A Survey on Large Language Models with some Insights on their Capabilities and Limitations

1.1 Motivations

• Transformer-based LLMs, including GPT and LLaMA, perform language tasks while also demonstrating commonsense reasoning, code generation, arithmetic, and other complex abilities.Their capabilities span text generation, question answering, translation, and summarization.Source: §1.1 · p. 4
• Exponential growth in data and computing has enabled larger models to capture nuanced linguistic patterns, cultural context, and domain-specific knowledge, while raising ethical and applicability concerns.The introduction identifies GPUs, TPUs, and distributed frameworks as enabling models with billions or trillions of parameters.Source: §1.1 · p. 4

1.2 Goals of the paper

• The paper surveys LLM development, pre-training, architectures, scaling laws, trade-offs, and applications across healthcare, finance, education, law, and scientific research.It also examines how model size, data, and architecture shape emergent abilities, generalization, elicitation, and limitations.Source: §1.2 · pp. 4–5

1.3 Content and organization

• The paper’s structure covers LLM foundations, training and adaptation, in-context learning, chain-of-thought and planning, CoT and PoT experiments, and responsible future application.The CoT and PoT analysis uses publicly available Llama models on GSM8k and gsm-hard reasoning tasks.Source: §1.3 · pp. 5–6

2.1 Definition and Overview

• LLMs use transformer architectures and increasing parameter or data scale to improve language-processing performance, while supporting generation, question answering, translation, summarization, and sentiment analysis.Their capabilities extend to in-context learning, instruction following, and step-by-step reasoning, although ethical concerns, environmental impact, bias, and hallucination remain challenges.Source: §2.1 · p. 7
• Large pre-trained models outperform smaller counterparts and can exhibit emergent abilities such as zero-shot, few-shot, and one-shot learning, though such abilities are not guaranteed in every LLM.Instruction following improves significantly for LaMDA-PT at 68B parameters, while PaLM requires at least 62B parameters to excel broadly, with task-dependent exceptions.Source: §2.1 · pp. 7–8
• Chain-of-thought prompting enables intermediate-step reasoning on complex tasks, with gains especially evident for PaLM and LaMDA variants above 60B parameters and more pronounced beyond 100B.Reported improvements vary by task, following GSM8k > MAWPS > SWAMP for PaLM, while the capability may be influenced by code training.Source: §2.1 · p. 9

2.2 Scaling Law

• Scaling generally improves performance across language tasks and strengthens long-range dependency handling, contextual understanding, and emergent capabilities, but its discontinuous nature remains debated.The scaling-law account describes linear performance growth with exponentially increasing model size, while alternative metrics can show more continuous improvement.Source: §2.2 · pp. 10, 12, Figure 2 · p. 12
• The KM and Chinchilla scaling laws prescribe different resource allocations: KM favors model size, whereas Chinchilla recommends increasing model and data sizes at similar scales.These laws were derived across broad ranges of model, data, and compute sizes, but scaling large models requires substantial processing power, memory, and distributed-training innovation.Source: §2.2 · pp. 11–12

2.3 Prominent Model Families

• Prominent LLM families shape language processing through distinct architectures, with BERT [65] marking a major advance by using bidirectional context in text processing.The survey focuses on prominent models larger than 10B parameters, including their capabilities, limitations, applications, and parameter effects.Source: §2.1 · p. 10, §2.3 · p. 12, §2.3.1 · p. 12

2.3.1 BERT

• BERT uses bidirectional transformer attention with masked language modeling and next-sentence prediction to improve contextual understanding and transfer across NLP tasks.Its large, complex training requirements and limited world knowledge or reasoning beyond the training data constrain its use.Source: §2.3.1 · pp. 13–14

2.3.2 T5

• T5 recasts translation, question answering, classification, and other NLP tasks as text-to-text generation, using a multitask mixture of supervised and unsupervised pre-training on C4 [99].This unified formulation supports one model, loss function, and set of hyperparameters across tasks, although its training is resource-intensive and may be suboptimal for specialized tasks.Source: §2.3.2 · pp. 14–15, Figure 5 · p. 14

2.3.3 GPT Series

• GPT-3 scaled to 175 billion parameters and introduced in-context learning for few-shot and zero-shot task performance using natural-language instructions and examples [88].GPT models use stacked transformer blocks with self-attention, and larger models generally capture subtler data patterns and perform better.Source: §2.3.3 · p. 15
• Across GPT models, scaling expands few-shot, zero-shot, emergent reasoning, creative, and multimodal capabilities, but hallucination, limited reliability, interpretability, and training-cost concerns persist.GPT-4 may have a limited context window and does not learn from experience, while GPT models remain data-hungry and difficult to interpret.Source: §2.3.3 · pp. 16, 18
• GPT-4 outperformed existing models by a considerable margin on the 57-subject MMLU benchmark and accepts interlaced image-text inputs while producing text outputs.Its development used scaling behavior to predict final performance, and GPT-4’s final loss was predicted accurately from runs using at most 10,000× less compute.Source: §2.3.3 · pp. 16–17
• OpenAI o1 improves reasoning through reinforcement learning and extended chain-of-thought computation, showing strengths in mathematics, coding, science, and planning but weaker generalization and autonomy in some settings.o1-preview was preferred to gpt-4o in reasoning-heavy categories but not some natural-language tasks; evaluated models did not surpass Claude 3.5 Sonnet on overall autonomy tasks.Source: §2.3.3 · pp. 18–19

2.3.4 Llama

• Llama provides efficient transformer-based models from 7B to 65B parameters, using pre-normalization, SwiGLU activations, rotary embeddings, and publicly available training data [330].Its design enables the 13B model to run on a single GPU, while web-derived data introduces acknowledged bias and toxicity concerns.Source: §2.3.4 · pp. 19–21
• Llama 2 extends the family from 7B to 70B parameters with a 40% larger public-data corpus, roughly 8% code, increased context length, and grouped-query attention [329].It retains Llama’s RMSNorm, SwiGLU, rotary embeddings, and AdamW-based training settings while focusing data selection on previously recognized bias and toxicity.Source: §2.3.4 · p. 21
• Code Llama [384] benefits from Llama 2 initialization for code generation, while Llama 3 improves capabilities through larger, higher-quality, more diverse training data.Code Llama was fine-tuned on 500B additional tokens, 85% of which were code; Llama 3 used about 15T pretraining tokens spanning general knowledge, mathematics and reasoning, code, and multilingual data.Source: §2.3.4 · pp. 22, 24

2.3.5 Gemma

• Gemma [372, 385] exceeds larger models such as Llama 2 (13B) on language understanding and reasoning while enabling deployment on common developer hardware.Its 2B and 7B pretrained and instruction-tuned variants use Multi-Query Attention, RoPE, GeGLU, and RMSNorm, though smaller size may limit complex-task performance and bias remains a concern.Source: §2.3.5 · pp. 24–25

2.3.6 Claude

• Claude 3 and Claude 3.5 Sonnet improve logical, mathematical, commonsense, ethical, and coding abilities, but remain vulnerable to ambiguity, abstract reasoning, and uncertainty.Claude 3.5 Sonnet achieves higher accuracy on GSM8K and MATH than earlier versions, with faster interaction and more versatile coding, while longer-context memory could reduce coherence errors.Source: §2.3.6 · pp. 25–26

2.4 Specialized Large Language Models

• Specialized LLMs are adapted through domain-specific pretraining, general-model adaptation, or hybrid approaches and evaluated with both broad and sector-specific benchmarks.They target fields such as healthcare and finance, measuring general reasoning, alignment with human intent, and application-specific performance.Source: §2.4 · pp. 26–27

2.4.1 LLMs in Healthcare

• In healthcare, LLMs support imaging, clinical decisions, documentation, drug discovery, personalized medicine, and patient engagement, but privacy, explainability, bias, and safety require oversight.Flan-PaLM achieved state-of-the-art accuracy across MultiMedQA datasets using few-shot, chain-of-thought, and self-consistency prompting, while instruction-tuned Med-PaLM narrowed the gap with clinicians [213].Source: §2.4.1 · pp. 27–28

2.4.2 LLMs in Finance

• Financial LLMs remain an early-stage field [380] supporting trading, risk management, customer service, and fraud detection, while privacy, interpretability, bias, and compliance remain challenges.Domain-specific and continual pretraining adapt general LLMs to financial jargon, data, and context, improving financial analysis and decision-making.Source: §2.4.2 · pp. 28–31
• Financial LLMs are evaluated on sentiment analysis, named entity recognition, question answering, text classification, and regulatory compliance across financial documents, reports, news, and regulatory texts.Source: §2.4.2 · p. 31
• FLANG-ELECTRA achieved 92% F1 in financial sentiment analysis, while FinMA-30B and GPT-4 reached 87% F1 with 5-shot prompting.The survey notes that general-purpose models can outperform fine-tuned models on some tasks but fail on others.Source: §2.4.2 · pp. 32–33
• BloombergGPT was the best-performing model among similarly sized models, although it lagged behind the much larger PaLM540B and BLOOM176B, while retaining general-purpose abilities.Its financial evaluation included domain-specific and internal benchmarks, with strong performance in sentiment analysis and NER/NED tasks.Source: §2.4.2 · p. 37, Table 16 · p. 37, Table 17 · p. 38

2.4.3 LLMs in Education

• In education, LLMs can generate personalized content, tutoring, feedback, grading, research assistance, administrative support, and language-learning help, achieving student-level performance on standardized tests across multiple subjects [370].Their use requires attention to information reliability, bias, privacy, security, and academic integrity.Source: §2.4.3 · pp. 38–39

2.4.4 LLMs in Law

• In law, GPT-3 statutory reasoning used zero-shot, dynamic few-shot, and chain-of-thought prompting, surpassing previous benchmarks but still struggling with simple synthetic statutes [253].These results indicate limitations in statutory reasoning despite GPT-3’s prior knowledge of the U.S. Code.Source: §2.4.4 · p. 40
• ChatGPT passed four University of Minnesota law exams with an average C+ but ranked near the bottom, while GPT-4 reached the top 10% on a simulated bar exam [370].ChatGPT performed better on essays than multiple-choice questions but struggled with issue spotting and applying rules to facts; legal prompt engineering improves complex legal comprehension and reasoning.Source: §2.4.4 · p. 40

2.4.5 LLMs in Scientific Research

• LLMs can support scientific research across literature review, hypothesis generation, data analysis, manuscript drafting, proofreading, and peer review [362].They are also explored for software engineering tasks including code suggestions, summarization, and automated program repair [323] [325] [352].Source: §2.4.5 · pp. 40–42
• GPT-4 generated specific materials-science hypotheses that combined disciplines and reflected academic trends, but errors require human evaluation and refinement [311].The findings argue against simple prompt mirroring while highlighting both creative insights and risks of unsupported or copied ideas.Source: §2.4.5 · p. 41
• GPT-4 can automate data cleaning, feature engineering, model selection, exploratory analysis, visualization, and analytical inference, streamlining data-science workflows [276].These capabilities extend LLM use beyond text generation into practical data-analysis stages.Source: §2.4.5 · p. 41
• LLMs can improve manuscript language, structure, readability, and academic conventions, but ChatGPT was over 10 times slower than DeepL and lacked highlighted or alternative suggestions.They can also generate abstracts, introductions, conclusions, document statistics, and age-adapted or story-like rewrites [255].Source: §2.4.5 · p. 41
• GPT-3 davinci-002 solved 70% of ToM tasks, while davinci-003 achieved 93%, suggesting nascent ToM-like abilities without implying human consciousness [287].The survey notes that later studies question the robustness of LLM abilities in planning, arithmetic, logic, ToM, and mathematical benchmarks [379] [331] [386] [269].Source: §2.4.5 · pp. 41–42

3 Foundations of Large Language Models

• LLMs learn language structure from vast text corpora, enabling strong performance across generation, classification, translation, and other NLP tasks.The section frames Transformer-based training, data selection, adaptation, tuning, and quantization as core foundations for development and deployment.Source: §3 · p. 42

3.1 Pre-training

• Pre-training builds broad language understanding before task-specific fine-tuning, but requires substantial computation and energy while potentially propagating corpus biases and fairness problems [65] [88] [73] [80].Transfer learning and efforts to improve contextual awareness and ethical sensitivity aim to reduce these limitations.Source: §3.1 · pp. 42–43

3.1.1 Unsupervised pre-training

• Unsupervised pre-training uses autoregressive next-token prediction with Transformer self-attention, whereas masked language modeling learns from both preceding and succeeding context [88, 334, 65].Autoregressive generation iteratively feeds sampled tokens back into the model, but its unidirectional objective cannot use future context during prediction.Source: §3.1.1 · p. 43

3.1.2 Supervised pre-training

• Supervised pre-training learns representations from labeled input-output pairs, aligning models with downstream tasks and domain knowledge [89] [74].It can reduce labeled data needs and improve generalization, but annotation costs and overfitting require regularization and validation [45].Source: §3.1.2 · pp. 43–44

3.1.3 Semi-supervised pre-training

• Semi-supervised pre-training combines abundant unlabeled data with scarce labeled data, first learning from vast unlabeled corpora and then refining models on smaller labeled datasets to improve target-task performance [42].Self-training uses pseudo-labels and consistency regularization promotes similar outputs for perturbed inputs, supporting accuracy, robustness, stability, and generalization [34].Source: §3.1.3 · pp. 44–45

3.2 Data sources

• LLM training data spans general sources such as webpages, conversations, and books, plus specialized multilingual, scientific, and programming-code datasets supporting general language modeling and task-specific capabilities [88] [99] [27].Crawled web data mixes high-quality and noisy text and requires careful preprocessing; overtraining on conversational data can cause performance decline.Source: §3.2.1 · pp. 46–47, §3.2.2 · p. 47

3.2.3 Commonly-used data sources.

• Commonly used pre-training data sources include BookCorpus [27], Gutenberg, CommonCrawl subsets [99] [55] [73] [82], The Pile [111], and ROOTS, spanning books, web pages, and diverse sources.BookCorpus contains over 11,000 books, Gutenberg over 70,000 free ebooks, The Pile 800GB of data, and ROOTS 1.61TB of text in 59 languages.Source: §3.2.3 · pp. 48–49

3.3 Data preprocessing

• Data preprocessing removes noise, irrelevant material, and potentially toxic information; quality filtering uses classifiers or heuristics, though classifier filtering may discard dialectal and colloquial language and reduce corpus diversity [155] [131] [299] [257] [349].Language-based filtering removes data outside the target language, and DataJuicer is proposed to improve processed-data quality.Source: §3.3 · p. 50, §3.3.1 · p. 50

3.3.1 Quality Filtering.

• Quality filtering removes low-quality data using metric-based, statistical, and keyword-based criteria, with perplexity measuring autoregressive language-model predictability but not being well-defined for masked models like BERT [65].Perplexity is the exponential average negative log-likelihood of a sequence and reflects prediction uniformity over corpus tokens.Source: §3.3.1 · pp. 50–51

3.3.2 Deduplication.

• Deduplication removes repeated documents, paragraphs, or sentences to reduce redundancy, improve diversity, prevent contamination, and mitigate instability, overfitting, and poor generalization [171] [330] [131] [349] [155] [150].It should precede train-validation-test splitting, and document-, paragraph-, and sentence-level strategies should be used together [155] [150].Source: §3.3.2 · p. 51

3.3.3 Privacy reduction.

• Privacy reduction anonymizes or obfuscates sensitive information such as names, addresses, and phone numbers, while deduplication can reduce privacy risks from duplicated PII [107].Web-collected user-generated content creates a substantial risk of privacy breaches.Source: §3.3.3 · p. 51

3.3.4 Tokenization.

• Tokenization affects model performance and perplexity; subword methods address vocabulary and morphology challenges, with BPE supporting rare words in GPT, BART, and LLaMA and WordPiece reducing out-of-vocabulary issues [35] [36].Unigram tokenization uses probabilistic segmentation and iterative pruning, supporting morphological variation and multilingual processing through SentencePiece, T5, and mBART [46].Source: §3.3.4 · pp. 51–53

3.4 LLM Adaptation

• LLM adaptation fine-tunes pretrained models for specific tasks or domains through instruction and alignment tuning while aiming to preserve generalization and align behavior with human preferences.Some fine-tuning studies suggest task performance may reflect approximate retrieval rather than newly unlocked reasoning capabilities.Source: §3.4 · p. 53, §3.4.1 · p. 53

3.4.1 Instruction Tuning

• Instruction tuning improves zero-shot generalization to unfamiliar tasks; FLAN substantially outperforms traditional models like GPT-3, while instruction-tuned PaLM enhances MMLU, BBH, TyDiQA, and MGSM performance at sizes of at least 62B [156].These gains are attributed to training across diverse datasets, while omitting task descriptions causes performance to decline.Source: §3.4.1 · pp. 53, 54, 55, 57
• Instruction tuning improves LLM performance, task generalization, and domain specialization across model scales and architectures, but higher-quality and more complex instructions matter more than simply increasing dataset size.Smaller tuned models can surpass larger untuned models, while complexity improvements benefit QA and excessive instruction scaling can degrade performance.Source: §3.4.1 · pp. 59–60

3.4.2 Alignment Tuning

• Alignment tuning steers LLM outputs toward helpfulness, honesty, and harmlessness, but human-defined criteria are subjective and alignment can impose an alignment tax [104].RLHF combines supervised fine-tuning, reward-model training, and PPO-based reinforcement learning from human preferences, although it is memory-intensive and hyperparameter-sensitive.Source: §3.4.2 · pp. 60–62

3.5 Architecture

• LLM architectures determine performance, efficiency, and scalability through encoder and decoder components, and are commonly classified as encoder-only, decoder-only, or encoder-decoder models.Transformer-based architectures use stacked layers containing multi-head self-attention and position-wise feed-forward networks, with residual connections and layer normalization [334].Source: §3.5 · pp. 62–63, §3.5.4 · p. 65

3.5.1 Encoder-decoder

• Encoder-decoder models separately encode input representations and autoregressively decode outputs conditioned on the encoded input through cross-attention.This separation supports diverse language tasks; BART [94] and T5 [99] are cited examples, while encoder-only models produce task-specific predictions rather than autoregressive sequences.Source: §3.5.1 · pp. 63–64

3.5.2 Casual decoder

• Causal decoders generate each token from preceding tokens using masked self-attention, making them suited to sequential generation tasks such as completion, language modeling, and text generation.GPT models [88], OPT, BLOOM [349], and Gopher [131] exemplify this architecture.Source: §3.5.2 · p. 64

3.5.3 Prefix decoder

• Prefix decoders permit bidirectional attention over prefix tokens while retaining unidirectional attention over generated tokens, enabling controlled generation under specified contexts or constraints.GPT-3 [88] showed in-context learning that GPT-1 and GPT-2 did not, suggesting that scaling increases capacity in this architecture.Source: §3.5.3 · pp. 64–65

3.5.4 Transformer Architecture

• The Transformer is the dominant LLM architecture because it captures long-range dependencies and complex language structures while scaling to billions or trillions of parameters [334] [88] [330].Its position-wise feed-forward sub-layer is a two-layer network with ReLU activation and learnable parameters W 1, W 2, b1, and b2.Source: §3.5.4 · pp. 65–66
• Normalization configurations trade off stability, training cost, and performance: pre-LN removes warm-up and is more stable, whereas post-LN can be unstable and Sandwich-LN may collapse training.Sandwich-LN places normalization before and after residual blocks, while pre-LN places it inside them and post-LN after them.Source: §3.5.4 · pp. 67–68
• Position embeddings encode token order through absolute, relative, RoPE, or ALiBi mechanisms, with ALiBi reported to improve extrapolation, stability, and convergence.ALiBi applies a predefined distance-based penalty directly to attention logits, while RoPE encodes positions through rotations in embedding space.Source: §3.5.4 · pp. 70–71
• Full attention captures global token relationships but has quadratic complexity, motivating sparse, multi-query/grouped-query, Flash, and paged alternatives for longer or more efficient sequences.Multi-query and grouped-query attention reduce memory bandwidth and computational cost with small sacrifices in model quality by sharing keys and values across heads or groups.Source: §3.5.4 · p. 72
• FlashAttention-2 achieves approximately a 2× speedup over the original FlashAttention by optimizing GPU thread-block and warp partitioning.FlashAttention improves attention speed and memory consumption by organizing input blocks in faster GPU memory.Source: §3.5.4 · pp. 72–73
• PagedAttention partitions dynamic KV caches into non-contiguous physical blocks, increasing GPU memory utilization and enabling efficient memory sharing during parallel sampling.Existing systems waste 60%-80% of memory through fragmentation and over-reservation.Source: §3.5.4 · p. 73
• The survey recommends pre-RMSNorm, SwiGLU or GeGLU, and RoPE or ALiBi for stronger generalization, training stability, and long-sequence performance.It also cautions that placing normalization immediately after embedding layers may degrade performance.Source: §3.5.4 · pp. 73–74

3.5.5 Emerging architectures

• S4 achieves state-of-the-art performance on long-range sequence-modelling tasks, addressing the Transformer’s quadratic token complexity with a fixed-size state space.Alternative approaches include Hyena’s sub-quadratic long convolutions and gating, as well as recursive mechanisms such as RWKV and RetNet.Source: §3.5.5 · pp. 74–75

3.6.1 Parameter-efficient model adaptation

• Adapter tuning approaches full fine-tuning on GLUE with only about 3.6% of parameters trained per task, while prefix tuning achieves comparable performance using about 0.1% of parameters.Adapters freeze the pretrained network and insert small bottleneck modules; prefix tuning optimizes task-specific continuous vectors while discarding its re-parameterization network after training.Source: §3.6.1 · pp. 76–78
• P-Tuning v2 matches full fine-tuning with 0.1%-3% of parameters tuned, while AdaLoRA improves low-budget adaptation by allocating parameters to important weight matrices.P-Tuning v2 applies continuous prompts across all layers, whereas AdaLoRA selectively prunes less-important singular-value updates; both address limitations of uniform or input-only adaptation.Source: §3.6.1 · pp. 79–80
• DyLoRA matches or exceeds static-rank LoRA across sentiment analysis, question answering, and natural language generation while selecting rank sizes dynamically without exhaustive search or retraining.It trains low-rank modules across a spectrum of ranks and adjusts the rank during inference, improving flexibility across tasks and rank constraints.Source: §3.6.1 · p. 80

3.6.2 Memory-efficient model adaptation

• Quantization lowers memory and computational costs by using reduced-precision weights and activations, but lower bit-widths generally increase storage savings alongside greater accuracy losses.Post-training quantization is easier to implement, whereas quantization-aware training can produce better results; uniform, non-uniform, and mixed-precision schemes provide different precision allocations.Source: §3.6.2 · pp. 80–81

4.1.1 ICL strategy

• ICL performs task recognition and learning from demonstrations without gradient updates, but its effectiveness requires careful selection, formatting, ordering, and scoring of examples [88].Training approaches include MetaICL, Symbol Tuning, self-supervised ICL, and PICL, while inference methods use similarity-based selection, entropy-based ordering, perplexity, channel scoring, or calibration.Source: §4.1.1 · pp. 81, 82, 84, 85, 86

4.1.2 ICL performance and origins

• ICL performance depends strongly on demonstration properties and scale: larger models can learn input-label mappings, but ICL still lags fine-tuning on most natural-language-understanding tasks [88].Relevant factors include domain source, label-space exposure, input-label format, demonstration complexity, similarity, and ordering; sufficiently large models can even follow flipped or semantically unrelated labels [345].Source: §4.1.2 · pp. 86–88

4.1.3 ICL future research

• Future ICL research targets stronger pretraining alignment, capability distillation, robustness, efficiency, scalability, knowledge updating, augmentation, and multimodal expansion [265].ICL remains sensitive to demonstration format and permutation, while context limits and the computational cost of many demonstrations constrain scalability; structured, dynamic, and extended-context prompting are proposed responses [141] [374].Source: §4.1.3 · p. 88

4.2.1 CoT strategy

• CoT prompting adds intermediate reasoning steps to ICL prompts, improving complex arithmetic, commonsense, and symbolic reasoning for sufficiently large models exceeding 100 billion parameters [227].Few-shot CoT uses annotated reasoning demonstrations, whereas zero-shot CoT generates reasoning steps directly; demonstration ordering has relatively minor impact.Source: §4.2.1 · p. 89
• DIVERSE improves CoT reasoning by generating multiple paths, filtering them with a weighted voting verifier, and checking individual reasoning steps rather than only complete chains [295].It selects M1 prompts and M2 reasoning paths per prompt, producing M1 × M2 diverse prompts before aggregating predictions.Source: §4.2.1 · pp. 89–90
• Complexity-based prompting most strongly affects LLM performance on complex reasoning tasks, while Complexity-based Consistency selects the majority answer among the K most complex reasoning chains.Prompts can use reasoning-step count, question length, or formula length as complexity indicators; question length can reduce annotation costs when reasoning steps are unavailable.Source: §4.2.1 · p. 91, Figure 41 · p. 92
• Auto-CoT clusters questions and selects representative demonstrations, generating diverse reasoning chains that improve arithmetic and symbolic reasoning performance.Zero-Shot-CoT uses a task-agnostic “Let’s think step by step” prompt, whereas Manual-CoT relies on manually designed reasoning demonstrations and is costly to create.Source: §4.2.1 · pp. 92–93

4.2.2 CoT performance and origins

• CoT helps step-by-step tasks such as arithmetic, commonsense, and symbolic reasoning, but can harm standard-prompting performance on other tasks and may emerge from code training rather than instruction tuning [232] [123] [81] [156].The proposed code-training explanation remains unconfirmed because publicly reported ablation experiments are lacking.Source: §4.2.2 · pp. 93–94

4.3 Program-of-Thoughts

• PoT expresses reasoning as executable Python programs, improving math performance over CoT while struggling with diverse algebraic questions in AQuA and semantic reasoning such as StrategyQA [259].Execution addresses arithmetic, complex-expression, and iteration limitations of LLMs, although zero-shot PoT may require suppressing “#” token logits to prevent reasoning in comments.Source: §4.3 · pp. 94–95

4.4.1 Commonsense knowledge

• LLMs perform commonsense question answering but rely heavily on memorization and distributional patterns, lacking reliable instance-level reasoning, autonomous planning, self-verification, and sufficient generalization [373, 378, 379].External solvers and verifiers in LLM-Modulo frameworks can support planning, while the LM with tutor method achieved 100% accuracy on the reported symbolic tasks, including OOD scenarios.Source: §4.4.1 · pp. 96–98
• LLM planning remains limited by unnecessary steps, missing domain-specific heuristics, and weak generalization to unseen scenarios or symbolic reasoning where action semantics diverge from natural language.These limitations can lead to suboptimal decision-making in planning tasks.Source: §4.4.1 · p. 99

4.4.2 Prompt and code based planning

• Prompt-based planning separates a planner, executor, and environment, with memory and environmental feedback supporting plan reuse, execution, and revision.Plans may be natural-language, symbolic, or programmatic; executors may be LLMs or program interpreters, while environments may be the LLM itself or external systems such as simulators and Minecraft.Source: §4.4.2 · pp. 99–100

4.4.3 Plan generation

• Plan generation decomposes complex tasks into ordered subtasks, using natural-language instructions, tool selection, modular prompting programs, or executable code to produce interpretable solutions.Natural-language plans can be ambiguous, whereas programmatic plans are more verifiable and executable; DECOMP uses a decomposer and modular subtask handlers, while Plan-and-Solve explicitly devises and executes subtasks.Source: §4.4.3 · pp. 100, 101, 104
• PaL and SELF-PLANNING improve reasoning and code-generation performance by combining generated plans with Python execution or staged planning, with stronger effects for larger models.PaL outperforms much larger CoT-based LLMs and is especially effective on GSM-HARD, while SELF-PLANNING improves performance across public code-generation datasets; interpreter execution avoids arithmetic and reasoning errors by design.Source: §4.4.3 · pp. 105–107

4.4.4 Feedback and plan refinement

• Feedback-based refinement follows a planning-execution-refinement loop using internal evaluations or external tools and environments, although reasoning-based approaches typically fix the planning order.Internal feedback includes plan scoring, comparative voting, and memory-preserved reflections, while external feedback includes interpreter errors, visual signals, environmental interactions, and agent communication.Source: §4.4.4 · p. 108
• Tree of Thoughts enables global planning by backtracking through alternative paths and adapting thought generation, assessment, and search across mathematical, creative, and lexical reasoning tasks.Its evaluation includes the Game of 24, Creative Wiring, and 5 × 5 mini crosswords, which require deductive, commonsense, mathematical, and systematic planning abilities.Source: §4.4.4 · pp. 108–110
• Feedback-driven planners revise imperfect plans after failures, with DEPS [344] using situation descriptions, explanations, and a learned goal selector to choose feasible sub-goals in Minecraft.This addresses long dependency chains and unreachable resources by refining plans according to the current state.Source: §4.4.4 · pp. 110–111
• Reflexion achieved 91% pass@1 on HumanEval versus 80% for GPT-4 by using verbal self-reflection and persistent memory to improve decisions across trials.Its Actor, Evaluator, Self-Reflection model, and memory combine trajectory context with feedback stored for later attempts.Source: §4.4.4 · pp. 111–113

4.4.5 LLM-modulo Framework

• LLMs often cannot independently plan or self-verify, so LLM-modulo combines them with hard and soft external critics that reject, refine, or validate candidate plans.Hard critics assess correctness constraints such as causality, timelines, resources, and unit tests, while LLMs can assist with soft criticism and plan reformatting.Source: §4.4.5 · pp. 115–116
• LLM-modulo improved Blocks World to 82% and Logistics to 70% within 15 back-prompting rounds, but reached about 10% on Mystery BW; travel planning improved 6× over a 0.7% baseline.The framework uses external critics and verifiers to assess candidate plans, while LLMs generate, reformulate, and revise them.Source: §4.4.5 · pp. 116–117

4.5 Retrieval-Augmented Generation

• RAG combines LLM generation with retrieval to update knowledge, reduce hallucinations, improve factual grounding, and adapt outputs to domains such as law, medicine, and engineering.Its paradigms progress from Naïve RAG to Advanced and Modular RAG, whose specialized search, fusion, memory, routing, and prediction modules support flexible retrieval-generation workflows.Source: §4.5 · pp. 117–118
• RAG systems comprise retrieval, generation, and augmentation components that connect external or internal knowledge with language-model generation.The standard pipeline indexes and retrieves relevant chunks before generation, while augmentation extends this process through iterative, recursive, or adaptive retrieval for complex reasoning.Source: Figure 66 · p. 120, §4.5 · pp. 120, 122
• Modular RAG reconfigures retrieval, reranking, context compression, fine-tuning, and task adapters to improve task-specific relevance and flexibility.Strategies include query rewriting, hybrid retrieval, dynamic frameworks, rerankers, token filtering, distillation, and alignment of retriever and generator preferences.Source: §4.5 · pp. 119–122
• Iterative, recursive, and adaptive retrieval support multi-step reasoning, ambiguous queries, and autonomous information gathering by repeatedly refining queries, contexts, or retrieval decisions.ITERRETGEN [320], IRCoT [223], ToC [284], Flare [250], Self-RAG [262], GraphToolformer [291], and WebGPT [124] exemplify these strategies.Source: §4.5 · pp. 122–123
• RAG evaluation considers retrieval precision, generation accuracy, and end-to-end performance, using benchmarks such as SQuAD [33] and Natural Questions.These metrics assess retrieved-data relevance, factual correctness, and overall coherence, fluency, and informativeness.Source: §4.5 · p. 123
• RAG remains limited by retrieval latency, dependence on data quality, and scalability challenges, motivating multimodal inputs, improved indexing, and tighter retrieval-generation integration.Future directions include supporting text, images, and audio while improving search efficiency and coupling between retrieval and generation modules.Source: §4.5 · p. 123

• The section examines how skills such as Chain-of-Thought (CoT) arise in large language models.Source: §5 · p. 124
• It summarizes evidence from experiments reported in scientific articles and papers.Source: §5 · p. 124
• It tests whether selected hypotheses are supported using publicly available models through LMStudio software on HuggingFace.Source: §5 · p. 124

• Scale alone does not determine CoT ability: OPT175B and BLOOM176B cannot perform it, whereas smaller UL220B [328] and Codex12B [108] can use CoT to improve performance.The source of CoT remains unclear despite several hypotheses.Source: §5.1 · p. 124
• Code in pretraining is a leading hypothesis for CoT because procedure-oriented programming resembles stepwise solving and object-oriented programming resembles decomposing complex tasks.Training on code may greatly increase chain-of-thought prompting ability, but this requires more thorough verification.Source: §5.1 · p. 124
• Code-trained code-davinci-002 consistently outperforms text-davinci-002 on CoT language tasks [360], including the MRPC evaluation summarized in Table 40.HELM also found strong language-reasoning abilities in models trained on or for code.Source: §5.1 · p. 124, Table 40 · p. 125
• Instruction tuning appears to elicit existing CoT abilities rather than introduce new ones, because its dataset is orders of magnitude smaller than pre-training data.text-davinci-002 could leverage CoT whereas text-davinci-001 could not, and PaLM [155] showed CoT after instruction tuning.Source: §5.1 · p. 124

5.2 Empirical evidences

• The experiments were resource-constrained and therefore focused on publicly available HuggingFace Llama models tested with CoT on GSM8k and gsm-hard, plus PoT on code-expressed reasoning.Tests used LMStudio on an Apple M1 Pro with 10 cores; closed-source and most very large models were unavailable.Source: §5.2 · pp. 125–126
• Llama 3 outperforms Llama 2, and CoT performance improves from 0-shot to 5-shot settings, supporting a strong association between pre-training code content and reasoning ability.The models have similar architectures and sizes, while code comprises 8% of Llama 2 pre-training data versus 17% for Llama 3; GSM8k improvements consistently increase with code percentage.Source: Table 41 · p. 126, §5.2 · pp. 126–127
• PoT improves performance on gsm-hard, with gains increasing alongside pre-training code content, but all models retain low success rates on this larger-number dataset.PoT extracts and executes Python code from model solutions; the parenthetical table results report success rates using this execution-based reasoning.Source: Table 41 · p. 126, §5.2 · pp. 126–127
• Comparisons involving Llama 13B show that model size is not decisive for CoT, although scaling can improve CoT ability when other factors are held comparable.Both models demonstrate CoT reasoning, while the authors use same-size and similar-architecture comparisons to focus on pre-training data.Source: §5.2 · pp. 125–126
• The sharp 0-shot gsm-hard drop suggests that models struggle to reason without recovering the underlying algorithm and instead rely on distributions represented in pre-training data.CoT generalization also weakens after a point, as indicated by the 5-shot gsm-hard results.Source: §5.2 · p. 127

5.3 Prompting

• Prompt and request parameters significantly affect model performance, and experimentation identified a satisfactory prompt format and configuration.Source: §5.3 · p. 127
• Top-p sampling is set to 0.9121 and the maximum token limit to 1024123.Source: §5.3 · p. 127
• The context uses the model’s maximum permitted length, while prompts follow the OpenAI API request format with role-based messages and task-specific system instructions.Source: §5.3 · p. 127

5.4 Examples of generated text

• In 0-shot experiments, the system prompt uses “Let’s think step by step” to initiate chain-of-thought reasoning without task examples.The prompt frames the model as a maths teacher and requests a single numerical solution on the last line.Source: §5.4 · p. 127
• The 5-shot setup places five user-assistant solution pairs before the final user question, allowing assistant messages to guide reasoning without triggering inference.Its simplified system prompt asks the model to think step by step because the examples provide the reasoning guidance.Source: §5.4 · p. 128

Llama2 7B GSM8k 0-shot

• Llama2 7B answers one GSM8k problem correctly with 2 dollars per strawberry-gum pack, but provides no reasoning.On another problem, it gives an incorrect answer rather than the correct answer 30.Source: Llama2 7B GSM8k 0-shot · pp. 128–129

Llama2 7B gsm-hard 0-shot

• Llama2 7B shows no demonstrated ability to answer any gsm-hard 0-shot prompt correctly, so no generated-text sample is reported.The paper explicitly withholds a sample because the model did not answer any prompts correctly.Source: Llama2 7B gsm-hard 0-shot · p. 129

Code Llama 7B GSM8k 0-shot

• Code Llama 7B correctly solves a GSM8k grooming-cost example with an answer of $70, and executing its code snippet returns the correct answer.In another example, its reasoning is coherent but misses the final multiplication, producing an incorrect result instead of 24.Source: Code Llama 7B GSM8k 0-shot · p. 129

Code Llama 7B gsm-hard 0-shot

• Code Llama 7B’s gsm-hard 0-shot output contains the correct answer 168, but its reasoning is noisy and gives multiple answers.The model also produces an incorrect rice calculation, reporting 20104579 kilograms instead of the correct answer 225514250.Source: Code Llama 7B gsm-hard 0-shot · pp. 130–131

Llama2 13B GSM8k 0-shot

• The chicken-feed prompt specifies 20 chickens receiving three cups each daily, with 15 cups in the morning and 25 in the afternoon.Source: Llama2 13B GSM8k 0-shot · p. 131
• The generated chicken-feed explanation introduces inconsistent details, including an unsupported claim about chickens eaten and leftover feed.Source: Llama2 13B GSM8k 0-shot · p. 131
• Llama2 13B correctly answers the chicken-feed problem with twenty cups in the final meal, although its reasoning is only partially coherent.Source: Llama2 13B GSM8k 0-shot · p. 131
• Llama2 13B incorrectly answers the penny-stack problem; the supplied correct answer is 62.Source: Llama2 13B GSM8k 0-shot · p. 132

Llama2 13B gsm-hard 0-shot

• On gsm-hard, Llama2 13B correctly computes that the candle becomes 8 centimeters shorter from 1:00 PM to 5:00 PM, but its reasoning is not coherent.Source: Llama2 13B gsm-hard 0-shot · p. 132
• Llama2 13B incorrectly solves the Grade 5 girl-scout problem; the supplied correct answer is 1806106.8000000003.Source: Llama2 13B gsm-hard 0-shot · p. 132
• The gsm-hard examples include permission to use a calculator while solving the problem.Source: Llama2 13B gsm-hard 0-shot · p. 132

Llama 3 7B GSM8k 0-shot

• The GSM8k examples show that the model can make its reasoning explicit through variable definitions, equations, and elimination of an unknown.For the coin problem, it defines G and S, writes two equations, and adds them to derive the result.Source: Llama 3 7B GSM8k 0-shot · p. 133
• On GSM8k 0-shot, Llama 3 7B correctly solves the coin problem by forming G + S = 110 and G - S = 30, yielding 70 gold coins.The response explicitly adds the equations to obtain 2G = 140, and the passage confirms the answer and reasoning are correct.Source: Llama 3 7B GSM8k 0-shot · p. 133
• On a second GSM8k 0-shot rice problem, the model begins a step-by-step multiplication procedure but produces an incorrect answer; the correct answer is 175.The response identifies 20 sacks and 25 kilograms per sack, but its calculation remains incomplete before the incorrect result is reported.Source: Llama 3 7B GSM8k 0-shot · pp. 133–134

Llama 3 7B gsm-hard 0-shot

• The gsm-hard examples show analogous procedural reasoning with rate-distance-time formulas, but arithmetic or transcription errors can invalidate the final result.The car example correctly applies D = rt, whereas the marble example changes the stated per-box quantity and reaches an incorrect estimate.Source: Llama 3 7B gsm-hard 0-shot · pp. 134–135
• On gsm-hard 0-shot, Llama 3 7B correctly computes 8 hours for the fast car and approximately 16 hours for the slow car traveling at half the speed.The reasoning uses D = rt and applies 480/60 and 480/30; the passage confirms the answer and reasoning are coherent.Source: Llama 3 7B gsm-hard 0-shot · pp. 134–135
• On a second gsm-hard 0-shot marble problem, the model estimates approximately 4.65001 * 10^9 marbles, but the answer is incorrect; the correct answer is 42474420.The response multiplies the box count by an altered per-box quantity and adds 20 marbles, leading to the erroneous estimate.Source: Llama 3 7B gsm-hard 0-shot · p. 135

GSM8K 0-shot

• The model correctly solves the pen-cost problem, deriving a final answer of $12 with coherent reasoning.It combines the pencil and eraser costs, equates that amount to one pen, and multiplies by eight.Source: Llama3.1 7B GSM8k 0-shot · p. 136
• The house-flipping example is incorrect: the model predicts a $10,000 loss, while the correct profit is $70,000.The error comes from calculating $80,000 × 1.5 = $120,000 without adding the original $80,000 again to obtain $200,000.Source: Llama3.1 7B GSM8k 0-shot · p. 137

gsm-hard 0-shot

• On the gsm-hard bridge problem, the model correctly computes that at most 83 boxes can be loaded and provides coherent reasoning.It first finds 1,245 pounds of remaining capacity and divides by the 15-pound box weight.Source: Llama3.1 7B GSM8k 0-shot · pp. 137–138
• The raspberry-count example is incorrect, with a predicted total of 4,046,647 instead of the correct 42,463,867.The model’s intermediate multiplication and final addition do not produce the correct total.Source: Llama3.1 7B GSM8k 0-shot · p. 138

Generated examples

• The juggling example reaches the correct answer, but its reasoning is not coherent.The supplied chain-of-thought reasoning includes an inconsistent calculation involving practice sessions and performance timing.Source: Llama2 7B GSM8k 5-shot · p. 139
• The restaurant-table problem is answered incorrectly: Llama2 7B gives 128, whereas the correct answer is 310.This example demonstrates an arithmetic failure on the GSM8k 5-shot task.Source: Llama2 7B GSM8k 5-shot · p. 139

Llama2 7B gsm-hard 5-shot

• Llama2 7B answered the gsm-hard mileage problem incorrectly, with the correct answer identified as 13.The example presents a multi-step word problem involving Monday, Tuesday, and Wednesday mileage.Source: Llama2 7B gsm-hard 5-shot · p. 139

Code Llama 7B GSM8k 5-shot

• Code Llama 7B produced a correct, coherent 2600-gram solution for a GSM8k candy-weight problem.The calculation used 25 × 40 g chocolate bars and 80 × 20 g candied apples.Source: Code Llama 7B GSM8k 5-shot · pp. 139–140
• Code Llama 7B used the correct loan formula but gave an incorrect result, reporting an answer other than the correct balance of 220.The stated calculation was $2200 − 165 × 12, but the model missed the correct arithmetic.Source: Code Llama 7B GSM8k 5-shot · p. 140

Code Llama 7B gsm-hard 5-shot

• Code Llama 7B generated coherent code reasoning that correctly solved the die-probability problem.The example’s stated solution was 0.25, and the generated code reasoning was judged correct and coherent.Source: Code Llama 7B gsm-hard 5-shot · p. 140
• Code Llama 7B generated code that would produce 31879055, but its textual answer was 43581690.0.The generated code was judged correct even though the model gave the wrong answer in the text.Source: Code Llama 7B gsm-hard 5-shot · p. 141

Llama2 13B GSM8k 5-shot

• Llama2 13B produced a correct answer for a GSM8k ping-pong problem, but its reasoning was not coherent.The generated response attempted to compute the total from first- and second-half scores but contained malformed reasoning.Source: Llama2 13B GSM8k 5-shot · p. 141
• Llama2 13B answered the toy-valuation problem incorrectly; the correct total was 50.Its generated solution contained malformed algebraic expressions and repeated corrupted calculations.Source: Llama2 13B GSM8k 5-shot · pp. 141–142

Llama2 13B gsm-hard 5-shot

• The Llama2 13B gsm-hard example presents a generated paraphrase rather than a valid solution, and its final answer is explicitly marked incorrect.The output reframes the school-trip scenario with altered wording and does not recover the correct negative value.Source: Llama2 13B gsm-hard 5-shot · p. 142
• Llama2 13B incorrectly answers a gsm-hard school-trip problem; the correct answer is -7494668.0.The generated reformulation changes the original problem wording and fails to produce the correct calculation.Source: Llama2 13B gsm-hard 5-shot · p. 142

Llama3 7B GSM8k 5-shot

• On GSM8k, Llama3 7B gives the correct answer 60 for the recipe-instructions problem, but its reasoning contains errors.The response computes the second recipe as twice the first and then sums the steps, despite an erroneous intermediate expression.Source: Llama3 7B GSM8k 5-shot · p. 142

Cross-example observation

• Across the examples, answer correctness does not consistently imply sound reasoning: one correct response is erroneous, while another is coherent despite noisy redundancy.The GSM8k recipe response is correct but flawed, whereas the gsm-hard tree response is judged correct and coherent.Source: Llama3 7B GSM8k 5-shot · p. 142, Llama3 7B gsm-hard 5-shot · p. 144

Llama3 7B GSM8k 5-shot

• Llama3 7B incorrectly answers 40 instead of 57 on a GSM8k grocery-bill problem.Its response includes inconsistent arithmetic and malformed intermediate totals before stating the final answer as 40.Source: Llama3 7B GSM8k 5-shot · pp. 142–143

Llama3 7B gsm-hard 5-shot

• Llama3 7B produces a correct answer of 34.0 on a gsm-hard tree-height problem, with coherent though noisy and redundant reasoning.The model derives tree heights of 6 and 11 feet before reporting 34.0.Source: Llama3 7B gsm-hard 5-shot · pp. 143–144
• A second Llama3 7B gsm-hard example is incomplete and incorrect, with the correct answer reported as 322886700.The response describes an unfinished or incorrect solution rather than computing the requested weekly distance.Source: Llama3 7B gsm-hard 5-shot · p. 144

Llama3.1 7B GSM8k 5-shot

• On a GSM8k robe problem, Llama3.1 7B correctly computes 1 bolt of white fiber and 3 bolts total with coherent reasoning.The response explicitly calculates 2 × 0.5 = 1 and then 2 + 1 = 3.Source: Llama3.1 7B GSM8k 5-shot · p. 145
• The GSM8k examples therefore pair successful elementary arithmetic with an incorrect multi-step geometric-progression calculation.The supplied outputs show both a correct, coherent solution and a longer but erroneous discounted-price computation.Source: Llama3.1 7B GSM8k 5-shot · p. 145
• On a GSM8k glasses problem, the model applies a geometric progression to successive discounted prices but outputs $27.65 instead of the correct $64.The reasoning identifies a 3/5 common ratio and sums 16 terms, but the reported solution is marked incorrect.Source: Llama3.1 7B GSM8k 5-shot · p. 145

Llama3.1 7B gsm-hard 5-shot

• On a gsm-hard running-and-walking problem, Llama3.1 7B generates executable-style calculations and correctly returns 18.0 with coherent reasoning.The code derives running and walking speeds, allocates two hours to running and four hours to walking, and sums the distances.Source: Llama3.1 7B gsm-hard 5-shot · pp. 145–146
• The gsm-hard examples similarly show a correct rate-and-time solution alongside an incorrect financial calculation involving implausible tie-price arithmetic.The contrast is evidenced by the correct 18.0 example and the tie-cost output marked incorrect against 800.Source: Llama3.1 7B gsm-hard 5-shot · pp. 145–146
• On a gsm-hard ties problem, the model produces executable-style code but returns 4829138.0 instead of the correct answer, 800.The output is explicitly labeled incorrect despite computing blue-tie and red-tie costs.Source: Llama3.1 7B gsm-hard 5-shot · p. 146
• A further gsm-hard water example describes Python execution producing a correct remaining-water answer, but the supplied passage does not state its numeric result.The calculation subtracts two girls’ one-sixth shares and six liters from the total.Source: Llama3.1 7B gsm-hard 5-shot · p. 146

PoT examples across models

• PoT produced correct executable solutions for selected gsm-hard problems in Llama2, Code Llama, and Llama3.1 examples, but correctness was inconsistent across examples.Correct outputs include 3244047.0999999996, 83, and 6282701.800000001, while other generated programs returned explicitly incorrect answers.Source: Llama2 7B gsm-hard 5-shot PoT · p. 147, Code Llama 7B gsm-hard 5-shot PoT · p. 148, Llama3.1 7B gsm-hard 5-shot PoT · pp. 149–150

Llama2 7B gsm-hard 5-shot PoT

• Llama2 7B correctly computed the three-month download total as 3244047.0999999996, but its ice-cream program returned the wrong answer because it multiplied rather than divided by servings per carton.The correct ice-cream formula is days ÷ servings per carton × cost per serving.Source: Llama2 7B gsm-hard 5-shot PoT · p. 147

PoT limitations

• The examples indicate that PoT can encode multi-step arithmetic and execute calculations, yet correctness depends on selecting the problem’s intended formula and including all required quantities.Observed failures involve dividing versus multiplying, omitting terms, miscounting periodic events, and confusing simple with compound interest.Source: Llama2 7B gsm-hard 5-shot PoT · p. 147, Code Llama 7B gsm-hard 5-shot PoT · p. 148, Llama3 7B gsm-hard 5-shot PoT · p. 149

Code Llama 7B gsm-hard 5-shot PoT

• Code Llama 7B correctly solved the bridge-loading problem with 83 boxes, while its treasure program omitted diamonds and rubies from the final sum and returned the wrong answer.The correct treasure formula is result = diamonds + rubies + emeralds, yielding 4029727.Source: Code Llama 7B gsm-hard 5-shot PoT · p. 148

Llama3 7B gsm-hard 5-shot PoT

• Executing one code snippet returned the correct answer, -4358966; a separate code snippet returned the wrong answer, and its code set legs_per_5_days = 0.2.Source: Llama3 7B gsm-hard 5-shot PoT · p. 149

Llama3.1 7B gsm-hard 5-shot PoT

• Llama3.1 7B correctly computed the history-test score as 6282701.800000001 using weighted question categories, but its interest program incorrectly applied compound-interest exponentiation.The correct simple-interest formula is total interest = brenda investment ∗interest rate per year ∗num years, producing 23628338.5.Source: Llama3.1 7B gsm-hard 5-shot PoT · pp. 149–150

6 Conclusions

• LLMs have transformed NLP, achieving strong performance in text generation, question answering, translation, summarization, and related tasks while demonstrating emergent capabilities.The paper examines how LLMs emerged, evolved, and were applied across fields.Source: §6 · p. 150, Summary of Key Findings · p. 150

Summary of Key Findings

• Scaling Laws indicate super-linear performance growth with parameters when datasets are sufficiently large, but scaling incurs substantial computational and environmental costs.The paper argues that future work must balance model size and efficiency.Source: Summary of Key Findings · p. 150
• LLMs show promise in healthcare, finance, education, law, and scientific research, with Med-PaLM supporting clinical decisions and FinGPT analyzing trends and managing risks.The paper also discusses LLM-Modulo for planning and RAG for retrieval-augmented generation.Source: Summary of Key Findings · pp. 150–151

Reflection on Capabilities and Limitations

• LLMs can hallucinate, produce misleading information, and mimic rather than perform genuine reasoning, limiting reliability in diagnosis, legal interpretation, planning, and multistep problem-solving.The paper also notes repetitive token generation and uncertainty about whether recent reasoning advances represent a breakthrough toward AGI.Source: Reflection on Capabilities and… · p. 151

Concluding Thoughts

• Realizing LLMs’ potential to transform industries and advance human knowledge requires addressing their ethical, technical, and practical challenges.The paper emphasizes bias, harmful or misleading content, and environmental impact alongside the need for continued responsible development.Source: Reflection on Capabilities and… · p. 151, Concluding Thoughts · p. 152

Future Research Directions

• Future research should pursue parameter-efficient and specialized training, integrate external tools and knowledge, and combine cognitive science, linguistics, and computer science to improve reasoning and alignment.External databases, calculations, and up-to-date information may improve factual accuracy, although progress toward AGI remains uncertain.Source: Future Research Directions · p. 151