A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models

In plain words

Large language models often write text that sounds right but is not backed by any source. This survey gathers more than thirty techniques that try to reduce that problem and sorts them into groups. One group changes how the model is prompted: fetch documents while it answers, have it check and rewrite its own output, or tune the prompt itself. The other group changes the model: alter how it picks each next word, feed it structured knowledge, add a faithfulness term to training, or fine-tune it on data meant to make it more factual.

The paper does not propose a new method or run its own experiments. It is a map: a taxonomy, a table listing each technique with its tasks and the models it was tested on, and a discussion of open limits. Read it to choose a family of approaches before you pick a specific technique to implement.

TL;DR

A survey, not a method. It covers 32 techniques for reducing hallucination in LLMs, in two families: prompt engineering (retrieval at three points in generation, self-refinement, prompt tuning) and model development (decoding, knowledge graphs, faithfulness losses, supervised fine-tuning) fig. 1 tab. 1. The paper runs no experiments, and its only quoted result is a 59.6% F1 for one technique, which is author-reported §2.1.1. Go to the primary paper for any technique before implementing it.

Problem

Generative LLMs produce fluent text that can contain claims with no support in their training data or in the input, and this is the main barrier to deploying them where errors matter, such as medical summaries or financial reports [abstract] §1. The paper attributes this to training on large, uneven web text, to misreading ambiguous prompts, and to the lack of live knowledge after training §1. It is a survey, so it defines no benchmark or evaluation protocol of its own.

Method

The paper's contribution is a taxonomy and a summary table, not an algorithm. The paper contains no algorithm block and no equations. The steps below restate how each surveyed technique works, as described in the paper. The one loop the text describes in enough detail to write down is LLM-Augmenter's. It is reconstructed below as pseudocode from that description, not copied from the paper.

# LLM-Augmenter loop, reconstructed from the survey's description  [§2.1.1]
# Input:  query (str); evidence_source; fixed_llm (black-box LLM)
# Output: response (str) that passed verification

def answer_with_verification(query, evidence_source, fixed_llm):
    # Retrieve evidence and reason over it into evidence chains  [§2.1.1]
    evidence_chain = retrieve_and_reason(query, evidence_source)
    prompt = build_prompt(query, evidence_chain)
    while True:
        # Ask the fixed LLM for a candidate grounded in the evidence  [§2.1.1]
        candidate = fixed_llm(prompt)
        # Check the candidate for invented evidence  [§2.1.1]
        feedback = verify(candidate, evidence_chain)
        if feedback is None:
            return candidate
        # Revise the prompt with the feedback and query again  [§2.1.1]
        prompt = revise_prompt(prompt, feedback)

Taxonomy. Techniques fall under two headings. Prompt engineering has three sub-groups: retrieval augmentation, self-refinement through feedback and reasoning, and prompt tuning §2 §2.1 fig. 1. Developing models has four: new decoding strategies, use of knowledge graphs, faithfulness-based loss functions, and supervised fine-tuning §3 fig. 1.

Retrieval augmentation (RAG), by when retrieval happens §2.1.1.

1. Before generation. The model receives retrieved evidence in its prompt. LLM-Augmenter wraps a fixed black-box LLM with plug-in modules: it retrieves evidence, queries the LLM, checks the answer for invented evidence, and rewrites the prompt with feedback until the answer passes §2.1.1. FreshPrompt adds search-engine results to a few-shot prompt, and the paper notes that the number and order of retrieved items affect correctness §2.1.1.
2. During generation. Retrieval runs sentence by sentence. Knowledge Retrieval uses token logits to flag likely hallucinations, checks them against retrieved text, and asks the model to repair the flagged phrase §2.1.1. The Decompose-and-Query framework (D&Q) breaks a question into steps, queries an external question-answer base through tools, and can backtrack and search again §2.1.1. EVER runs a generate, validate, rectify loop at each step §2.1.1.
3. After generation. The full answer is checked and then edited. RARR runs research and then revision to attach evidence while keeping the original text §2.1.1. High-entropy word spotting uses an open model to find uncertain words, groups consecutive ones, masks them, and replaces them with a model that is less prone to hallucinate §2.1.1.
4. End-to-end. RAG pairs a pretrained seq2seq generator (BART) with a dense retriever (DPR) over a Wikipedia index. Both are trained jointly, and the latent documents are marginalised either per output or per token §2.1.1.

Self-refinement through feedback and reasoning §2.1.2. Each entry prompts or trains the model to critique and revise its own output. The techniques listed are prompting GPT-3 for reliability across four facets, ChatProtect for self-contradictions, an iterative self-reflection loop for medical QA, structured comparative reasoning for text preference, Mind's Mirror for distilling self-evaluation into small models, DRESS for natural-language feedback on vision-language models, MixAlign for clarifying the link between a question and retrieved references, Chain-of-Verification (CoVe) for drafting, planning checks, answering them independently and rewriting, and CoNLI, which judges sentences and entities with a chain of natural language inference steps and then post-edits the text §2.1.2.

Prompt tuning §2.1.3. A learned retriever selects prompts for a new task (UPRISE), and a synthetic task is used to tune a system message through prefix-tuning before transfer to summarisation (SynTra) §2.1.3.

Decoding strategies §3.1. CAD contrasts output distributions with and without the context, which helps when the context conflicts with the model's prior knowledge §3.1. DoLa contrasts logits projected from later and earlier layers, based on the paper's observation that factual knowledge is localised in certain layers §3.1. ITI selects a sparse set of attention heads with high probing accuracy for truthfulness and shifts activations along truth-correlated directions at every generated token §3.1.

Knowledge graphs §3.2. RHO grounds dialogue responses in linked entities and relation predicates from a graph and re-ranks candidates with a reasoning model §3.2. FLEEK extracts facts, writes questions, retrieves evidence from curated graphs and the open web, and proposes revisions, aimed at human graders §3.2.

Faithfulness-based losses §3.3. THAM adds a regulariser derived from mutual information between the response model and a hallucination model, to reduce indiscriminate copying in video-grounded dialogue §3.3. The mFACT loss weighting scores each training summary for faithfulness and weights its loss by that score, for cross-lingual summarisation §3.3.

Supervised fine-tuning §3.4. HALO estimates hallucination severity with sentence-level entailment and reduces it by knowledge injection and teacher-student data from a stronger model §3.4. HAR builds counterfactual training data from the model's own hallucinations §3.4. One entry fine-tunes Llama-2 with automatic fact-checking and DPO preferences §3.4. BEINFO fine-tunes on dialogues with the true source plus randomly sampled facts §3.4. R-Tuning finds questions the model is uncertain about and appends refusal phrases to them §3.4. TWEAK re-ranks decoding candidates by how well their hypotheses support the input facts, with a Hypothesis Verification Model, and needs no retraining §3.4.

Summary table. Table 1 lists the 32 techniques in the paper's order, with the family, tasks, evaluated models, and whether the work covers detection. Cell values are reproduced from the paper's table; the paper's text gives no further details, so consult the cited work.

Source for every row: tab. 1. Family labels follow the paper's fig. 1 and §2 to §3.

Hyperparameters

Not applicable. The survey defines no training or inference settings of its own §2 §3. Settings for each surveyed technique are in the cited works and are not restated here.

Results

Author-reported. The paper runs no experiments of its own. Its only quoted numeric result in the body is the D&Q framework's 59.6% F1 on HotPotQA (question-only), along with a claim that D&Q is competitive with GPT-3.5 on ChitChatQA §2.1.1. The survey's table tab. 1 lists tasks, models and metrics per technique but gives no result values.

Paper vs code

No official code.

The arXiv source has no code, repository or project-page links. The paper does not name any repository in its abstract or body, and the arXiv comments field was not part of the fetched metadata. A GitHub search for the title's key terms returned only third-party curated lists (awesome-style collections of hallucination papers). None belongs to the paper's authors, so none is used.

Pitfalls

• Not an implementable method. Nothing here is enough to implement any of the 32 techniques. Take each one from its primary paper, and use the card to choose which family fits.
• Contradictory framing in §3. The §3 introduction says the techniques avoid fine-tuning, but §3.4 is entirely about supervised fine-tuning §3 §3.4.
• Taxonomy does not follow mechanism. ITI edits activations rather than choosing tokens, yet it is filed under decoding §3.1. CAD is also filed under decoding, and TWEAK, a decoding-time re-ranker, sits under supervised fine-tuning §3.1 §3.4. CoNLI is described as detection followed by post-editing, yet it is filed under self-refinement §2.1.2. FLEEK targets human graders rather than a model §3.2. Use the table's families as navigation, not as a mechanism description.
• "Detection" is undefined. Table 1 marks each work with a tick or cross for detection, but the paper never defines what counts as detection tab. 1.
• No stopping rule. The text gives LLM-Augmenter's loop without an iteration cap, a fallback when verification never passes, or a cost bound §2.1.1. An implementation needs to add these.
• Metrics do not compare. Each row uses different metrics and datasets, and many cells read "Details not provided" or "Manual". Do not rank the techniques by the table tab. 1.
• Limits are listed by the authors but unevenly. Several rows note that a work covers only English, one model family, or a small sample tab. 1. The table limits column is the most specific source of these caveats.
• Source version. This card follows the v3 source in the fetched LaTeX package, which the arXiv metadata dates to 2024-01-08.
• Section numbers are counted from the LaTeX order (§1 Introduction, §2 Hallucination Mitigation, §3 Developing Models, §4 Conclusion, §5 Discussion and Limitations). The paper's compiled layout should match, but the section numbers here were derived from the source and not checked against a PDF.

Lineage

• Builds on: not stated. The survey reviews the works in tab. 1.
• Compared against: not stated. The survey summarises each work and does not benchmark them against one another §2 §5.
• Superseded by: not stated.