PaperlayerView exact revision on arXiv

Source-linked AI summary

The Curious Case of Factual (Mis)Alignment between LLMs' Short- and Long-Form Answers

Saad Obaid ul Islam, Anne Lauscher, Goran Glavaš

arXiv:2510.11218v3cs.CLcs.AI

TL;DR

Existing evaluations measure short- and long-form factuality separately, leaving consistency across query complexity untested. SLAQ compares answers to the same facts in isolated and combined queries; across models, factual correctness often differs between formats, and alignment frequently reflects shared incorrectness.

  • Problem

    Existing benchmarks evaluate short- and long-form factuality in isolation, leaving unanswered whether models give consistent answers to the same facts across query complexity.

    Source: §1 · p. 1

  • Method

    SLAQ compares model answers to the same factual questions asked independently and combined into long queries, assessing correctness and short–long alignment.

    Source: §1 · p. 1, §3.1 · p. 4

  • Results

    Models show systematic short–long misalignment; most correctness alignment reflects both answers being incorrect, and short-query accuracy is generally higher.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    Factual consistency across query complexity is an important dimension of LLM reliability that simple-query performance alone does not establish.

    Source: §1 · p. 2, §6 · p. 8

    The dataset and evaluation framework cover English and do not address free-form long-form generation.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Large language models (LLMs) can correctly answer "When was Einstein born?" yet fail to provide the same date when writing about Einstein's life revealing a fundamental inconsistency in how models access factual knowledge across task complexities. While models display impressive accuracy on factual question-answering benchmarks, the reliability gap between simple and complex queries remains poorly understood, eroding their trustworthiness. In this work, we introduce Short-Long Form Alignment for Factual Question Answering (SLAQ), a controlled evaluation framework that compares LLMs' answers to the same factual questions asked (a) in isolation (short) vs. (b) integrated into complex queries (long). Looking at 16 LLMs across 600 queries, we find a systematic misalignment of answers to the corresponding short and long queries. We further uncover position-dependent accuracy loss and momentum effects where consecutive correct or incorrect answers create self-reinforcing patterns. Through mechanistic analysis, we find that aligned facts activate overlapping model internals, and that metrics based on mechanistic similarity can predict short-long answer alignment with up to 78% accuracy. Our work establishes factual consistency over query complexity as an important aspect of LLMs' trustworthiness and challenges current evaluation practices, which implicitly assume that good performance for simple factual queries implies reliability in more complex knowledge-seeking tasks too.

1 Introduction

The introduction identifies a gap in evaluating whether factual answers stay consistent as queries become more complex. It presents SLAQ as a controlled comparison and reports systematic short–long misalignment, with behavioral patterns and internal pathway similarity offering further evidence.

1 Introduction

  • Existing open-domain benchmarks assess short- and long-form factuality separately, leaving consistency across query complexity unmeasured.The introduction asks whether a model gives the same answer to the same factual question when query complexity varies.Source: §1 · p. 1
  • SLAQ compares answers to the same fact-seeking questions asked alone or combined with four other topically related questions.Models answer each short query independently and the combined long query; correctness comparisons distinguish knowledge gaps from retrieval failures.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • Across 16 LLMs, most short–long correctness alignment comes from both answers being incorrect, while short-query accuracy is generally higher.Consequently, many misalignments pair a correct short answer with an incorrect answer to the corresponding question in the long response.Source: §1 · p. 2
  • Consecutive correct answers increase the likelihood of later accuracy, while errors tend to cascade and compound.The paper describes these as momentum effects in factual answering.Source: §1 · p. 2
  • Six pathway-similarity metrics predict short–long answer alignment with 80% accuracy (ROC-AUC: 0.85).Attention-head rank correlation is the most predictive feature; aligned answers also show more similar pathways and stronger correlations in component-importance rankings.Source: §1 · p. 2

2 Background and Related Work

Related work covers factuality evaluation across short and long answers, observed hallucination patterns, and mechanistic interpretability. The paper distinguishes its focus on factual consistency across query formats from prior work evaluating those formats separately.

2 Background and Related Work

  • Open-domain factuality benchmarks include TruthfulQA (Lin et al., 2022) and SimpleQA (Wei et al., 2024a) for single factoid answers, and FactScore (Min et al., 2023) for Wikipedia biographies.LongFact (Wei et al., 2024b) and UNCLE (Yang et al., 2025b) extend long-form evaluation across diverse domains.Source: §2 · p. 3
  • UNCLE (Yang et al., 2025b) also pairs short and long queries, but evaluates them in isolation and studies uncertainty expression rather than factual consistency between formats.Source: §2 · p. 3
  • Prior work describes snowballing of false assertions, input-position sensitivity, and increased hallucinations near the end of generated summaries.The reported position effects concern evidence location in a long context and output location in long-document summarization, respectively.Source: §2 · p. 3
  • Mechanistic interpretability localizes output-relevant components through activation patching and assembles minimal circuits using greedy search.This work uses zero-ablation for computational efficiency; earlier circuit studies include ROME (Meng et al., 2022) and collaborative retrieval circuits (Yao et al., 2024).Source: §2 · p. 3

3 Factual Consistency over Query Complexity

SLAQ evaluates whether factual answers remain consistent when questions are asked alone or embedded in more complex queries. Its task definition distinguishes correctness alignment from answer equivalence, while its dataset combines manually checked questions across diverse topics.

3 Factual Consistency over Query Complexity

  • SLAQ frames factual answer consistency across query complexity as a benchmark task.Its goal is to measure whether responses to the same fact-seeking questions remain factually equivalent across query formats.Source: §3 · p. 3

3.1 Task Definition and Metrics

  • Each topic contains N related facts, with a short question for each fact and a long query that combines the questions.Models respond independently to the long query and each short question; S_k and L_k indicate whether the corresponding short- and long-query answers are correct.Source: §3.1 · pp. 3–4
  • Alignment means matching correctness labels for a fact, not semantic equivalence between the two answers.Thus, two incorrect answers count as aligned even if they differ in content.Source: §3.1 · p. 4
  • FS and FL measure short- and long-form accuracy, while Align± distinguishes aligned-correct, aligned-incorrect, and misaligned cases.The signed scores are +1 for both correct, −1 for both incorrect, and 0 for differing correctness labels.Source: §3.1 · p. 4

3.2 Dataset

  • SLAQ samples 600 popular and obscure English Wikipedia pages across 15 categories, yielding 600 topics and 3,000 short questions.The dataset includes 1,071 entity-based and 1,929 non-entity facts.Source: §3.2 · p. 4
  • OpenAI o3-mini-high generated five short questions and one long query per topic, and researchers manually checked all generated questions.The generator introduced errors in 77 of 3,600 query-answer pairs (2.14%); prompts that failed open-domain answerability or factual-grounding criteria were adjusted.Source: §3.2 · p. 4

4 Benchmarking and Evaluations

The evaluation finds that factual accuracy is generally higher for short than long queries, while high short-long alignment often reflects answers that are wrong in both formats. Consecutive successes and errors also shape accuracy within long-form responses.

4.1 Experimental Setup

  • 92.0% of short-query and 94.8% of long-query judge labels agree with human annotations; evaluation covers five model families from 1B to 12B parameters.Models use greedy decoding, and the order of five sub-questions is randomized across five permutations; the judge assesses semantic equivalence to gold answers.Source: §4.1 · p. 5

4.2 Results and Analysis

  • Most models achieve 30–50% factual accuracy on both query formats, with short-query accuracy consistently higher than long-query accuracy.Larger models perform only modestly better, and the short-form accuracy measure FS exceeds the long-form measure FL.Source: §4.2 · pp. 5–6
  • 73–78% raw alignment coexists with signed alignment of -0.01 to -0.51, indicating that aligned answers more often share incorrectness than correctness.Models show negative signed alignment because they provide wrong answers in both formats more often than correct answers in both.Source: §4.2 · pp. 5–6
  • After four consecutive correct facts, accuracy rises from 30% to 57%; after three errors, it falls from 45% to 24%.Each additional success adds roughly 7% to subsequent accuracy; these momentum patterns are shown by long-form sub-fact position and preceding streak length.Source: §4.2 · p. 6, Figure 3 · p. 6

5 Mechanistic Analysis

The analysis tests whether factual alignment across short and long queries corresponds to shared internal mechanisms, then asks whether those mechanisms predict alignment. Aligned facts show greater mechanistic similarity, and attention-based similarity is a strong alignment predictor.

5 Mechanistic Analysis

  • The analysis is motivated by whether behavioral differences reflect distinct internal mechanisms, and hypothesizes greater similarity for facts answered correctly in both formats than for facts correct in only one.The latter group defines misalignment for this comparison.Source: §5 · p. 6

5.1 Preliminaries: Component Importance via Zero-Ablation

  • Zero-ablation measures each component’s effect on the gold-token logit, then selects high-importance components until they recover at least 90% of the baseline logit.This yields the minimal component set used for each answer token.Source: §5.1 · p. 6
  • The analysis compares component sets with Containment and IoU, and compares component-importance scores using Pearson and Spearman correlations.Containment measures sharing relative to the smaller circuit; IoU measures overlap symmetrically, while the correlations compare score values and component rankings.Source: §5.1 · p. 7
  • Earth Mover’s Distance aggregates token-level component similarities into an answer-level score by optimally aligning tokens across multi-token answers.It accommodates differing token boundaries and token order.Source: §5.1 · p. 7

5.2 Mechanistic Comparison

  • The mechanistic comparison analyzes 240 short–long fact pairs across four Qwen models, split evenly between aligned and misaligned pairs.The study extracts minimal component sets and importance scores, then uses EMD to aggregate token similarities.Source: §5.2 · p. 7
  • Aligned facts show significantly higher mechanistic similarity than misaligned facts across all six metrics (all p < 0.001).IoU is 0.309 vs 0.261 and Containment is 0.552 vs 0.493; attention metrics show the largest relative gaps.Source: §5.2 · pp. 7–8

5.3 Model Internals as Predictors

  • Spearman attention is the strongest individual alignment predictor (ROC-AUC = 0.86, Accuracy = 0.79), while combined features reach Accuracy = 0.80.The combined model has ROC-AUC = 0.85; Spearman Attention has the largest coefficient (1.53).Source: §5.3 · p. 8, Figure 5 · p. 8

6 Conclusion

The conclusion frames factual consistency across query complexity as a dimension of LLM reliability. It identifies systematic misalignment and calls for studying free-form long-form alignment and circuit interventions.

6 Conclusion

  • The paper establishes factual consistency over query complexity as a critical dimension of LLM reliability and reports systematic short–long misalignment.It also reports that models are often aligned in being incorrect and that correctness exhibits momentum effects.Source: §6 · p. 8
  • Future work should test factual alignment between short queries and free-form long-form queries, and investigate targeted circuit-pattern interventions.Source: §6 · p. 9

Limitations

The paper identifies limitations in SLAQ’s data, evaluation, scope, and mechanistic method.

Limitations

  • SLAQ uses synthetically generated data rather than fully human-annotated data, balancing quality and resource costs.Source: Limitations · p. 9
  • The dataset covers only English, uses LLM-as-a-judge factuality assessment, and does not address free-form long-form generation.The authors report high agreement between Gemini-2.5-Flash and human annotations in their experiments.Source: Limitations · p. 9
  • The mechanistic analysis uses zero ablation rather than counterfactual activation patching, which the authors describe as potentially more precise.Constructing counterfactual prompts for complex, multi-token tasks is challenging and labor-intensive.Source: Limitations · p. 9

A.1 Dataset Construction

SLAQ’s dataset is built from Wikipedia articles sampled across categories and popularity levels, then used to generate paired short and long factual queries. The generation process constrains the five short answers and the long answer to the article’s facts.

A.1 Dataset Construction

  • Articles are sampled from 15 curated categories, filtered to at least 1,000 characters, and split by pageviews into more- and less-popular sets.Within each category, pages are sorted by pageviews before splitting.Source: §A.1 · p. 11
  • For each article, o3-mini-high generates five self-contained, single-fact short questions and answers, then a long question and answer covering those same facts.Short answers are under 10 words and derived strictly from the article; the long answer must synthesize only the five short answers.Source: §A.1 · p. 11
  • The generation prompt bans ambiguous “pick one of many” questions and enforces a strict, single-line output format.Constructing the dataset with o3-mini-high cost $18.57.Source: §A.1 · p. 12
  • The dataset’s topic categories range from Aesthetics and Algebra to Theology and Virtual Reality.Source: §A.1 · p. 12

A.2 LLM-as-a-judge

The authors evaluate short- and long-form answers with an LLM judge and progressively develop correctness criteria, then test those criteria on a separate set.

A.2 LLM-as-a-judge

  • The LLM judge reports 92.% accuracy on short-form responses and 94% on long-form responses across 500 atomic facts.The evaluation covers 50 samples, comprising 250 short questions and 50 long questions.Source: §A.2 · p. 12
  • Incorrect answers include factual errors, semantic dissimilarity, irrelevance or omission, contradiction, and hallucination; correctness includes semantic equivalence, subset or superset equivalence, sparse world-knowledge overrides, and correct vagueness.The criteria were developed on 50 validation samples and tested separately on 50 samples.Source: §A.2 · p. 12

A.3 Inference

Responses are generated with greedy decoding on H100 GPUs, and the mechanistic analysis finds greater overlap between the pathways supporting aligned answers. The analysis uses zero ablation and manually pairs short-form facts with corresponding long-form spans.

A.3 Inference

  • Short- and long-form answers are generated through the Hugging Face API with greedy decoding on NVIDIA H100 GPUs, requiring 192 GPU hours.Source: §A.3 · p. 12

A.4 Mechanistic Interpretability

  • The mechanistic analysis applies zero ablation, then manually pairs each short-form fact with its minimal corresponding span in the long-form answer.The manual pairing required 16 hours.Source: §A.4 · p. 12
  • Aligned responses exhibit higher mechanistic overlap under both the Hungarian algorithm and Earth Mover’s Distance.Source: §A.4 · p. 12

Related papers

Loading 2510.11218v3…