Challenges and Applications of Large Language Models

• The review asks which LLM problems remain unresolved and where LLMs are being applied, including how challenges constrain those applications.Source: §1 · p. 1
• It groups challenges into three categories—Design, Behavior, and Science—and explores applications in chatbots, computational biology, programming, creative and knowledge work, law, medicine, reasoning, robotics, and social sciences.Source: §1 · p. 2
• The paper is an opinionated review for technical researchers familiar with LLMs, focusing on text-trained models and excluding political, philosophical, and moral perspectives.Source: §1 · p. 2

2.1 Unfathomable Datasets

• Modern pre-training datasets are too large for thorough human quality assessment, leaving practitioners with only a nebulous understanding of their training data.Collection therefore relies heavily on heuristics about data sources and filtering.Source: §2.1 · p. 2
• Deduplication matters because near-duplicates degrade performance and memorization: NearDup reduced emitted memorization rates 10x on C4, while SemDeDup improved over NearDup.Lee et al. [294] report that over 1% of unprompted emitted tokens belonged to memorized sequences, including one repeated 61,036 times.Source: §2.1 · p. 2
• Benchmark contamination can inflate evaluation metrics through memorization, yet detecting all overlaps is difficult across web-crawled corpora and generated benchmarks.Dodge et al. [125] found input-and-label contamination in generation and knowledge-completion tasks and input-only contamination in GLUE; Jacovi et al. [237] propose encryption and training-exclusion controls.Source: §2.1 · p. 2
• Pre-training corpora may contain extractable PII, including phone numbers, emails, and secret API keys, because strong performance often depends on memorizing training data [146].Carlini et al. [65] [67], Lukas et al. [344], and Kulkarni demonstrated such extraction through prompting.Source: §2.1 · p. 3
• Domain and task mixtures remain difficult to balance: heterogeneous pre-training data broadly improves transfer, while poor mixtures can cause weak transfer or spurious correlations.Longpre et al. [341] find heterogeneity likely more important than measured document quality or size, whereas multitask fine-tuning can suffer negative transfer and catastrophic forgetting [239].Source: §2.1 · pp. 3–4

2.2 Tokenizer-Reliance

• Tokenizer reliance creates unequal costs because equivalent information requires different token counts across languages, potentially overcharging users while producing subpar results.Ahia et al. [13] and Petrov et al. [426] identify this disparity, especially affecting users where APIs are less affordable.Source: §2.2 · p. 4
• Tokenizer–corpus coupling causes glitch tokens from mismatched training data and requires retraining when corpora change, while multilingual and low-resource tokenization remains challenging.Greedy subword methods favor shared scripts and larger training-data segments, producing suboptimal tokenization for low-resource languages [92] [676].Source: §2.2 · p. 4
• Byte-level and tokenization-free alternatives show promising results, including Charformer’s comparable performance with less computational overhead and Canine’s advantage over a comparable tokenization-based encoder.Byte-level inputs still encode Latin languages more compactly than Chinese, Japanese, or Korean, while local patches can reduce byte-level computational costs.Source: §2.2 · pp. 5–6

High Pre-Training Costs

• Pre-training a single LLM can require hundreds of thousands of compute hours, millions of dollars, and energy equivalent to several typical US families’ annual use.Scaling laws imply diminishing returns: reducing error from 3% to 2% may require an order of magnitude more data or compute.Source: §2.3 · p. 6

Compute-Optimal Training Recipes

• OpenAI predicted full-size GPT-4 performance from smaller models using at most 10,000x less compute, supporting compute-optimal scaling recipes.However, scaling coefficients disagree on whether model size or data should grow faster, and pre-training loss may not predict downstream performance.Source: §2.3 · p. 6
• Performance prediction remains difficult because scaling laws differ between upstream and downstream settings, with model shape affecting downstream fine-tuning beyond model size.Inference costs also motivate training smaller models more intensively to reduce future usage expenses.Source: §2.3 · p. 6

Pre-Training Objectives

• Pre-training objectives trade data efficiency against task suitability: bidirectional attention favors NLU, next-token prediction favors NLG, and intermediate objectives seek both.Span corruption speeds training through shorter sequences, while Mixture of Denoisers improves continued training for relatively few steps.Source: §2.3 · pp. 7–8

Parallelism Strategies

• Pipeline parallelism, exemplified by GPipe [226], reduces waiting and improves computational utilization relative to naive model parallelism.Asynchronous methods can improve utilization further but may suffer instability from stale gradients.Source: §2.3 · p. 9

Miscellaneous

• Stacking layers reduced compute by 40%, while weight averaging saved up to 4,200 GPU hours for a 6.9B-parameter model.Smaller-model speedup algorithms have not yet scaled and offer limited gains against budget-adjusted baselines.Source: §2.3 · pp. 9–10

Large Memory Requirements

• Full-model fine-tuning requires large memory for parameters, activations, gradients, and optimizer statistics, often restricting access to institutions with large device clusters.Fine-tuning entire LLMs requires memory comparable to pre-training, making it infeasible for many practitioners.Source: §2.4 · p. 10

Overhead of Storing and Loading Fine-Tuned LLMs

• Full-model adaptation also requires storing and loading a separate complete model copy for each task, creating computational and memory overhead.The fixed base model in PEFT can instead be reused across tasks during inference.Source: §2.4 · p. 10, Figure 5 · p. 11

Parameter-efficient fine-tuning

• Adapters match full fine-tuning competitively on 26 text-classification tasks while updating 3% of parameters, whereas bias-only tuning updates less than 1% with competitive GLUE performance.Adapters add learnable layers while keeping the rest of the network fixed; bias-only tuning updates only model bias terms.Source: §2.4 · p. 10
• Prompt-tuning achieves competitive performance with fewer than 100 safety-classification examples and five multilingual question-answering examples, while updating only learned soft prompts.Soft prompts contain thousands rather than millions of parameters, though tuning them still requires backpropagation through the network.Source: §2.4 · pp. 10–11
• Quantized LoRA enables fine-tuning a 65B model on one 48GB GPU, whereas regular training requires over 780GB of GPU memory.Quantized LoRA reduces memory usage by representing Transformer weight updates with an additive low-rank decomposition.Source: §2.4 · p. 11

Compute Requirements

• Despite reducing memory complexity, PEFT still requires full-network forward and backward passes and gradient computation, limiting personalization on smaller devices.MeZO reduces memory to the inference footprint and supports non-differentiable objectives, while (IA)3 outperforms full fine-tuning using 0.01% of parameters.Source: §2.4 · p. 11

2.5 High Inference Latency

• LLM inference latency arises from one-token-at-a-time decoding, large model and transient-state memory footprints, and attention’s quadratic scaling.Transient states include attention key and value tensors; attention scaling is discussed separately in Sec. 2.6.Source: §2.5 · p. 11
• Efficient attention reduces memory bandwidth or attention complexity through multi-query attention, FlashAttention [107], sharding [432], sparsification, and approximations.Performer [85] can severely underperform vanilla self-attention, especially for large models [220].Source: §2.5 · p. 12
• Quantization reduces memory or increases throughput, with LLM.int8() [117] preserving performance through Int8 weights and higher-precision handling of outliers.Other methods compress weights to 3–4 bits, enabling 175B-parameter models to run on a single GPU [153].Source: §2.5 · p. 12
• Pruning removes model weights, but unstructured sparsity is harder than structured sparsity to convert into practical hardware speedups.SparseGPT [152] prunes up to 60% of parameters while roughly preserving performance, and Wanda matches its performance with one forward pass; both support n:m sparsity and GPU speedups.Source: §2.5 · pp. 12–13
• Mixture-of-Experts models accelerate inference by activating only selected experts, reducing computation and potentially communication across devices.Switch Transformer uses one expert and achieved up to 7x pre-training speedups, while sMLP achieved a 2x pre-training speed improvement; MoE models can suffer expert collapse.Source: §2.5 · p. 13
• Implicit activation sparsity increases with model size, while post-training MoEfication can speed inference by 2x.Cascading differently sized models also dominates the accuracy–cost Pareto frontier [75].Source: §2.5 · p. 13
• Speculative sampling accelerates decoding by using a smaller draft model before the larger target model verifies and accepts tokens.Other latency-oriented strategies include blockwise parallel generation and skipping earlier layers for tokens that are easier to predict later in sequences [114].Source: §2.5 · p. 13

2.5.1 Software

• Software frameworks reduce LLM costs through optimized implementations, memory pooling, parallelism, and distributed or decentralized execution.FlexGen enables inference with 175B-parameter models on one GPU using aggregated GPU, CPU, and disk resources; vLLM uses PagedAttention to reduce decoding memory and transfers.Source: §2.5.1 · p. 14

Limited Context Length

• Long-context LLMs can degrade severely despite architectural support, with decoder-only models showing a U-shaped performance curve that favors relevant information at the beginning or end.Commercial closed-API models often fulfill their long-context promise, whereas many open-source models do not; GPT-3.5 struggles to retrieve information from the middle of the context.Source: §2.6 · p. 14

Efficient Attention Mechanisms

• Efficient attention mechanisms reduce the quadratic cost of long inputs, enabling contexts from 12,000 tokens with Transient Global attention to 64,000 tokens with CoLT5.Other approaches report 4.9x–12.1x autoregressive speedups, linear complexity, or hierarchical processing of inputs up to 1B tokens.Source: §2.6 · p. 15

Length Generalization

• Length generalization remains largely unsolved because positional embeddings can make performance depend on sequence positions unseen during training, despite progress from relative encodings and ALiBi.Causal LLMs without positional encodings can nevertheless remain competitive, possibly because the causal attention mask leaks positional information.Source: §2.6 · p. 15
• Interpolation plus brief fine-tuning extends RoPE-based pretrained LLMs to context windows of up to 32,768 tokens, while ALiBi matches longer-trained sinusoidal models at doubled test length.However, ALiBi, RoPE, and absolute positional encodings perform poorly on length generalization across 10 reasoning tasks, and ALiBi shows no large gains in Galactica.Source: §2.6 · p. 16

Transformer Alternatives

• Transformer alternatives use state-space, convolutional, hybrid, or recurrent designs to scale context efficiently while approaching Transformer quality, including Hyena at 131,000 tokens and RWKV at 14B parameters.H3 addresses SSM weaknesses in recalling and comparing tokens; Block-State Transformer reports up to 10x speedups; RWKV matches similarly sized Transformer LLMs.Source: §2.6 · p. 17

Prompt Brittleness

• Prompt wording, example selection, ordering, length, and formatting can cause dramatic output changes, leaving robustness to prompt style and format unresolved.Practitioners therefore design prompts ad hoc, with limited theoretical understanding of why particular phrasings work.Source: §2.7 · p. 17

Single-Turn Prompting

• Prompting methods include in-context learning, Chain-of-Thought reasoning, zero-shot and automatically generated reasoning variants, and expert impersonation, which improves domain-specific accuracy.Chain-of-Thought constructs intermediate reasoning steps, while Salewski et al. [473] report gains from asking models to impersonate domain experts.Source: §2.7 · pp. 17, 19

In-Context Learning

• Few-shot prompting substantially outperforms zero-shot learning, even when demonstration labels are randomly flipped, suggesting demonstrations convey task labels and input distributions.Min et al. argue that input-label associations are not decisive for performance, whereas demonstrations remain useful for identifying the task and its distribution.Source: §2.7 · p. 18
• LLMs are highly sensitive to the permutation order of few-shot demonstrations, despite evidence that their input-label associations may not drive few-shot performance.Lu et al. show that changing demonstration order affects model behavior.Source: §2.7 · p. 18

Multi-Turn Prompting

• Multi-turn methods iteratively compose prompts and answers, including majority-vote aggregation, self-consistency, decomposition, external tool use, refinement, and Tree of Thoughts search.Tree of Thoughts [639] maintains multiple reasoning paths and supports self-evaluation, lookahead, breadth-first or depth-first search, and backtracking.Source: §2.7 · p. 19

Controlled Generation

• Controlled generation modifies inference rather than prompt text, using classifier-free guidance, prompt changes during decoding, or related sampling strategies to steer fixed-prompt outputs.Sanchez et al. [474] up-weight prompt importance during generation, while Roush considers alternating prompts; negative prompting has produced negative initial results.Source: §2.7 · p. 19

Hallucinations

• Hallucinations are fluent outputs that contradict source content intrinsically or remain unverifiable because the source underdetermines correctness extrinsically.Extrinsic hallucinations are not necessarily erroneous, but their claims cannot be grounded or contradicted by the provided information.Source: §2.8 · p. 20

How to Measure Hallucinations

• Factuality evaluation is difficult because generations mix supported and unsupported information, motivating atomic-fact decomposition and external evidence checks.FactualityPrompts uses factual and nonfactual prompts with named-entity and textual-entailment metrics, while hallucination snowballing describes overcommitment to early mistakes [365] [664].Source: §2.8 · p. 20

Retrieval Augmentation

• Retrieval augmentation separates knowledge storage from processing by retrieving top-k documents and supplying them with the prompt, supporting modular updates and interpretability.Approaches include REALM, RAG, adaptive memory architectures, ATLAS [236], large retrieval databases [52], and instruction-trained retrieval datasets.Source: §2.8 · pp. 20–21
• Over 60% on average and up to 85% on out-of-distribution data, retrieval-augmented models reduce hallucinated GPT-3 responses.Shuster et al. [507] study retrieval-augmented architectures, which ground inputs in externally retrieved documents.Source: §2.8 · p. 20
• Standard retrieval augmentation can still hallucinate after finding relevant documents, and retrievers may fail to locate passages that directly answer complex questions [262].A browsing example retrieves two related papers but produces an inaccurate title and summary for the second paper.Source: §2.8 · p. 21

Decoding Strategies

• In response generation, increased diversity is positively correlated with hallucinations [136], and the trade-off between diversity and quality remains largely unsolved.Diverse beam search and confident decoding [552] try to reduce decoding-level hallucinations, while uncertainty-aware beam search [620] penalizes high predictive uncertainty during decoding.Source: §2.8 · p. 22

Alignment methods

• Alignment methods include human-feedback pre-training, instruction fine-tuning, and RLHF, while detection uses evaluation, auditing, red teaming, and mechanistic interpretability.RLHF learns a reward model from human-ranked generations and optimizes the language model against it, often after supervised fine-tuning for complex tasks.Source: §2.9 · pp. 22–23
• RLHF aligns models across summarization, assistant behavior, and tool use, but can amplify political or religious views and increase stated resistance to shutdown.DPO removes reward-model fitting and reinforcement learning, while fine-tuning on 1,000 selected prompts and responses can outperform RLHF-trained models, motivating the Superficial Alignment Hypothesis.Source: §2.9 · p. 23

Self-improvement

• Self-improvement can align models through self-critique, feedback-conditioned training, and automatic repair of unethical outputs, while also supporting tool use, coding, and computer tasks.Chain of Hindsight improves human-preference alignment, summarization, and dialogue; Toolformer self-generates and filters examples for API use.Source: §2.9 · pp. 23–24

Evaluation and auditing

• Automated evaluations agree highly with crowd workers and uncover diverse harmful behaviors, including inverse scaling, stronger political views, and greater desires for resource acquisition or goal preservation.Evaluation and auditing are intended to inform training, deployment, security, and policy decisions, including risks from offensive cyber capabilities and manipulation.Source: §2.9 · p. 24

Red teaming

• Red teaming finds harmful outputs, privacy leaks, and training-data leakage; RLHF models become harder to red-team as they scale, whereas other models do not.Methods increasingly automate adversarial prompt generation, Bayesian test selection, or red teaming from abstract specifications when harmful behavior cannot be precisely classified beforehand.Source: §2.9 · pp. 24–25

Emergent capabilities

• Emergent capabilities matter for safety because deception may arise at larger model sizes in both pretrained and RLHF models, although some reported emergence may reflect metric choices.Models can have strategic incentives to deceive, and phishing campaigns show deceptive behavior can already be elicited from LLMs.Source: §2.9 · p. 25

Mechanistic interpretability

• Mechanistic interpretability reverse-engineers model components and can detect malicious inputs with high accuracy, recover latent knowledge, and reduce prompt sensitivity by half.Recent methods identify behavior-explaining subgraphs, decode layerwise hidden states, and recover knowledge without human supervision or model outputs.Source: §2.9 · pp. 25–26

Risks and open problems

• LLMs inherit social biases and face ambiguous toxicity judgments, while prompt injections can leak private data or override safeguards; possible agentic goals and beliefs add further safety risks.Bias evaluations cover race, gender, ethnicity, and age and include a debiasing toolkit; prompt-injection failures are attributed to competing safety and capability objectives.Source: §2.9 · pp. 26–27

Challenge

• Outdated or inaccurate pre-training facts are difficult to correct because retraining is expensive and fine-tuning may not reliably replace old knowledge.These challenges are especially relevant when information changes over time, such as political leadership.Source: §2.10 · p. 27

Model editing limitations

• Existing model editing methods have limited effectiveness for isolated knowledge updates and can create unintended associations [642] [205].Their low specificity makes it difficult to update one faulty fact while consistently changing related information without altering unrelated knowledge; Hoelscher-Obermaier et al. [205] report unintended associations.Source: §2.10 · p. 27

Approaches

• Model editing modifies model parameters or uses an external post-edit model, while retrieval-augmented language modeling relies on updateable non-parametric knowledge sources [642].The latter approach is omitted here and discussed in Sec. 2.8.Source: §2.10 · p. 27

Modifying Model Parameters

• Parameter-modification methods either locate and edit the relevant weights [102] or use meta-learning to predict weight updates [111].These are respectively locate-then-edit and meta-learning approaches.Source: §2.10 · p. 27

Preserving Model Parameters

• Parameter-preserving methods use post-edit models or add new weights [127] [227], including adapters that selectively apply edits in latent space [191].Hartvigsen et al. [191] use a similarity-based mechanism to decide when the adapter should be used.Source: §2.10 · p. 27

Model editing limitations

• Model editing methods show weak non-trivial generalization and variable performance across architectures, with ROME and MEMIT working well empirically only on decoder-only LLMs.These limitations reduce the methods’ performance and applicability across model architectures [642].Source: §2.10 · p. 27

Retrieval-augmented language modeling

• Retrieval-augmented models can reflect updated knowledge at inference time by swapping non-parametric indices, including information about changed world leaders and temporal updates in both directions [236].Lewis et al. update knowledge by replacing the model’s non-parametric memory, while Izacard et al. [236] swap the retrieval index.Source: §2.10 · p. 27

• Small task or prompt modifications can reverse an LLM’s result, making holistic assessment difficult without rigorous benchmarks covering diverse inputs.Source: §2.11 · p. 27
• Holistic suites such as HELM [318] standardize evaluation across scenarios and tasks while seeking broad coverage of capabilities and risks.Source: §2.11 · p. 27
• GPT-4 achieves human-level performance on several SAT, LSAT, and mathematics competition tests in the AGIEval benchmark [679].Source: §2.11 · p. 28
• MMLU results vary significantly with evaluation method, while larger and instruction-fine-tuned models are more sensitive to minor prompt variations [197].Source: §2.11 · p. 28

Reliance on Static, Human-Written Ground Truth

• Static, human-written ground truth is scarce for expert domains, while surpassing human benchmark performance makes human-level comparisons harder.This limits evaluation in programming and mathematics and other domains where expert-written answers are difficult to obtain.Source: §2.12 · p. 28
• Older static benchmarks become outdated, saturated, or overfit as LLM capabilities change, requiring continual replacement or more dynamic evaluation measures [113] [447] [263].Updating benchmarks often still depends on human-written ground truth.Source: §2.12 · p. 28

Dynamic Evaluation

• Dynamic evaluation work adds tasks to BIG-Bench, including programmatically evaluated tasks, and develops evaluations without humans in the loop [519].These approaches aim to keep evaluation aligned with changing model capabilities.Source: §2.12 · p. 28

Model-generated evaluation tasks

• Model-generated tasks can cover arbitrary evaluation axes, but candidate order can strongly affect assessments, motivating evidence-first prompting and averaging scores across swapped positions [581].Reward models trained on human preferences can filter generated datasets for quality.Source: §2.12 · p. 28

Model-generated scores

• Model-generated judges can grade other models when producing correct answers is difficult, but results vary by judge model and remain insufficiently robust to replace human judgment [325] [586] [667] [156].The approach relies on evaluating answer correctness or relative quality rather than generating a reference answer.Source: §2.12 · pp. 28–29

Motivation

• Detection matters for limiting misinformation, plagiarism, impersonation, scams, and inferior generated text entering future training data, yet improved fluency makes classification more difficult.These risks include fake authoritative narratives, plagiarism-tool bypasses, style imitation, phishing, and accidental training-data contamination [574] [602] [439].Source: §2.13 · p. 29

Post-hoc Detectors

• Post-hoc detectors classify arbitrary text using token improbability visualization, energy-based discrimination, authorship attribution, or DetectGPT’s negative-curvature hypothesis.DetectGPT identifies generated passages by approximating regions where minor rewrites receive lower model probability than the original sample.Source: §2.13 · p. 29

Watermarking

• Watermarking embeds hidden, algorithmically identifiable token patterns through green-list sampling, with soft variants for low-entropy tokens and methods for black-box or secret-key detection.Kirchenbauer et al. [268] seed vocabulary partitions by the previous token; related work addresses black-box APIs, identifiable distributions, and secret-key-only detection.Source: §2.13 · p. 29
• Watermarks remain detectable after human rewriting, non-watermarked rewriting, or insertion into longer human documents given sufficient tokens, but detection reliability remains disputed.Sadasivan et al. [469] argue detection is unreliable in practice and theory, including watermark extraction and spoofing attacks that misclassify human text.Source: §2.13 · pp. 29–30

Paraphrasing Attacks

• Paraphrasing reduced DetectGPT’s detection accuracy from 70.3% to 4.6%, while semantic-retrieval defenses remained fairly robust because paraphrasing preserves meaning.Krishna et al. [280] trained an 11B paraphrase model with controls for lexical diversity and reordering, then proposed storing generations for later semantic retrieval.Source: §2.13 · p. 30

2.14 Tasks Not Solvable By Scale

• Inverse scaling can reflect memorization, undesirable pattern imitation, distractor-task learning, or spurious few-shot correlations rather than a permanent scaling limit.Across 11 datasets, four retained inverse scaling at higher compute while six became U-shaped, with performance eventually improving; testing methodology may also explain some effects.Source: §2.14 · p. 30
• Compositional and formal-semantic tasks expose limits of shortcut-based learning: larger models improve subproblem solving without closing the compositionality gap, while systematic reasoning and quantifier understanding may not emerge.Proposed mechanisms include linearized path matching, compounded early errors, and insufficient understanding of universal quantifiers.Source: §2.14 · pp. 30–31
• Counterfactual tasks reveal a memorization-like effect: LLM performance worsens as altered input-output conditions become less common, leaving unclear whether scaling increases memorization or improves coverage.The paper identifies proportional dataset scaling as a possible source of more uncommon-condition examples, but presents this as a future direction.Source: §2.14 · p. 31

2.15 Lacking Experimental Designs

• Many LLM papers lack controlled ablations, impeding scientific comprehension because large design spaces make it difficult to attribute performance differences to individual factors.Examples include untested claims about corpus quality, inherited untuned hyperparameters, and omitted implementation details such as optimizer-state resets.Source: §2.15 · p. 31
• One costly remedy is varying one factor at a time, while zero-shot hyperparameter transfer with µP was empirically verified up to a 6.7B model but remains unverified for other factors.The suggested controlled variations include model size and context length.Source: §2.15 · p. 31
• LLM experimental design spaces combine data, tokenizer, objective, architectural, and hyperparameter choices, causing factorial experiments and training costs to grow combinatorially.Insufficient coverage can therefore inhibit understanding of what makes an LLM perform well, especially given large parameter counts and long feedback cycles.Source: §2.15 · p. 33
• When full experimentation is infeasible, the paper recommends richer evaluation distributions and efficient design-space methods such as dimensionality-reduced Bayesian optimization and treatment-effect estimation.Distributional reporting is described as less biased and more robust to outliers than point estimates.Source: §2.15 · pp. 31, 33

Lack of Reproducibility

• The paper defines repeatability as the ability to reproduce experimental outcomes exactly and distinguishes it from broader uses of “reproducibility”.Non-reproducible results can waste resources when researchers build upon them.Source: §2.16 · p. 33

Training Repeatability

• Distributed LLM training is nondeterministic because parallel scheduling, lock-free schemes, floating-point summation order, and optimized operations can alter results.These effects are especially consequential for algorithms such as stochastic gradient descent that are not order-invariant.Source: §2.16 · p. 33
• Reproducing pre-training data can fail when users crawl changing web sources from dataset indexes rather than downloading static dumps [64].Monetary, privacy, and legal restrictions can require this decentralized collection process.Source: §2.16 · p. 33

Inference Repeatability

• API-served models are difficult to repeat because providers can modify, retrain, or replace models, discontinue older versions, and produce stochastic outputs even at zero temperature.These models are typically accessed through stochastic, black-box APIs, limiting experimental control.Source: §2.16 · pp. 33–34
• GPT-4’s prime-number accuracy fell from 97.6% in March 2023 to 2.4% in June, while GPT-3.5 improved over the same period [76].Across four diverse tasks, GPT-3.5 and GPT-4 performances varied substantially within three months.Source: §2.16 · p. 34

3.1 Chatbots

• Chatbot RLHF outperforms dialogue-prompted and supervised-fine-tuned approaches on output preference and rule-violation rate, while latency and multi-turn memory remain constraints.Chatbots combine retrieval, multi-turn interaction, and generation; synthetic conversations can fine-tune models with performance only slightly below human-generated datasets.Source: §3.1 · pp. 34, 36

3.2 Computational Biology

• LLMs support computational biology through sequence embeddings for structure prediction, sequence generation, classification, mutation analysis, protein design, and antibody evolution.These applications use nontext biological sequences but often evaluate on smaller or specialized datasets rather than real-world design settings.Source: §3.2.1 · p. 36, §3.2 · p. 36

3.2.1 Protein Embeddings

• xTrimoPGLM-100B outperforms existing approaches on 13 of 15 evaluated protein tasks, while protein-language-model applicability to drug design remains unclear.ESM-2 scaling improves structure-prediction performance, whereas ESMFold trades lower benchmark performance than AlphaFold2 for an order-of-magnitude faster inference time.Source: §3.2.1 · pp. 36–37

3.2.2 Genomic Analysis

• HyenaDNA models genomic sequences up to 1 million tokens and outperform Transformer-based models with orders of magnitude more parameters.Current genomic models remain constrained by genomes whose sequences exceed available context windows.Source: §3.2.2 · p. 37

3.3 Computer Programming

• Computer-programming LLMs generate and complete code across languages, including Python, multilingual programming, competitive programming, and repository-level tasks.Programming applications increasingly also use general chatbots partially trained on code datasets.Source: §3.3 · p. 37, §3.3.1 · p. 38

3.3.1 Code Generation

• Code-generation LLMs produce programs from specifications, with filtered-data Phi-1 achieving near-SOTA HumanEval performance using over an order of magnitude fewer parameters and tokens.Repository-scale code generation remains limited by context windows, motivating retrieval-based frameworks and API definitions that supply relevant context more compactly.Source: §3.3.1 · pp. 38–39

3.3.2 Code Infilling and Generation

• SantaCoder outperforms InCoder on both HumanEval generation and infilling, while infilling supports modifying, reviewing, and debugging existing code.InCoder uniquely performs single- and multi-line infilling despite underperforming the best Codex and CodeGen models on HumanEval generation.Source: §3.3.2 · p. 39

3.4 Creative Work

• Creative-work applications generate stories and scripts using hierarchical prompting, recursive revision, detailed outlining, candidate ranking, and controllers to maintain coherence over long outputs.Dramatron supports complete scripts without requiring a human in the loop, while Re3 and DOC target stories extending thousands of words.Source: §3.4 · pp. 39–40
• LLMs support collaborative poetry, stories, news reels, interactive fiction, idea generation, and controllable image generation, with idea-generation performance comparable to 100 human participants.Long-form creative applications commonly decompose work into short-form subtasks because limited context windows prevent retaining the entire generated work.Source: §3.4 · p. 40

3.5 Knowledge Work

• On DM/CNN and XSUM, instruction-fine-tuned InstructGPT performs best across faithfulness, relevance, and coherence, while zero-shot InstructGPT-3 performs comparably to six freelance writers across all three metrics.On qualitative sections, GPT-3.5 achieved 50–70% accuracy, significantly ahead of random guessing and approaching human-level scores. BloombergGPT outperforms OPT, GPT-NeoX, and BLOOM on financial domain-specific tasks.Source: §3.5 · p. 41

3.6.1 Legal Question Answering and Comprehension

• Zero-shot GPT-4 achieves SOTA and passing scores on the full UBE, while GPT-3.5 achieves SOTA on COLIEE legal entailment and SARA but varies substantially with prompting.GPT-3.5 scored 50% on the UBE multiple-choice component, and ChatGPT performed like a passing C+ law student; legal data also become outdated as laws and precedents change.Source: §3.6.1 · p. 42

3.6.2 Case Prediction and Legal Text Generation

• GPT-based case prediction does not match SOTA, while Transformer-XL and GPT-2 judicial-text models generate opinions that human participants struggle to distinguish from genuine text.Earlier attention-based models outperform bag-of-words SVMs on several ECHR prediction tasks, indicating potential despite sparse LLM usage.Source: §3.6.2 · pp. 42–43

3.7 Medicine

• Med-PaLM 2 achieves current SOTA on MultiMedQA, while GPT-4 significantly outperforms GPT-3.5 across medical benchmarks but remains vulnerable to erroneous and biased generations.Medical applications span question answering, information extraction, indexing, triage, and health-record management, but hallucinations and bias constrain safety-critical use.Source: §3.7.1 · p. 43, §3.7 · p. 43

3.7.2 Medical Information Retrieval

• A distilled PubMedBERT model outperforms GPT-3.5 and GPT-4 on adverse drug-event extraction and matches specialized supervised models.Medical information-retrieval systems use prompting, structured-output conversion, specialized T5 models, or distillation to extract dosages, disambiguate acronyms, and identify entities and relations.Source: §3.7.2 · p. 44

3.8 Reasoning

• Instruction fine-tuning, self-improvement, and chain-of-thought prompting improve mathematical reasoning, but the best GPT-3.5 model produces valid Blocksworld plans only 5% of the time versus 78% for humans.Process-based fine-tuning yields fewer reasoning errors despite similar final-answer correctness, while causal-reasoning results are inconsistent and some models perform near random on implicatures.Source: §3.8 · pp. 44–45
• Reasoning applications extend beyond mathematics to analogy, theorem proving, causal inference, and planning, but existing models often fail to generalize or match human performance.GPT-3.5 generally outperforms undergraduates on matrix and verbal analogies, yet performs poorly on letter-string analogies, causal tasks, implicatures, and common-sense planning.Source: §3.8 · p. 45
• LLMs show some complex reasoning ability, but perform relatively poorly on several reasoning tasks and benchmarks compared with their human-level performance elsewhere [164] [61] [263].They may recite causal knowledge embedded in training data rather than perform causal reasoning [253].Source: §3.8 · p. 45

3.9 Robotics and Embodied Agents

• In robotics, LLMs provide high-level planning by translating instructions into executable tasks, incorporating vision and environment summaries, or generating code with human oversight [14] [129].PaLM-E outperforms the original SayCan results by using images for state prediction, failure awareness, and long-horizon planning [129], while GPT-4-based VOYAGER autonomously explores Minecraft, acquires skills, and completes tasks beyond ReAct, Reflexion, and AutoGPT [579].Source: §3.9 · pp. 45–46

3.10 Social Sciences & Psychology

• Social-science applications use LLMs to model human behavior, analyze behavioral characteristics, and simulate social relationships [22] [176] [211] [126] [367] [409].Figure 15 groups these applications into human-behavior modeling, analysis of LLM behavior, and social-relationship simulation [12, 211, 414, 408].Source: §3.10 · p. 46, Figure 15 · p. 46

3.10.1 Modeling Human Behavior

• GPT-3 often approximates human behavioral responses, but results range from close alignment to replicating only 37.5% of human effects, with training exposure and bias posing risks [22, 211, 176, 12, 270].Aher et al. [12] found larger models aligned more closely with humans, whereas Park et al. [409] attributed weaker replication to fundamentally different cognitive systems.Source: §3.10.1 · pp. 46–47

3.10.2 Analyzing Behavioral Characteristics of LLMs

• LLM personality and cognition assessments find human-aligned traits, temperature-dependent personalities, creativity trade-offs, and uneven Theory-of-Mind performance across models [367] [523] [162].GPT-3.5 and GPT-4 performed like 6- and 7-year-old children on ToM tasks, respectively, while most tested models struggled with basic ToM and GPT-4 was closest to humans [162].Source: §3.10.2 · pp. 47–48

3.10.3 Simulating Social Relationships

• LLM-based communities of 25 artificial agents can develop emergent social behaviors, including relationships and event attendance, without human interaction.Agents interact through natural language in a digital environment while retaining distinct character traits.Source: §3.10.3 · p. 48

3.11 Synthetic Data Generation

• GPT-3-generated labels can reduce labeling costs and train smaller models; they outperform human labels at small budgets, but higher-quality human labels produce better models at higher budgets across nine NLP tasks.InPars-generated retrieval examples outperform only pre-trained models and perform worse than models fine-tuned on MS MARCO.Source: §3.11 · pp. 48–49
• Synthetic data generation benefits from randomness and coherent examples, but generating inputs and targets risks unrepresentative distributions and model collapse [182].Labeling existing data reduces this risk, although hallucinated labels remain possible; fully or partially synthetic datasets make hallucinated distributions more significant [123] [182] [503].Source: §3.11 · p. 49

4 Related Work

• Concurrent surveys address LLMs broadly [673], augmented language models [363], AutoML applications [555], generated-text detection [539], and evaluation tasks [72].These works cover associated topics, tool-using and reasoning-capable models, methods for improving LLMs, detection techniques, and evaluation tasks.Source: §4 · p. 49
• Prior surveys and evaluations also examine LLM applications in chatbots, computational biology [217], programming, medicine, law, knowledge work [140], and reasoning [223].The literature includes work focused on individual application domains rather than only general LLM topics.Source: §4 · p. 49

• The authors aim to guide future research by highlighting existing limitations and support cross-domain idea transfer through an overview of approaches used in applied areas.The paper also targets further research across application domains.Source: §5 · p. 49