Hallucinate at the Last in Long Response Generation: A Case Study on Long Document Summarization

1 Introduction

• For long summaries, faithfulness significantly decreases toward the end, where hallucinations tend to concentrate; the authors call this positional pattern “hallucinate at the last.”Figure 1 compares faithfulness scores for summaries of around 50, 200, and 800 words generated with GPT-4o mini.Source: Figure 1 · p. 1, §1 · p. 2
• The study targets a gap in research on the positional distribution of factual errors in long responses generated from long contexts.Prior work had examined positional biases in processing input contexts, while output-position errors—especially for long responses from long contexts—were less studied.Source: §1 · pp. 1–2
• The investigation focuses on long-context-aware long-response generation, using long document summarization as its case study.The authors contrast this setting with prior work focused primarily on long inputs or short outputs.Source: §1 · p. 2
• The paper empirically characterizes the late-stage hallucination pattern, investigates potential contributing factors, and explores mitigation strategies tailored to positional errors.Its stated research questions address where hallucinations occur, what contributes to them, and how they might be mitigated.Source: §1 · p. 2

2 Generating Summaries

• Summaries were generated from original documents, and outputs with excessive length, repetitions, or repeated paragraphs were excluded to support fair comparisons.The authors note that controlling summary length became harder with longer input contexts and larger target ranges.Source: §2 · p. 2

2.1 Input Context Lengths

• The evaluation used input contexts over 4K tokens, extending to 12K and 16K tokens for Llama.Source: §2.1 · p. 2

2.2 Output Length Categories

• The study contrasts standard summaries of 100–200 words with long summaries up to 30% of the input context length.The two output categories support comparison of positional patterns across standard and substantially longer summaries.Source: §2.2 · p. 2

2.3 Overall Faithfulness

• FineSurE was used for a preliminary overall-faithfulness check of Llama3.1-8B-Instruct summaries on Wikipedia, with sentence-level source comparisons.The passage reports alignment with CNNDM results, where Llama3-70B-Instruct achieved 85.5% faithfulness; Table 1 is scoped to overall faithfulness on Wikipedia.Source: §2.3 · pp. 2–3, Table 1 · p. 3

3.1 Experimental Setup

• The setup tests four LLMs across four datasets, sampling 20 examples per dataset; all summaries use greedy decoding.The models include Llama3.1-8B-Instruct, Gemma3-12B-Instruct, Qwen2.5-7B-Instruct, and GPT-4o mini.Source: §3.1 · p. 3

3.2 Evaluation Metric

• Faithfulness scores average filtered atomic-fact scores after source-document checking; sensitivity captures positional differences across five equal-length bins.Scores range from 0 to 1, and positive sensitivity indicates lower faithfulness in the final bin than earlier bins.Source: §3.2 · p. 3

3.3 Hallucinate at the Last

• In long summaries, faithfulness declines toward the end for three models, eventually below 0.75, while Qwen dips mid-summary then recovers.Figure 2 compares faithfulness scores across models on Wikipedia.Source: §3.3 · p. 4, Figure 2 · p. 4
• Human evaluation is consistent with Llama having the highest sensitivity; annotators agreed on 515 of 543 facts (94.8% raw agreement).Table 4 reports human-evaluated faithfulness and sensitivity for Llama long summaries on Wikipedia.Source: §3.3 · pp. 4–5, Table 4 · p. 5
• With Llama, sensitivity remains high at 12K and 16K context lengths, so positional discrepancies are not significantly affected by longer input contexts.Table 3 covers standard and long summaries at both context lengths.Source: §3.3 · p. 4, Table 3 · p. 5
• Across datasets, later summary segments are least faithful, and the decline becomes more severe in longer summaries.Figure 3 compares Llama3.1-8B-Instruct summaries across domains.Source: Figure 3 · p. 4, §3.3 · p. 5
• For Llama and Gemma, faithfulness is lowest near the end across decoding strategies; η-sampling takes Llama from above 0.8 early to below 0.7 in the final bin.Qwen generally does not show this end decline; its lowest faithfulness is in the middle.Source: §3.3 · p. 5, Figure 4 · p. 5

4 Why Do Hallucinations Frequently Occur at the Last?

• The section considers summary structure and increasing attention to previously generated tokens, rather than the original context, as possible contributors to end-position hallucinations.The first hypothesis is that key information appears early in summaries, leaving less important content for later; the second proposes an attention shift as outputs grow.Source: §4 · p. 6, §4.1 · p. 6, §4.2 · p. 6

4.1 Is it Intrinsic to Summarization?

• Generated CNNDM summaries fall below 0.65 faithfulness in the final segment, while human references dip mid-summary and recover.This comparison suggests the end-position pattern is not solely due to summarization’s inherent structure.Source: §4.1 · p. 6

4.2 Is it Intrinsic to Attention?

• Llama assigns nearly three times more attention to its final summary sentence than to the first or middle, while Qwen’s attention is similar across all three positions.The authors report that greater attention to previously generated text correlates with a higher likelihood of hallucination.Source: §4.2 · p. 6

Experimental Setup

• The evaluation compares four mitigation methods against Llama3.1-8B-Instruct on Wikipedia summaries: chunk-and-merge summarization, sparse attention, long-output fine-tuning with DPO, and context-aware decoding.The methods are BOOOOKSCORE, MINFERENCE, LONGWRITER-LLAMA3.1-8B, and ADACAD.Source: §5 · p. 7

Results & Analysis

• BOOOOKSCORE has the lowest sensitivity (0.0) and highest faithfulness, especially in the final bin, avoiding the sharp end-of-summary decline seen with other methods.The authors report that it also improves overall summary quality; Figure 7 shows faithfulness trends across increasing output lengths, and Table 6 reports faithfulness and sensitivity.Source: §5 · p. 8, Figure 7 · p. 8, Table 6 · p. 8
• BOOOOKSCORE is a practical mitigation for models without sliding-window attention, but does not fundamentally resolve the issue because it relies on chunking and merging.The authors call for methods that maintain faithfulness during direct generation from long contexts.Source: §5 · p. 8

• Most prior hallucination work focuses on detection or reduction rather than the distribution of errors within long generated sequences.Related studies of long-output generation often use tasks not grounded in contextual input, limiting their applicability to real-world summarization.Source: §6 · p. 8
• Studies of large vision-language models report that hallucinations tend to increase toward the end of generated image descriptions.Source: §6 · p. 8
• The study presents the first dedicated investigation of where hallucinations occur in LLM-based long document summarization, including positional patterns, contributing factors, and mitigation.Source: §6 · p. 8

• Hallucinations disproportionately increase toward the end of long outputs, and the bias is amplified in longer summaries.Source: §7 · p. 9
• The study investigated contributing factors and explored targeted mitigation strategies for the Hallucinate at the Last phenomenon.Source: §7 · p. 9
• The findings highlight output faithfulness and motivate future research into positionally aware generation techniques.Source: §7 · p. 9

Limitations

• Books, dialogues, movie scripts, and meeting transcripts were not investigated.Books were excluded because they were excessively long; the other domains lacked datasets with suitable input lengths or sufficient sample sizes.Source: Limitations · p. 9
• The four studied domains were selected because their datasets provided input contexts of the desired length.Source: Limitations · p. 9
• The authors present the work as a foundation for future research on long-form output generation, particularly summarization.Source: Limitations · p. 9

Ethics Statement

• The study used publicly available datasets and reports that experiments were conducted consistently without data or result manipulation or omission.Source: Ethics Statement · p. 9

A Attention Weight Calculation

• Attention measurements average the attention from each 100-token block to the summary’s first, middle, and last sentences.The full sequence comprises the input context and generated summary, and the target sentences have lengths T1, T2, and T3.Source: A Attention Weight Calculation · p. 13

B Details for Atomic Fact Generation

• The authors manually labeled 1,569 atomic facts as true or false, agreeing on 1,541 for 98.2% raw agreement.They report raw agreement because chance-corrected coefficients can be unstable or difficult to interpret for binary labels with skewed class distributions.Source: B Details for Atomic Fact Generation · p. 13

C Details for Model Usage

• The study used Llama3.1-8B-Instruct, Gemma3-12B-Instruct, Qwen2.5-7B-Instruct, and GPT-4o mini; summaries used greedy decoding with float16 precision.Atomic-fact evaluation used Llama3.1-8B-Instruct with vLLM; the appendix also reports results with five-bin and ten-bin settings.Source: C Details for Model Usage · pp. 13–14
• The Hallucinate at the Last trend persists when results are calculated with ten bins instead of five.Source: C Details for Model Usage · p. 14

D More Results with Different Methods

• The appendix compares faithfulness and sensitivity using NLI, ROUGE-L, and LLM-based FineSurE in addition to MiniCheck.It reports results for standard and long summaries across output bins on Wikipedia; the supplied passages do not state ROUGE-L outcomes.Source: D More Results with Different Methods · p. 14

NLI Based

• NLI results show high sensitivity for Llama and GPT, while Qwen occasionally has sensitivity values below zero.Source: D More Results with Different Methods · p. 14

LLM Based

• Llama and GPT show their highest FineSurE error rates in the final bin for standard and long summaries, while Qwen does not.Source: D More Results with Different Methods · p. 14

E More Results on Varying Datasets

• For standard summaries, below-zero sensitivities occur frequently and affect more than half of PubMed values; for long summaries, they are rare and absent in PubMed.Source: E More Results on Varying Datasets · p. 14
• NLI-based sensitivity for standard summaries across domains includes values below zero.Source: E More Results on Varying Datasets · pp. 14–15

F Details for Mitigation Methods

• BOOOOKSCORE summarizes 2048-token chunks with Llama3.1-8B-Instruct and hierarchically merges the partial summaries; LONGWRITER uses its Llama3.1-8B model with greedy decoding and bfloat16.Source: F Details for Mitigation Methods · p. 15

G Details for Generating Summaries

• Because longer contexts and larger target ranges made output length difficult to control, the analysis retained comparable-length summaries, including 28 sentences for 4K contexts and 40 for 8K.Source: G Details for Generating Summaries · p. 15

H Use of Large Language Model

• ChatGPT-5.1 was used for grammar correction and clarity refinement, not content generation or research ideation.Source: H Use of Large Language Model · p. 16