Computer Science Literature Review Example: Annotated Sample

In computer science, "literature reviews" are usually called survey papers or state-of-the-art (SOTA) reviews. They do not follow the APA-style paragraph structure common in social sciences. Instead, CS reviews focus on comparing methods, benchmarking results, and identifying unsolved technical problems. The writing is dense, precise, and often supported by comparison tables.

This page gives you a complete illustrative paragraph from a master's-level CS literature review, with every sentence labeled for its function.

Annotated Computer Science Literature Review Sample

Topic: Transformer-Based Models for Code Generation: A Comparative Analysis

Sentence 1

"Since the introduction of the Transformer architecture (Vaswani et al., 2017), neural code generation has shifted from recurrent models (LSTMs, GRUs) to attention-based approaches that jointly model source code syntax and natural language intent."

Sentence 2

"Codex (Chen et al., 2021) demonstrated that large-scale language models pretrained on public GitHub repositories could generate functionally correct code, achieving 28.8% pass@1 on HumanEval — a milestone that established pretrain-and-finetune as the dominant paradigm."

Sentence 3

"Subsequent models — including AlphaCode (Li et al., 2022), CodeT5 (Wang et al., 2021), and InCoder (Fried et al., 2022) — have iterated on this paradigm by incorporating fill-in-the-middle training, larger context windows, and multi-language pretraining, yet none have substantially improved pass@1 beyond Codex's baseline on competitive programming benchmarks."

Sentence 4

"A persistent technical limitation across these models is their inability to reason about program semantics beyond surface-level token patterns — they generate syntactically plausible code that fails on execution because they do not model control flow, data dependencies, or runtime behaviour (Jain et al., 2022; Li et al., 2022)."

Sentence 5

"No existing model has successfully integrated static analysis or symbolic execution techniques into the generation pipeline to enforce semantic correctness during decoding — a gap this thesis addresses through a neuro-symbolic approach combining Transformer-based generation with constraint-guided beam search."

Annotation Summary Table

Key insight: Sentence 5 combines the gap and study positioning in one sentence. In CS, this is standard because the field is solution-oriented — you do not just identify a problem and leave it for future work; you immediately propose your technical approach. Compare this to psychology (gap and positioning are separate sentences) or nursing (gap is PICO-framed before positioning).

What Makes Computer Science Literature Reviews Unique

• Benchmark-driven evidence Sentence 2 gives exact metric: "28.8% pass@1 on HumanEval"

• Model comparison, not study comparison Sentence 3 compares Codex vs. AlphaCode vs. CodeT5 — models are the unit of analysis

• Technical limitation as gap pathway Sentence 4: models cannot reason about semantics — this is a technical, not empirical, flaw

• Gap + solution in one sentence Sentence 5 names the missing technique and the proposed approach simultaneously

• Foundational paper as anchor Sentence 1 cites Vaswani et al. (2017) — the paper that started the paradigm

Common Mistakes in CS Literature Reviews

• No benchmark numbers — only qualitative claims about model performance "Codex performed well on code generation tasks" Give the metric: "Codex achieved 28.8% pass@1 on HumanEval (Chen et al., 2021), establishing the pretrain-and-finetune paradigm's ceiling on competitive programming."

• Comparing models with vastly different parameter counts without acknowledging scale Comparing a 125M-parameter model's result directly to a 12B-parameter model's without comment Acknowledge confounding: "While Codex (12B) outperforms CodeT5 (220M) on HumanEval, the comparison is confounded by parameter scale — architectural contributions must be isolated through controlled ablations."

• Gap is "more data" or "larger model" instead of a technical limitation "Future work should train on larger datasets" Technical gap: "No existing generation pipeline has integrated static analysis (e.g., abstract syntax tree validation) into the decoding process to enforce semantic correctness at inference time."

• Not citing the foundational architecture paper Writing about Transformers without citing Vaswani et al. (2017) Always anchor the paradigm: "Since the introduction of the Transformer architecture (Vaswani et al., 2017), neural code generation has shifted from recurrent to attention-based models."

• Ignoring concurrent work from the same release window Only citing the most famous model and omitting contemporaneous alternatives Map the full landscape: "Within a 12-month period, Codex (Chen et al., 2021), AlphaCode (Li et al., 2022), and InCoder (Fried et al., 2022) all proposed different approaches to the same problem — their differences lie in training objectives, not just scale."

CS Writing Formula (Discipline-Specific Logic)

CS SOTA reviews follow a paradigm → flagship → iteration ceiling → technical limitation → solution proposal logic. Crucially, gap and solution are often combined in one sentence because the field is engineering-oriented.

Sentence 1 (Paradigm anchor): Since [foundational architecture paper] (Author, Year), [task domain] has shifted from [old paradigm] to [new paradigm].

Sentence 2 (Flagship model + benchmark): [Flagship Model] (Author, Year) demonstrated [capability], achieving [exact metric on exact benchmark] — establishing [approach] as the dominant paradigm.

Sentence 3 (Iteration ceiling): Subsequent models — including [A], [B], [C] — have [technical contribution: training objective, architecture change], yet none have substantially improved [metric] beyond [baseline] on [benchmark].

Sentence 4 (Technical limitation): A persistent limitation is [fundamental technical flaw — what the approach cannot do at an architectural level] (Author, Year).

Sentence 5 (Gap + solution combined): No existing [system/architecture] has integrated [missing technical component] into [pipeline stage] — a gap this thesis addresses through [proposed method: e.g., neuro-symbolic decoding, constraint-guided search].

Why this formula differs: Sentence 2 must include a benchmark number — without it, the sentence is not CS writing. Sentence 3 compares models as the unit of analysis, not studies. Sentence 5 combines gap and solution in one sentence because CS is solution-oriented: you name what is missing and immediately propose your technical approach. There is no separate "positioning" sentence as in psychology or nursing.

How Acade Supports This Discipline-Specific Review

Acade is an AI academic research and writing assistant. It can help plan literature searches, organize notes, compare studies, develop an outline, draft synthesis-focused prose, translate academic material, and polish writing. Where citation features are included in the selected plan, it can also help format references.

Acade cannot verify that a gap is real, guarantee citation accuracy, replace disciplinary or clinical judgment, or take responsibility for authorship and submission. Check every source and follow institutional, ethical, and journal requirements. Product features, credits, and plan availability may change.

Tell Acade your computer-science task, method families, datasets, metrics, baselines, and verified papers. Ask it to identify the function of each sentence, show where evidence or synthesis is missing, and propose the next revision without inventing sources or results.

Compare My Research Methods

Related Discipline Examples

• Business Literature Review Example

• Education Literature Review Example

Related Research Workflows

• Best Academic Databases

• How to Find Research Papers

Compare All Annotated Examples

Return to the Annotated Literature Review Example hub to compare sentence-level writing across all eight disciplines, or use the complete dissertation and thesis guide for the broader literature-review workflow.