DxHF: Claim-Level Preference Interface

• DxHF is a novel interface technique that decomposes long responses into atomic claims for effective RLHF preference annotation.
• The system integrates claim decomposition, relevance ranking, and cross-response linking to reduce cognitive load during human evaluations.
• Empirical studies show DxHF improves annotation accuracy by up to 6.4 percentage points despite modest increases in feedback time.

Searching arXiv for the DxHF paper and closely related preference-alignment interface work to ground the article in current literature. DxHF is a user interface and interaction technique for collecting higher-quality human preference data for LLM alignment by decomposing long responses into atomic claims and supporting comparison at the claim level rather than only at the paragraph level (Shi et al., 24 Jul 2025). It was introduced as an HCI contribution to reinforcement learning from human feedback (RLHF)-style data collection, with the stated goal of improving the quality of human judgments that later feed alignment methods rather than changing the learning algorithm itself. The term is potentially ambiguous: in several heart-failure papers, “DxHF” is also used as shorthand for diagnosis-oriented heart-failure research, including decompensation detection, ECG-based congestive heart failure classification, cardiomyocyte dysfunction classification, risk prediction, and diagnosis-trajectory analysis (Cakmak et al., 2021, Kutlu et al., 2017, Mehdi et al., 2 Jul 2025, Zvuloni et al., 20 Dec 2025, González et al., 2024, Murris et al., 2024). In the specific sense established by the 2025 paper, however, DxHF denotes the interactive decomposition interface for long-form preference comparison (Shi et al., 24 Jul 2025).

1. Definition and problem setting

DxHF was proposed for the standard preference-annotation setup in which, given a query $Q$, a model produces two candidate responses $A$ and $B$, and an annotator supplies a preference $\text{Pref}(A,B \mid Q)$ based on which response is more helpful (Shi et al., 24 Jul 2025). The paper formalizes the undecomposed responses as

$$A = \{t^a_1, t^a_2, ..., t^a_n\}, \qquad B = \{t^b_1, t^b_2, ..., t^b_m\},$$

where each $t$ is a sentence in a long-form response. The practical difficulty, as framed by the paper, is that annotators must read two full texts, remember their contents, identify corresponding parts, compare them mentally, and then convert that comparison into a single overall preference.

The system targets cases in which the responses are long, structurally complex, or unfamiliar. The paper explicitly links these conditions to lowered feedback quality: annotators may not fully read both responses, long texts strain memory and attention, multiple intertwined points are hard to isolate, and back-and-forth scanning is cognitively expensive. DxHF therefore treats the annotation interface itself as a source of noise in alignment pipelines. This suggests a reframing of alignment data collection: preference quality is not only a property of the annotator or the model outputs, but also of the representational structure imposed by the interface.

The contribution is situated relative to RLHF and DPO-style pipelines rather than as an alternative to them. DxHF does not introduce a new reward model, new post-training objective, or new preference-learning loss. Its intervention point is upstream: it seeks to improve the human judgments that later become alignment supervision (Shi et al., 24 Jul 2025).

2. Decomposition principle and claim-level comparison

The conceptual core of DxHF is the decomposition principle. Instead of forcing direct comparison of two long-form responses, the system converts each response from a sequence of sentences into a sequence of atomic claims:

$$A = \{c^a_1, c^a_2, ..., c^a_N\}, \qquad B = \{c^b_1, c^b_2, ..., c^b_M\}.$$

Each $c$ is intended to express one piece of information in a simple format (Shi et al., 24 Jul 2025). The annotator can then compare local units such as

$\text{p}(c^a_i, c^b_j \mid Q)$

for corresponding claims, or

$\text{p}(c^a_i, * \mid Q)$

when one side contains a claim with no counterpart on the other.

The paper identifies three resulting workflow changes. Reading becomes selective and simpler because claim lists are easier to skim than paragraphs. Difference finding is externalized because similar or corresponding content is made visible rather than reconstructed mentally. Comparison becomes claim-wise rather than paragraph-wise because each unit carries one idea rather than a mixture of points.

DxHF also emphasizes an “integrity principle”: decomposition should preserve the original meaning as much as possible. Claims are kept near the original text and maintain narrative order unless the user explicitly switches to ranked mode (Shi et al., 24 Jul 2025). This is important because the interface is not intended to replace the source response with an abstract representation; it is intended to expose internal structure while preserving access to the original wording.

The paper’s example illustrates the intended granularity. The sentence

“In addition to his acting roles, he has written and directed two short films and is currently in development on his feature debut.”

is decomposed into four claims: “He has acting roles,” “He has written two short films,” “He has directed two short films,” and “He is currently in development on his feature debut” (Shi et al., 24 Jul 2025). The purpose is not summarization in the ordinary sense, but isolation of individual propositions that can be compared across responses.

3. Interface organization and interaction design

DxHF is organized around three regions: a left vertical list of claims for response $A$0, a right vertical list of claims for response $A$1, and a middle region containing links and keyword labels connecting related claims (Shi et al., 24 Jul 2025). The paper states that the interface was designed for low learning cost in crowdsourcing settings, so the visual structure remains close to ordinary side-by-side comparison rather than introducing a heavily specialized analytic environment.

The first major interface element is the decomposed claim list. Each response appears as a sequence of individual claims, each placed in its own visual unit. The decomposed view can be shown alongside the original responses, reflecting a Focus+Context rationale: users can inspect fine-grained decomposition without losing the source text. The decomposition is also optional. DxHF uses an accordion-fold metaphor in which annotators can remain in the conventional side-by-side layout or click a “decompose” button to unfold the claim-based comparison view.

The second major element is contextual relevance encoding. DxHF computes a relevance score for each claim and renders it using text opacity: more relevant claims appear darker, less relevant claims lighter gray but still legible (Shi et al., 24 Jul 2025). The paper reports that opacity was chosen over background-color highlighting because it was less distracting and reduced clutter. Claims can be shown either in original narrative order or in ranked order, and a sorting toggle $A$2 allows reordering by relevance.

The third major element is cross-response linking. Semantically similar claims across the two responses are connected visually, and each connection is labeled with a keyword intended to summarize the shared topic or meaning. Links with the same keyword can be grouped and bundled using a control labeled $A$3. In grouped mode, the middle keyword list can display counts for repeated topics. The paper states that direct-link alternatives were explored but produced more clutter than bundled keyword-centered linking (Shi et al., 24 Jul 2025).

Interaction is primarily hover-based. Hovering a keyword highlights the connected claims on both sides and the corresponding segments in the original texts. Hovering a claim highlights all links associated with that claim and the corresponding linked claims on the opposite side. The interaction model is therefore designed to support rapid movement between overview and detail rather than a strictly linear reading process.

4. Computational pipeline and formalization

The technical pipeline has three main stages: decomposition, ranking, and linking (Shi et al., 24 Jul 2025).

First, each response is segmented into sentences using NLTK sentence tokenization. Second, each sentence $A$4 is passed to GPT-4 for decomposition into atomic claims,

$A$5

The appendix prompt instructs GPT-4 to break the sentence into independent claims, extract all claims, use words from the original text, prefer a subject–predicate–object form, replace first occurrences of pronouns with antecedents, avoid adding new words, and preserve meaning (Shi et al., 24 Jul 2025).

Third, DxHF computes contextual relevance for each claim using a Cross-Encoder:

$A$6

with $A$7. The score drives opacity encoding, optional sorting, and thresholding in the simulation study. The paper interprets $A$8 as how much a claim contributes to answering the conversation.

Fourth, the system computes semantic similarity across claims from $A$9 and $B$0 using cosine similarity on textual embeddings. The main text states that BERT is used for textual embeddings, and the simulation setup uses a similarity threshold of $B$1 for including links in the set $B$2 (Shi et al., 24 Jul 2025). For each linked claim pair, GPT-4 generates a short keyword summarizing the shared meaning of the two claims, conditioned on the conversation.

The paper also formalizes a simulated-annotator framework used for technical evaluation. Under plain decomposition, the score of a response is the sum of the helpfulness scores of its claims:

$B$3

For decomposition plus ranking, only claims above a relevance threshold are included:

$B$4

with $B$5 in the experiments (Shi et al., 24 Jul 2025). For decomposition plus ranking plus linking, the score additionally includes linked claims in $B$6. The paper notes minor indexing inconsistencies in these equations, but the intended meaning is that relevant and linked claims contribute to the overall response score.

The final simulated preference is sampled using a Boltzmann-rational model:

$B$7

Here $B$8 is a rationality parameter: lower $B$9 yields noisier decisions, while higher $\text{Pref}(A,B \mid Q)$0 yields more consistent choice of the better response (Shi et al., 24 Jul 2025). This formalization is central because it operationalizes the paper’s hypothesis that decomposition is especially useful when annotators are uncertain.

5. Experimental evidence

The paper reports three evaluations: a technical simulation study, a crowdsourcing user study, and an ablation study (Shi et al., 24 Jul 2025). The simulation uses 50 filtered instances from HH-RLHF. It compares four strategies: baseline direct comparison, decomposition, decomposition plus ranking, and decomposition plus ranking plus linking. The reported qualitative result is that decomposition-based strategies generally outperform baseline across rationality levels, with decomposition plus ranking plus linking performing best. The paper does not provide exact per-$\text{Pref}(A,B \mid Q)$1 values in the text, only the trend and summary.

The crowdsourcing study is the main human-subject evaluation. It included 160 Prolific participants, of whom 80 were female; average age was $\text{Pref}(A,B \mid Q)$2 with $\text{Pref}(A,B \mid Q)$3; all were native English speakers; 155 had prior ChatGPT experience; and average ChatGPT-like experience was $\text{Pref}(A,B \mid Q)$4 years with $\text{Pref}(A,B \mid Q)$5 (Shi et al., 24 Jul 2025). Participants used both a baseline side-by-side interface and DxHF in a within-subjects design, with order counterbalanced. Each session included a tutorial, one demonstration example, and five comparison tasks. For each task, participants chose the more helpful response and rated confidence on a 1–5 Likert scale.

The paper defines feedback accuracy against HH-RLHF labels, specifically the most popular votes for machine-generated responses. Baseline accuracy was $\text{Pref}(A,B \mid Q)$6, whereas DxHF accuracy was $\text{Pref}(A,B \mid Q)$7, an absolute improvement of 4.7 percentage points. The reported statistical test was a Wilcoxon signed-rank test with

$\text{Pref}(A,B \mid Q)$8

Among participants expressing less than maximum certainty, excluding those with confidence 5, the subset size was 138 and DxHF was 6.4 percentage points more accurate than baseline, with

$\text{Pref}(A,B \mid Q)$9

DxHF increased average feedback time by 18 seconds, and 95% of feedback sessions were completed in under 200 seconds (Shi et al., 24 Jul 2025).

The ablation study added 36 participants and compared full DxHF, DxHF without linking, and DxHF without ranking. Usefulness ratings were $$A = \{t^a_1, t^a_2, ..., t^a_n\}, \qquad B = \{t^b_1, t^b_2, ..., t^b_m\},$$0 for full DxHF, $$A = \{t^a_1, t^a_2, ..., t^a_n\}, \qquad B = \{t^b_1, t^b_2, ..., t^b_m\},$$1 without linking, and $$A = \{t^a_1, t^a_2, ..., t^a_n\}, \qquad B = \{t^b_1, t^b_2, ..., t^b_m\},$$2 without ranking, with repeated-measures ANOVA

$$A = \{t^a_1, t^a_2, ..., t^a_n\}, \qquad B = \{t^b_1, t^b_2, ..., t^b_m\},$$3

Accuracy was $$A = \{t^a_1, t^a_2, ..., t^a_n\}, \qquad B = \{t^b_1, t^b_2, ..., t^b_m\},$$4 for full DxHF, $$A = \{t^a_1, t^a_2, ..., t^a_n\}, \qquad B = \{t^b_1, t^b_2, ..., t^b_m\},$$5 for DxHF without linking, and the paper reports $$A = \{t^a_1, t^a_2, ..., t^a_n\}, \qquad B = \{t^b_1, t^b_2, ..., t^b_m\},$$6 for DxHF without ranking (Shi et al., 24 Jul 2025). This suggests that both ranking and linking contribute, with the full interface performing best.

Evaluation	Main setup	Reported outcome
Simulation	4 strategies on 50 HH-RLHF instances	Decomposition + ranking + linking was best across rationality levels
Crowdsourcing study	160 participants, within-subjects	Baseline $$A = \{t^a_1, t^a_2, ..., t^a_n\}, \qquad B = \{t^b_1, t^b_2, ..., t^b_m\},$$7, DxHF $$A = \{t^a_1, t^a_2, ..., t^a_n\}, \qquad B = \{t^b_1, t^b_2, ..., t^b_m\},$$8, +18 seconds average time
Less-certain subset	138 participants	DxHF +6.4 percentage points over baseline
Ablation study	36 participants	Full DxHF $$A = \{t^a_1, t^a_2, ..., t^a_n\}, \qquad B = \{t^b_1, t^b_2, ..., t^b_m\},$$9, w/o linking $t$0, w/o ranking reported as $t$1

The empirical pattern reported across all three evaluations is consistent with the paper’s central claim: decomposition is most helpful when the task is difficult or when users are uncertain. The paper interprets this as evidence that interface design can materially improve the quality of human feedback used in alignment.

6. Limitations, trade-offs, and broader significance

The clearest trade-off is speed. DxHF improves accuracy, but it adds an average of 18 seconds to feedback time (Shi et al., 24 Jul 2025). The paper treats this as an operational design choice rather than a defect in isolation: whether the accuracy gain justifies the extra time depends on the annotation context.

The method is also explicitly limited in scope. The paper states that DxHF is effective for factual or task-oriented comparisons but is not well suited to judgments requiring holistic assessment of coherence, tone, style, or overall impression (Shi et al., 24 Jul 2025). That limitation follows directly from the decomposition strategy itself. When quality is aggregative across local informational units, claim decomposition is well matched to the task. When quality emerges from global discourse organization or rhetoric, claim decomposition may obscure relevant properties. A user quotation reported in the discussion notes that “analysis on its own can be misleading without also reading the entire text separately.”

A second limitation is dependence on decomposition quality. DxHF relies on GPT-4-generated claims and GPT-4-generated keywords. If these are incomplete, unfaithful, or awkwardly phrased, the interface can mislead users. The paper attempts to mitigate this with prompt constraints, semantic fidelity, and continued access to the original text, but the dependency remains structural (Shi et al., 24 Jul 2025).

A third limitation is the possibility of interface-induced bias. Relevance highlighting, decomposition, and linking are themselves outputs of model-based processing. The paper acknowledges that these mechanisms could steer attention toward some aspects of a response and away from others. The fold/unfold design, preservation of narrative order, and transparency of access to full text are presented as mitigations rather than eliminations of this risk.

The simulation study has its own limitations. The authors note that simulated strategies aggregate all included claims too uniformly, whereas real users selectively inspect only some claims (Shi et al., 24 Jul 2025). They attribute the plateau at higher rationality partly to this simplification. This suggests that the simulation is useful as a directional analysis rather than as a literal model of human comparison behavior.

The broader significance claimed for DxHF is methodological and infrastructural. The paper argues that alignment quality is partly an HCI problem: if the interface for collecting preferences is poorly structured, then the resulting labels will also be poor, regardless of downstream optimization. DxHF therefore exemplifies a broader principle within alignment research: annotation interfaces are not neutral containers for human judgment but active determinants of feedback quality (Shi et al., 24 Jul 2025). Within that framing, DxHF is best understood as a structured preference-elicitation interface for long-form outputs, optimized for claim-level comparison under uncertainty rather than as a general replacement for holistic human evaluation.