llm summaries systematically distort topic distributions
content_hash: b7a83e20aa016f20 ---# LLM Summaries Systematically Distort Topic Distributions
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/shreeya1510/llm-summarisation-bias
Area: NLP Understanding
Keywords: LLM summarization, topic distribution distortion, prior bias, faithfulness evaluation, multi-document summarization
TL;DR¶
A systematic benchmark reveals that large language models systematically and stubbornly distort source topic distributions in multi-document summarization, persistently amplifying topics aligned with their pre-trained importance priors—even when entirely absent from input reviews—and resistant to mitigation via scaling or prompt engineering.
Background & Motivation¶
In an era increasingly powered by generative AI, commercial organizations and consumers routinely depend on large language models to condense large volumes of customer feedback for product development, resource allocation, and purchasing choices. A fundamental premise underpinning this widespread reliance is that model-generated summaries faithfully and proportionally capture the dominant themes and sentiment distribution of the underlying corpus. However, prior investigations into LLM summarization fidelity have overwhelmingly concentrated on fine-grained factual inconsistency or sentiment aggregation failures (such as insensitivity to positive-to-negative review ratios, position biases, and over-representation of majority stances), leaving a critical dimension unaddressed: whether LLMs preserve the macro-level topic distributions present in the source documents.
When topic proportions shift within a body of reviews, a reliable summarizer should adapt its coverage proportionally to reflect customer emphasis accurately. Instead, current abstractive multi-document summarization systems frequently resort to fluent yet over-generalized synthesis, heavily relying on background language patterns and representations acquired during pre-training. Under long-context inputs and multi-instance processing constraints, LLMs struggle with selective attention and coverage balancing. This leads them to fixate on pre-conceived canonical themes while suppressing minority yet operationally critical feedback (such as acute pricing complaints or technical bugs), ultimately generating misleading summaries that distort real-world decision-making.
To rigorously quantify and diagnose this distortion, this work establishes a controlled, multi-domain evaluation framework that explicitly grounds source topic proportions and systematically probes topic preservation across 11 diverse LLMs. Core idea: by constructing a controlled benchmark of 50,000 synthetic reviews across 50 domains and 10 topics alongside real-world validation, this paper demonstrates that LLM summaries suffer from systematic, directional topic distribution distortion driven by pre-trained domain importance priors—amplifying canonical topics even when completely omitted from source reviews, and remaining intractable across frontier model scaling and extensive prompt interventions.
Method¶
Overall Architecture¶
The paper introduces a two-stage evaluation framework designed to eliminate confounds such as ambiguous topic boundaries, multi-topic sentence entanglement, and positional bias. The pipeline operates across three primary stages: first, constructing synthetic customer review datasets (surveys) where every review is strictly single-topic and sentiment-balanced; second, passing randomly shuffled review sets into a candidate LLM to generate multi-document summaries under controlled formatting; and third, employing a structured LLM judge to extract the implied topic distribution from the summary and computing divergence metrics against the ground truth.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Controlled Survey Construction<br/>50 domains × 10 topics × balanced sentiment"] --> B["Multi-Document Summary Generation<br/>shuffled order to eliminate position bias"]
B --> C["LLM Judge Distribution Estimation<br/>extract predicted proportions & normalize"]
C --> D["Prior Correlation & Robustness Validation<br/>probing pre-trained domain prior mechanisms"]
Key Designs¶
1. Controlled Benchmark Construction: Single-Topic Disentanglement and Exact Proportion Injection To assess model responsiveness to topic proportions in the absence of confounding variables, the benchmark eliminates the ambiguity of multi-aspect sentences common in real-world reviews. The authors curated 50 consumer-facing domains (e.g., Restaurants & Cafés, Laptops, Airlines, Online Tutoring), defining 10 mutually exclusive, domain-representative topics for each. For every topic, 100 short reviews (1-2 sentences) were generated with an even split of positive and negative sentiment, totaling 50,000 synthetic reviews. The corpus underwent rigorous deduplication (only 1 near-duplicate pair found with Jaccard > 0.8 across 1.2M within-topic comparisons) and sentiment verification (100% agreement across two independent external LLM classifiers). In the standard primary experiment, surveys of 40 reviews (4 reviews per topic: 2 positive, 2 negative) were sampled, establishing an exact uniform ground truth distribution of 10% per topic.
2. Controlled Multi-Document Summary Generation: Multi-Model Evaluation and Order Randomization Across 11 LLMs—encompassing GPT-5-mini, GPT-5.4-mini, Llama-3-8B, Llama-3.3-70B, OLMo-3.1-32B, Qwen3-14B/80B, alongside frontier models GPT-5.4, Gemini-3.1-Pro, DeepSeek-V3.2, and Claude Sonnet 4.6—each model was prompted to produce a multi-document summary faithfully reflecting the distribution of topics present in the input. To rule out "Lost in the Middle" and primacy/recency position effects in long contexts, reviews were fully shuffled before being provided to the model. Sampling temperature was fixed at 0.0 across all main experiments to ensure deterministic reproducibility.
3. Structured LLM Judge Distribution Estimation and Multi-Metric Divergence To extract implicit topic proportions from unstructured narrative text, an LLM judge was provided with the 10 topic definitions for the target domain and prompted to assign percentage values reflecting topic emphasis within the summary, normalized to 100%. A No-Summarisation (No-Sum) baseline—where the judge estimated topic proportions directly from raw reviews—yielded an SMAPE of 16.3%, demonstrating that the judge itself does not introduce the observed distortion. Evaluation relies on the scale-invariant, symmetric Symmetric Mean Absolute Percentage Error (SMAPE) and Jensen-Shannon Divergence (JSD). For ground truth proportions \(a_i\) and predicted proportions \(p_i\):
4. Prior Belief Correlation and Topic Omission Stress-Testing To investigate the root cause of why specific topics are repeatedly amplified, the authors formulated the "pre-trained prior belief" hypothesis. In a zero-shot isolation test without access to any customer reviews, LLMs were instructed to rank the 10 domain topics by their inherent importance to customer satisfaction. The Spearman rank correlation \(\rho\) between these zero-shot importance rankings and the mean predicted proportions in summaries was measured across all 50 domains. Furthermore, the authors introduced an extreme Topic Omission experiment: the most over-predicted topic from the main experiment was completely removed from the review set (reducing surveys from 40 to 36 reviews with 0% ground truth for that topic) to verify whether models could discern true absence or would hallucinate coverage from prior expectations.
Key Experimental Results¶
Main Results¶
The primary benchmark compares 7 models across 50 domains in the same-model judge configuration, benchmarked against the No-Summarisation baseline.
| Model | Parameter Scale / Access | SMAPE (%) ↓ | JSD ↓ | Degradation vs No-Sum |
|---|---|---|---|---|
| No-Sum Baseline (GPT-5-mini) | - | 16.3 ± 5.9 | 0.009 ± 0.005 | - |
| Qwen3-14B | 14B (NF4, Local) | 34.8 ± 5.2 | 0.022 ± 0.006 | +18.5% |
| Qwen3-80B | 80B (API) | 35.3 ± 8.0 | 0.028 ± 0.012 | +19.0% |
| GPT-5.4-mini | Undisclosed (API) | 39.4 ± 6.5 | 0.030 ± 0.009 | +23.1% |
| GPT-5-mini | Undisclosed (API) | 40.8 ± 4.7 | 0.030 ± 0.007 | +24.5% |
| OLMo-3.1-32B | 32B (API) | 41.9 ± 6.1 | 0.032 ± 0.009 | +25.6% |
| Llama-3.3-70B | 70B (API) | 47.2 ± 6.8 | 0.041 ± 0.011 | +30.9% |
| Llama-3-8B | 8B (API) | 59.2 ± 10.0 | 0.064 ± 0.023 | +42.9% |
Ablation Study¶
On a 10-domain representative subset, the authors conducted extensive stress-tests evaluating input distribution skewness, frontier model scaling, complete topic omission, and prompt engineering strategies.
| Config / Perturbation Dimension | Evaluated Model / Strategy | Key Metric | Behavior & Note |
|---|---|---|---|
| Skewed Distribution (25% down to 2.5%) | GPT-5.4-mini Llama-3.3-70B |
SMAPE: 36.7% → 51.7% SMAPE: 46.0% → 59.0% |
Error worsens under long-tail inputs; models largely ignore input frequency and preserve prior emphasis |
| Frontier Model Scaling | GPT-5.4 Claude Sonnet 4.6 DeepSeek-V3.2 Gemini-3.1-Pro |
SMAPE: 32.1 ± 4.0% SMAPE: 31.8 ± 5.2% SMAPE: 29.6 ± 5.1% SMAPE: 25.3 ± 8.9% |
Significant distortion persists across frontier models (>25%); scaling parameters alone cannot resolve prior bias |
| Topic Omission (0% Ground Truth) | GPT-5.4-mini (GPT) Llama-3.3-70B (Llama) |
Mean assigned share: 7.9 ± 4.2% (GPT) Mean assigned share: 8.8 ± 4.2% (Llama) |
Topics completely absent from input still receive non-zero allocations (e.g. food quality in Restaurants: 14.2% / 17.3%) |
| Summarisation Prompt Variants | Default (Proportional) Simple (3-para) Accuracy (Strict Factual) Chain-of-Thought (CoT) |
SMAPE (GPT-5.4-mini): 38.0 ± 10.1% 34.7 ± 9.9% 33.2 ± 8.6% 45.7 ± 24.6% |
Strict factual prompt yields minor gains; CoT pre-counting causes ironic rebound, spiking error to 45.7% |
| Corpus Size Scaling (Review Count r) | r = 20, 40, 80, 160 | SMAPE: 36.7% / 37.2% / 36.0% / 37.3% (GPT) | SMAPE remains entirely flat as review count octuples, confirming bias is not driven by context capacity limits |
| Real-World Dataset Validation | SemEval & Airline Reviews | GPT-5.4-mini: 42.1 ± 9.4% Llama-3.3-70B: 60.4 ± 10.5% |
Benchmarks on real restaurant, hotel, laptop, and airline reviews confirm the identical prior amplification pattern |
Key Findings¶
- High Cross-Model Consensus on Amplified Topics: Topic distortion is highly directional rather than random information decay. Across 7 models and 10 surveys per domain (70 runs total), the same topic consistently dominated as the most over-predicted: in Furniture Stores, "product quality" was the top-amplified topic in 62 of 70 runs (chance expectation is 7); in Online Tutoring, "tutor quality" dominated 57 of 70 runs; and in Airlines, "staff service" led in 53 of 70 runs.
- Strong Alignment with Internal Pre-trained Priors: Across all 50 domains, zero-shot importance rankings correlated strongly with summary topic allocations (average Spearman \(\rho = 0.68\), with 42 of 50 domains exceeding \(\rho > 0.5\), reaching 0.93 in Home Cleaning and 0.72 in Restaurants). Summaries reflect what pre-training corpora highlight as canonical for each domain.
- Hallucinatory "Phantom Emphasis" on Omitted Topics: Even when the top amplified topic was completely deleted from the source corpus, LLMs opened their summaries with boilerplate framing asserting its dominance (e.g., "many comments focused on food quality and menu breadth"), leading judges to assign 14.2% (GPT) and 17.3% (Llama) to a topic with 0 mentions.
- The Ironic Rebound of Chain-of-Thought: Standard prompt interventions (accuracy instructions, varying summary length from 1 to 5 paragraphs, varying temperature) failed to eliminate the bias. Crucially, introducing Chain-of-Thought prompting to explicitly calculate topic frequencies before summarization backfired, triggering an ironic rebound effect where SMAPE degraded severely to 45.7% (GPT) and 58.0% (Llama).
Highlights & Insights¶
- Novel Problem Formulation: First large-scale investigation identifying macro-level "topic distribution distortion" as a distinct, severe dimension of unfaithfulness in LLM multi-document summarization, transcending sentence-level hallucination.
- Airtight Experimental Disentanglement: The decoupling of single-topic synthetic reviews, 50 domains, No-Sum baseline validation, and topic omission tests establishes an unassailable causal link between pre-trained priors and summary distortion.
- Crucial Caution for Downstream Analytics: Strongly demonstrates that using LLMs as black-box aggregators for sentiment analysis, customer satisfaction tracking, or public policy consultation risks manufacturing consensus around stereotypes while ignoring true user feedback shifts.
Limitations & Future Work¶
- Domain and Language Scope: All experiments are conducted on English customer feedback, leaving open whether similar topic distortions occur in non-English reviews or other genres like scientific literature and legal documents.
- Single-Topic Review Assumption: While crucial for controlled evaluation, real-world customer reviews often entangle multiple topics within single utterances, which may interact with prior distortion in more intricate ways.
- Absence of Architecture-Level Solutions: The paper conclusively demonstrates that prompt engineering and parameter scaling fail, but does not propose a decoding-time or fine-tuning intervention to enforce strict topic proportionality.
Related Work & Insights¶
- vs Factual Consistency and Hallucination Benchmarks (Peters & Chin-Yee, 2025; Maynez et al., 2020): While conventional faithfulness metrics focus on localized entity or relational hallucinations, this work uncovers systematic distributional hallucination where the overall narrative framework distorts reality despite local sentence plausibility.
- vs Long-Context Processing and Position Biases (Liu et al., 2023; Chen et al., 2026): Unlike position effects (e.g., Lost in the Middle) that vary with document ordering or corpus length, topic distortion remains invariant to review shuffling and review count variations (20 to 160).
- vs Opinion and Sentiment Aggregation (DeYoung et al., 2024; Aghaebe et al., 2026): Extends prior findings on sentiment ratio insensitivity by isolating the orthogonal, structural dimension of topic distribution preservation under balanced sentiment controls.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Pioneering evaluation paradigm focusing on macro topic proportion fidelity in multi-document summarization.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive experiments across 50 domains, 11 models, 50,000 synthetic reviews, real-world ABSA datasets, and extensive robustness stress-tests.
- Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally structured, logically airtight, and effective in methodically disproving competing explanations.
- Value: ⭐⭐⭐⭐⭐ Provides an urgent, high-impact warning for industrial practitioners and researchers relying on LLM-based customer intelligence and survey synthesis.