes language-model outputs as conditioned on prompts for the task and world.Source: §3.4.1 · pp. 31–32
  • Counterfactual task variants often reduce language-model performance, while pretrained models may produce completions that contradict established world knowledge.Across 11 tasks, GPT-4, Claude, and PaLM-2 often rely on narrow, context-specific procedures that transfer poorly; causal-discovery performance also remains inconsistent and prompt-sensitive.Source: Type Causality and Actual Causality · p. 31, §3.4.1 · p. 32

3.5 Visual Reasoning

The section surveys visual reasoning frameworks and models, spanning visual perception and task evaluation through systems that process 3D data. It describes 3D reasoning advances while noting that the field remains at an early stage.

3.5 Visual Reasoning

  • G-VUE evaluates visual understanding across Perceive, Ground, Reason, and Act, using 12 tasks that include reconstruction, reasoning, manipulation, and interaction.The section also describes visual reasoning as drawing conclusions from visual stimuli.Source: §3.5 · p. 32, Figure 9 · p. 33
  • VLGrammar jointly induces language and image grammars, AeNER addresses dynamic spatio-temporal reasoning, and LISA segments objects from implicit text queries.LISA combines LLM world knowledge with image understanding to locate queried objects.Source: §3.5 · p. 33

3.5.1 3D Reasoning

  • 3D-LLM, PointLLM, and 3D-VisTa apply geometric, visual, and textual information across 3D tasks, including question answering, grounding, captioning, and situated reasoning.PointLLM outperforms 2D baselines on object classification and captioning; the field remains burgeoning, with further work proposed on intricate scenes and real-world applications.Source: §3.5.1 · pp. 33–34

3.6 Audio Reasoning

The section presents self-supervised speech representations as a foundation for audio reasoning across downstream tasks. It distinguishes discriminative and generative speech evaluation and surveys model families and benchmarks.

3.6 Audio Reasoning

  • Speech reasoning uses self-supervised representations from a single foundation model that can be applied across downstream tasks.Speech is identified as the major source of audio reasoning data.Source: §3.6 · p. 34

3.6.1 Speech

  • Speech tasks divide into discriminative decisions and generative speech production, evaluated respectively by SUPERB and SUPERB-SG.SUPERB covers ten tasks across Content, Speaker, Semantics, and Paralinguistics; SUPERB-SG covers five generative tasks, including translation, out-of-domain ASR, voice conversion, separation, and enhancement.Source: §3.6.1 · pp. 34–35
  • Self-supervised speech models are categorized as generative reconstruction, contrastive discrimination, or predictive teacher-student approaches.Transformer-Encoder and Conformer-Encoder architectures are widely adopted in speech foundation models.Source: §3.6.1 · p. 35

3.7 Multimodal Reasoning

The section surveys reasoning across text, images, video, and other modalities, covering alignment, generation, multimodal-to-text tasks, and understanding. It describes representative model approaches and notes dataset-integration and training-cost challenges.

3.7 Multimodal Reasoning

  • Multimodal reasoning integrates multiple information modalities, and the section highlights instruction tuning, in-context learning, and LLM-aided visual reasoning as key techniques.The survey presents multimodal reasoning as a route to broader problem-solving and more flexible user interaction.Source: §3.7 · p. 35, Figure 11 · p. 36

Image-Text Alignment

  • CLIP aligns image and text representations, BLIP-2 uses a Q-Former with a fixed image encoder, and Flamingo fuses pretrained vision and language backbones.BLIP-2 avoids fine-tuning the vision encoder.Source: Image-Text Alignment · p. 36

Text-to-image Generation

  • Text-to-image systems include diffusion-based approaches such as Stable Diffusion and ImageGen, while PixArt is a Transformer-driven diffusion model.Latent diffusion models reduce computational demands relative to pixel-based diffusion models while maintaining competitive performance.Source: Text-to-image Generation · p. 37

Multimodal-to-text Generation

  • Multimodal-to-text systems use image inputs for outputs such as answers, captions, and other text, with approaches including visual prefixes and interlaced image-text prompts.Flamingo-80B supports few-shot multimodal tasks; Frozen and MAGMA preserve a frozen language model while incorporating visual information.Source: Multimodal-to-text Generation · pp. 37–38

3.7.3 Multimodal Understanding

  • Multimodal understanding includes instruction-following, chart reasoning, object detection, image quality assessment, and video understanding.DePlot translates plots into text before reasoning, MatCha combines chart derendering and math-reasoning pretraining, and LLaMA-VID represents each frame with two tokens.Source: §3.7.3 · pp. 38–39
  • Combining text, images, tables, and audio is challenging when datasets contain mismatches or missing data, and multimodal training requires substantial computation.The survey also notes universal representations as a potential application in materials science, chemistry, and biology.Source: §3.7.3 · p. 39

3.8 Agent Reasoning

The survey organizes agent reasoning around planning, interaction with environments, embodiment, and collaboration among agents. It reviews approaches ranging from static plans to feedback-driven adaptation, and connects these ideas to robotics and autonomous driving.

3.8 Agent Reasoning

  • Agent reasoning spans language-based task planning, dual-process planning, world-model search, and robotic policy learning, with classical robotics methods and safety also identified as relevant.SwiftSage pairs a fast SWIFT planner with a deliberate SAGE module; RAP uses Monte Carlo Tree Search, while MimicPlay learns robotic policies from human play data.Source: §3.8 · pp. 39–40

3.8.1 Introspective Reasoning

  • Introspective methods generate plans from internal knowledge: PAL delegates program execution to a runtime, while ProgPrompt and Code-as-Policies structure program-like action generation.The survey notes that introspective planning may not adapt well when dynamic or uncertain environments require feedback.Source: §3.8.1 · pp. 40–41

3.8.2 Extrospective Reasoning

  • Extrospective reasoning iteratively uses environmental observations and execution feedback to revise plans, supporting responses to changing circumstances and unexpected events.Examples include ReAct, which combines reasoning traces and actions, and Statler, which maintains a persistent representation of world state.Source: §3.8.2 · pp. 41–42

3.8.3 Embodied Reasoning

  • Embodied reasoning connects visual and linguistic understanding to robot action; surveyed systems include generalist policies, large-scale robotic models, and vision-language control.RT-1 is trained on over 130,000 episodes across more than 700 tasks, while RT-2 adds multi-stage semantic reasoning and RT-X extends to cross-embodiment settings.Source: §3.8.3 · pp. 43–44

3.8.4 Multi-agent Reasoning

  • Multi-agent reasoning involves agents interpreting others’ actions and intentions, with surveyed debate approaches reporting improved reasoning abilities and factual accuracy.Other examples include LLM-based self-adjusting agents and federated LLMs for collaborative model development while keeping data decentralized.Source: §3.8.4 · pp. 44–45

3.8.5 Reasoning in Autonomous Driving

  • Autonomous-driving work applies language models to interpretable driving, tool-mediated decisions, motion prediction, and language-supervised traffic scenarios.The survey identifies using logical reasoning to explain driving behaviors as a challenge and proposes common sense and world knowledge as potential resources for corner cases, explainability, and safety.Source: §3.8.5 · p. 45

3.9 Other Tasks and Applications

The survey reviews foundation-model reasoning across scientific, medical, biological, and programming tasks, alongside distinct forms such as long-chain, abstract, and defeasible reasoning. It reports strong results in some applications but also identifies limitations and open challenges.

3.9.1 Theory of Mind (ToM)

  • The survey presents two explanations for ToM-like task performance: it may arise alongside advancing language skills or from models exploiting undiscovered language patterns.The latter account suggests language regularities may support ToM task solutions without explicit engagement of ToM.Source: §3.9.1 · p. 46

3.9.2 Weather Forecasting

  • Weather models show strong forecasting results: MetNet-2 predicts high-resolution precipitation up to 12 hours ahead, while Pangu-Weather outperforms ECMWF’s operational system across assessed variables.Pangu-Weather uses 39 years of global weather data, earth-specific priors, and hierarchical temporal aggregation to mitigate accumulation errors.Source: §3.9.2 · p. 46

3.9.3 Medical Reasoning

  • Medical reasoning results include Med PaLM 2 scoring 86.5% on MedQA and GPT-4 scoring 86.7% on the USMLE, alongside foundation models for medical-image and multimodal tasks.The survey calls for rigorous verification and examination of biomedical reasoning outputs, as well as regulations for legitimate and safe use.Source: §3.9.3 · pp. 46–47

3.9.4 Bioinformatics Reasoning

  • Bioinformatics applications span DNA, RNA, and protein analysis, including protein-structure prediction and design, ligand design, genomic question answering, and chemical-synthesis condition recommendation.Examples include AlphaFold for protein structures, Uni-RNA for RNA predictions, and GeneGPT for genomics questions using the NCBI API.Source: §3.9.4 · pp. 47–48

3.9.5 Code Generation

  • Code-generation research covers natural-language-to-code models, code completion, infilling, multilingual generation, instruction tuning, and retrieval-guided generation.The survey also reports that stronger near-duplicate filtering improved SantaCoder performance, while selecting files from repositories with more than five GitHub stars tended to reduce it.Source: §3.9.5 · pp. 48–49

3.9.6 Long-Chain Reasoning

  • Fine-tune-CoT uses reasoning samples from very large teacher models to train smaller models, which surpassed prompt-based baselines and even the teacher on numerous tasks.Long-chain reasoning remains challenging, despite its potential applications in agent and embodied reasoning, decision-making, planning, and question-answering.Source: §3.9.6 · pp. 49–50

3.9.7 Abstract Reasoning

  • Gendron et al. (2023) find that state-of-the-art LLMs perform notably worse on abstract reasoning than on other natural-language tasks.The survey describes abstract reasoning as identifying and applying general patterns from limited data.Source: §3.9.7 · p. 50

3.9.8 Defeasible Reasoning

  • Defeasible reasoning allows conclusions to be revised with new evidence; surveyed resources include datasets for commonsense, natural-language inference, social norms, and contradictory information.BoardgameQA includes 1,000 training, 500 validation, and 1,000 testing examples for each variation; Table 5 reports commonsense-benchmark question counts and choices.Source: §3.9.8 · pp. 50–51, Table 5 · p. 51

3.10 Benchmarks, Datasets, and Metrics

This section surveys benchmarks, datasets, and metrics for evaluating commonsense, mathematical, question-answering, logical, causal, visual, audio, multimodal, embodied, autonomous-driving, and code-generation reasoning, and reports findings on multimodal-model limitations and resources for assessing them.

3.10.1 Commensense Reasoning

  • Commonsense evaluation spans physical, social, and temporal reasoning, using multiple-choice and generative tasks; Rainbow integrates six tasks, while PHYRE, CConS, and SummEdits provide specialized benchmarks.SummEdits spans 10 domains and reports a 20-fold efficiency improvement per sample over previous benchmarks.Source: §3.10.1 · p. 52

Metrics

  • Evaluation uses task-specific metrics: accuracy for multiple choice, cautious BLEU for generation, weighted AUCCESS for PHYRE, and measures including CLD, execution accuracy, MAP/MRR, top-k accuracy, and preference.AUCCESS weights success across attempts more heavily when fewer attempts are needed; BLEU may not align with human judgments.Source: Metrics · pp. 52, 59, §3.10.4 · pp. 60–61

Math Word Problems (MWPs)

  • Math word-problem benchmarks range from annotated algebra corpora to step-solved MATH, table-grounded TabMWP, and multilingual MGSM evaluation.Math23K-F and MAWPS-F annotate formula use at each reasoning step, while FOMAS separates formula learning from reasoning; SVAMP is harder than ASDiv-A despite lower lexical diversity.Source: Math Word Problems (MWPs) · pp. 53–54

Geometry Problem Solving

  • Geometry benchmarks range from shaded-area problems to GeoQA's program-annotated multimodal problems and UniGeo's multi-step proving tasks.GeoQA contains 4,998 problems; UniGeo adds 9,543 proving problems labeled with reasons and mathematical expressions.Source: Geometry Problem Solving · pp. 54–55

Math Question Answering Datasets

  • Math-question-answering resources include rationale-oriented AQuA and MathQA, table and hybrid QA such as Spider and HybridQA, and multi-document HotpotQA.MMQA requires cross-modal reasoning for 35.7% of its questions, while FinQA pairs financial questions with detailed reasoning steps.Source: Math Question Answering Datasets · pp. 55–59

3.10.3 Logical Reasoning

  • Logical-reasoning benchmarks range from ProofWriter's closed- and open-world variants to FOLIO's first-order-logic annotations and BIG-Bench's multi-step ordering task.ProofWriter extends RuleTaker; FOLIO includes 1,435 examples, and LogicalDeduction asks models to order three to seven objects from clues.Source: §3.10.3 · p. 59

3.10.4 Causal Reasoning

  • Causal-reasoning resources include domain-spanning cause-effect pairs, specialized medical and Arctic sea-ice datasets, and CRASS counterfactual questions; evaluations use top-k accuracy and preference.CRASS contains 275 multiple-choice counterfactual instances; Percentage of Preference assesses whether logical completions align with human judgments.Source: §3.10.4 · pp. 60–61

3.10.5 Visual Reasoning

  • Visual-reasoning datasets cover part-level 3D semantics and physical properties, broad diagnostic tasks in CLEVR, and visual questions requiring external knowledge in OK-VQA.PTR includes around 70,000 synthetic RGB-D images; CLEVR has 100,000 rendered images and about one million generated questions.Source: §3.10.5 · p. 61

3.10.6 Audio Reasoning

  • SUPERB covers discriminative audio tasks and SUPERB-SG generative tasks, while multilingual speech corpora support SSL pretraining and labeled-data fine-tuning.Resources include Libri-light’s approximately 60,000 English hours, Chinese Didi corpora of roughly 10,000 hours each, and multilingual corpora such as VoxPopuli, Multilingual LibriSpeech, and Common Voice.Source: §3.10.6 · pp. 61–62

3.10.7 Multimodal Reasoning

  • In text-centric tasks, multimodal models excel at semantically meaningful word recognition but struggle with meaningless character combinations and intricate image details, underperforming domain-specific methods.The evaluation covered text recognition, visual question answering, key information extraction, and handwritten mathematical expression recognition; performance was consistent across text lengths.Source: §3.10.7 · p. 62
  • Multimodal evaluation resources include PASCAL-50S and ABSTRACT-50S for image captioning, LVLM-eHub across six capability categories, and RAVEN and ARC for concept-based abstraction assessment.LVLM-eHub evaluates eight LVLMs using 47 datasets and an online arena; LAMM-Dataset and LAMM-Benchmark address 2D image and 3D point-cloud understanding, while RNSS measures table similarity and CIDEr evaluates image captioning.Source: §3.10.7 · pp. 63–64
  • Object hallucination remains common in VLP and VLM models; POPE addresses evaluation sensitivity, while black-box queries can strengthen targeted adversarial evasion.Higher CIDEr scores may accompany more unfaithful outputs, and visual instructions can affect hallucinations; T2I-CompBench evaluates compositional prompt capabilities.Source: §3.10.7 · pp. 63–64

3.10.8 Embodied Reasoning

  • Embodied-reasoning platforms support agent development and evaluation across simulated household tasks, navigation, interaction, instruction following, and question answering.RoboTHOR, VirtualHome, Gibson, iGibson, BEHAVIOR-1K, Habitat, and Habitat-Lab provide simulation or development capabilities; these platforms can also be used to evaluate LLMs on robotics tasks.Source: §3.10.8 · pp. 64–66

3.10.9 Autonomous Driving

  • DriveLM structures perception, prediction, and planning questions and answers as a graph to investigate LLM reasoning for explainable planning and safer autonomous-driving decisions.Compared with nuScenes QA and HAD, it extends logical reasoning to more tasks and a wider diversity of scenarios.Source: §3.10.9 · pp. 66–67

3.10.10 Code Generation

  • Code-generation benchmarks span 10,000 competition problems in APPS, 164 handwritten problems in HumanEval, 23,914 textual synthesis problems in MathQA-Python, and 974 entry-level tasks in MBPP.The section questions BLEU’s reliability for code generation and reports that human natural-language feedback significantly reduced MathQA-Python error rates; MBPP emphasizes imperative control flow, whereas MathQA-Python uses more intricate descriptions.Source: §3.10.10 · p. 67

4 Foundation Model Techniques

This section surveys foundation-model techniques for reasoning, spanning pre-training, fine-tuning, alignment training, mixture-of-experts, in-context learning, and autonomous agents. It covers pre-training data and architectures, reasoning-data synthesis, parameter-efficient adaptation, and alignment approaches that combine instruction data, supervised fine-tuning, and human-preference training.

4.1.1 Data Source

  • Pre-training spans text, image, and multimodal data, with examples including ROOTS (1.6 TB), ImageNet and ImageNet21K supervised datasets, and LAION-5B’s 5.85 billion CLIP-filtered image-text pairs.ImageBind targets a joint embedding across six modalities.Source: Figure 14 · p. 68, Text Data · p. 69, Image Data · p. 69

Data for Reasoning

  • Code and scientific-paper data can strengthen reasoning: code-trained models outperform natural-language-trained models on structured commonsense tasks, even without source code.CoCoGen (Madaan et al., 2022) reports this comparison; Galactica (Taylor et al., 2022), trained on scientific papers and other sources, demonstrates stronger performance across scientific tasks than existing models.Source: Data for Reasoning · p. 71

4.1.2 Network Architecture

Other Architectures

  • Alternative architectures target Transformer costs and long contexts through efficient state-space or attention variants, linear attention, and dilated attention.RWKV supports parallel training with constant inference complexity, LongNet scales sequence length beyond a billion tokens, and Streaming-LLM extends it to 4 million tokens without additional fine-tuning.Source: Other Architectures · pp. 73–74

Synthesis Data

  • LLM-generated Chain-of-Thought data supports reasoning fine-tuning, while math instruction datasets and teacher explanations provide varied training signals.Limited benchmark availability and the cost of manually creating benchmarks motivate synthesized data. Approaches include sampling and selecting rationales, self-generated solutions, question bootstrapping, hybrid CoT and Process-of-Thought rationales, and explanation tuning; Magister et al. (2023) report improved student performance across arithmetic, commonsense, and symbolic reasoning.Source: Synthesis Data · pp. 74–75, Benchmark Data · p. 74

4.2.2 Parameter-Efficient Fine-tuning

  • Parameter-efficient fine-tuning adapts large models without updating all parameters, using adapters, low-rank updates, prompts, or selected existing parameters.QLoRA enables fine-tuning models with up to 65 billion parameters on one 48GB GPU; partial-tuning methods include Bitfit, Child-Tuning, SSF, and DiffFit.Source: §4.2.2 · p. 75, Adapter Tuning · p. 76, Low-Rank Adaptation · p. 77

Mixture-of-Modality Adaption

  • MMA uses efficient adapters and dynamic routing for vision-language reasoning, while LLaVA links a vision encoder to an LLM for vision-language tasks.LaVIN, built with MMA, is described as more training-efficient and stronger in reasoning than existing multimodal LLMs; LLaVA-1.5 achieves top-tier results across 11 benchmark tasks.Source: Mixture-of-Modality Adaption · pp. 78–79

4.3 Alignment Training

  • Alignment training targets unintended behaviors such as fabricated, biased, or offensive outputs and failure to follow instructions, which arise from a mismatch between next-token prediction and helpful, safe instruction-following.The section frames alignment as finetuning language models to better correspond with human values.Source: §4.3 · p. 79

4.3.1 Data Source

  • Instruction data is represented as d_k = (i_k, y_k), with i_k denoting an instruction and y_k its corresponding response.Source: §4.3.1 · p. 79

Human Data

  • Human instruction data spans curated datasets, benchmarks, and multilingual prompts, including databricks-dolly-15k’s 15,000 instructions and xP3’s 46 languages and 16 NLP tasks.Other examples include OpenAssistant’s more than 10,000 dialogues involving over 13,000 annotators and Super-NaturalInstructions’ more than 1.5k tasks.Source: Human Data · pp. 79–80

Synthesis Data

  • LLM-generated instructions offer an alternative to costly, time-consuming human data collection, with datasets including Unnatural Instructions’ approximately 240,000 instances and LaMini-LM’s 2.58 million pairs.Self-Instruct (Wang et al., 2022c) uses ChatGPT in-context learning; other synthetic-data resources include GPT4All, Alpaca-GPT4, UltraChat, and CoT-Collection.Source: Synthesis Data · pp. 80–81

4.3.2 Training Pipeline

  • The pipeline first applies supervised fine-tuning (SFT) on high-quality instruction data, then may add human-preference training; SFT uses cross-entropy against correct responses but cannot distinguish response quality in detail.Figure 17 depicts online preference training with reinforcement learning and offline approaches that directly optimize the policy.Source: §4.3.2 · p. 81, Figure 17 · p. 81

Online Human Preference Training

  • RLHF (Ouyang et al., 2022) uses SFT, human rankings of response pairs to train a reward model, and PPO (Schulman et al., 2017) to refine the policy, though PPO can be unstable.RAFT (Dong et al., 2023a) instead selects reward-ranked responses for SFT, while APA (Zhu et al., 2023a) uses squared error on estimated advantages.Source: Online Human Preference Training · pp. 81–82

Offline Human Preference Training

  • Offline preference learning avoids complex interactions among policy, behavior, reward, and value models; DPO (Rafailov et al., 2023) implicitly optimizes the RLHF objective, while PRO (Song et al., 2023b) adds SFT regularization.SLiC (Zhao et al., 2022a) calibrates sequence likelihoods, RRHF (Yuan et al., 2023b) uses ranking loss as a simpler PPO alternative, and AFT (Wang et al., 2023m) aligns Chain-of-Thought responses using correctness labels and constraint alignment loss.Source: Offline Human Preference Training · p. 82

4.4 Mixture of Experts (MoE)

This section surveys model architectures and prompting strategies used to support reasoning, from sparse expert selection to iterative refinement and autonomous tool use. It describes how these approaches organize demonstrations, reasoning paths, feedback, and tools to produce or improve outputs.

4.4 Mixture of Experts (MoE)

  • Sparse gating routes each input to selected experts, while MoE architectures scale model capacity and have been adapted to transformers, vision models, and network compression.The surveyed variants include Top-K routing, Top-2 gating, single-expert selection, and a 1.2-trillion-parameter GaLM variant.Source: §4.4 · p. 83, Figure 18 · p. 83

4.5 In-Context Learning

  • In-context learning places a task description and selected demonstrations in a natural-language prompt, letting models perform tasks without gradient updates.Unlike ICL, instruction tuning fine-tunes the model; instruction tuning can also improve ICL, particularly in zero-shot settings.Source: §4.5 · pp. 84–85

4.5.1 Demonstration Example Selection

  • Demonstration selection ranges from retrieving relevant examples to selecting diverse or jointly relevant-and-diverse sets, with LLM-based methods scoring, ranking, or learning example interactions.Approaches include k-NN retrieval, EPR’s two-stage retrieval and ranking, and CEIL’s use of Determinantal Point Processes.Source: §4.5.1 · pp. 85–86

Zero-Shot CoT; Few-Shot CoT

  • Zero-shot CoT elicits reasoning with added instructions, while few-shot CoT uses prompted reasoning paths and variants that structure, cluster, or expand those paths.Plan-and-Solve prompts the model to plan before solving; Auto-CoT generates and clusters reasoning paths to select demonstrations.Source: Zero-Shot CoT · p. 86, Few-Shot CoT · p. 86

Multiple Paths Aggregation

  • Multiple-path methods aggregate or select candidate reasoning, using voting, model choice, or search; adaptive sampling and planning methods address the cost of sampling full paths.Self-Consistency uses majority voting over sampled paths, while RAP uses Monte Carlo Tree Search to explore reasoning space.Source: Multiple Paths Aggregation · pp. 86–87

4.5.3 Multi-Round Prompting; Learned Refiners; Prompted Refiners

  • Multi-round prompting iteratively refines responses, with learned refiners trained on feedback and prompted refiners using self-feedback, tools, or multiple model roles.The surveyed approaches include separating a generator from a corrector, iterative self-refinement, and collaboration among models assigned distinct subtasks.Source: §4.5.3 · p. 87, Learned Refiners · p. 87, Prompted Refiners · p. 88

4.6 Autonomous Agent

  • Autonomous-agent frameworks use LLMs to plan tasks and orchestrate tools, with approaches spanning program generation, tool selection, tool-augmented reasoning, and multimodal assistance.Other surveyed designs add episodic memory, tool creation, lifelong skill libraries, or teams of specialized agents.Source: §4.6 · pp. 88–90, Figure 20 · p. 89

5 Discussion: Challenges, Limitations, and Risks

The discussion surveys challenges and risks in reasoning with foundation models, alongside possible approaches and research directions. It covers reliability, context length, multimodal reasoning, efficiency, human preferences, and multilingual support.

Hallucinations

  • Hallucinations can arise from biased or erroneous pre-training data and models’ inability to acknowledge knowledge limits; mitigation research has not eliminated them.The paper attributes fabricated answers to questions beyond a model’s understanding and notes that complete elimination remains challenging given language complexity.Source: Hallucinations · p. 90

Context Length

  • Larger context windows can capture longer-range dependencies, but their computational costs and possible tradeoff with generalization constrain their use.The discussion notes quadratic cost growth in earlier studies and identifies preserving performance and generalization as an open challenge.Source: Context Length · pp. 90–91

Multimodal Learning

  • Multimodal reasoning is presented as a way to improve performance and support broader understanding, with further modalities and formal-methods applications identified as opportunities.The paper names video, audio, 3D data, and multiple images, and discusses combining foundation models with formal methods for software and hardware systems.Source: Multimodal Learning · pp. 91–92

Efficiency and Cost

  • Large architectures and training data make foundation models costly to train and deploy, while complex reasoning can slow inference; pruning, compression, quantization, distillation, and low-rank factorization are proposed efficiency techniques.These techniques are described as reducing model size, computation, or memory, with low-rank methods intended to improve efficiency without significant performance loss.Source: Efficiency and Cost · pp. 92–93

Human Preference

  • Responsible model behavior is framed as requiring diverse perspectives in training, feedback-informed adaptation, and checks against real-world evidence and trusted sources.The discussion also points to Constitutional AI and training that seeks approval across diverse perspectives.Source: Human Preference · p. 93

Multilingual Support

  • Reasoning support remains uneven across languages, with comprehensive knowledge sources concentrated in a few languages and robust multilingual reasoning models still lacking.The section cites English-pivot translation-retrieval-translation and cross-lingual thought prompting as approaches in related work.Source: Multilingual Support · pp. 93–94

6 Future Directions

The future-directions section identifies further research needs in reasoning with foundation models, emphasizing safety, interpretability, autonomous agents, scientific applications, and superintelligence alignment.

6.1 Safety and Privacy

  • Reasoning models raise safety and privacy concerns, including robustness gaps, backdoors, poisoning, disinformation, privacy leakage, and unauthorized disclosure.The section describes differential privacy and adversarial training as protective approaches, and discusses watermarking for proprietary language models.Source: §6.1 · p. 94

6.2 Interpretability and Transparency

  • Complex, opaque architectures make model behavior, decision processes, and potential biases difficult to understand, motivating interpretability and transparency work.The section says interpretability frameworks can help address transparency and accountability concerns.Source: §6.2 · pp. 94–95

6.3 Autonomous Language Agents

  • For autonomous agents, the paper highlights task generation, value systems, and world models, alongside interactive systems that adapt to user input and preferences.Examples discussed include LLM-based planning in Minecraft and agents that generate action sequences from natural-language instructions.Source: §6.3 · pp. 95–96

6.4 Reasoning for Science

  • Future science applications include reasoning from auditory information and multimodal medical reasoning, including gene-sequence analysis for diagnosis and treatment exploration.The section also names environmental monitoring, audio scene understanding, genetic-disorder identification, and personalized medicine.Source: §6.4 · p. 96

6.5 Super Alignment

  • The paper argues that human supervision may not scale to superintelligent systems, and describes a roughly human-level automated alignment researcher as one possible approach.It characterizes control and alignment of systems beyond human intelligence as an unsolved challenge.Source: §6.5 · pp. 96–97

7 Conclusion

The conclusion describes increasing complexity and efficacy in foundation-model reasoning while emphasizing that its strengths, limitations, and risks require critical attention. It reports no consensus among the surveyed papers on achieving consistently superhuman reasoning.

7 Conclusion

  • The survey describes a progression in foundation-model reasoning toward greater complexity and efficacy.The conclusion also calls for objective recognition of large models’ strengths and limitations.Source: §7 · p. 97
  • The surveyed literature has not reached consensus on how to achieve consistently superhuman reasoning, such as winning an IMO medal or solving open mathematical problems.Source: §7 · p. 97
  • The conclusion calls for critical development that acknowledges risks and limitations to support robust and reliable reasoning systems.Source: §7 · p. 97
Loading 2312.11562v5…