One-sentence paper summaries now sit in the places readers look first: search results, indexing pages, conference front matter, and paper feeds. These summaries are usually written by a neural generator that treats text as a flat string of words. That design invites one specific, costly error. The generator keeps the main claim ("our method improves accuracy") and drops the clause that limits it ("on low-resource languages", "when labels are noisy", "may"). The result is fluent, short, and an overclaim.
We propose building these summaries by deletion from the paper's own sentences rather than by free generation. Each sentence is parsed into a dependency tree, a graph in which each word points to the word it grammatically modifies. An integer program then picks which words to keep under a word budget. It carries one new hard rule on top of the classical ones: if a claim's main verb survives, every word that narrows that claim's scope (conditions, negations, hedges, restricting prepositional phrases) must survive with it. In two sentences: summaries at extreme compression should be cut out of the source with grammar-respecting scissors, and the scissors should be forbidden from cutting a claim away from its conditions. We test whether this lowers the overclaiming rate of scientific TLDRs compared with LLM generation and post-hoc pruning, at an acceptable cost in content overlap.
Core Claim
Classical dependency-constrained compression enforces that a retained word keeps its head. That guarantees grammaticality but says nothing about meaning-critical dependents. The head-presence constraint of Constraint-Based Sentence Compression: An Integer Programming Approach and the obligatory-argument constraints of Global Inference for Sentence Compression: An Integer Linear Programming Approach both treat adverbial clauses and modifiers as freely deletable, and these are exactly the units that carry scientific scope. The recent factuality line goes further and deletes modifiers on purpose. Improving Factual Consistency in Abstractive Summarization with Sentence Structure Pruning (SSP) prunes 20 types of dependency units from news summaries to remove hallucinated content. That is the right move when the modifier was invented, and the wrong move when the modifier is a true condition from the source.
Our claim is that, at TLDR-scale compression of scientific text, the binding failure is dropping conditions rather than inventing facts. A single "scope-coupling" constraint inside a tree-pruning program removes that failure by construction while staying close to abstractive systems on content overlap. Let $x_i \in \{0,1\}$ indicate whether word $i$ of a candidate source sentence is kept, $s_i$ be a salience score for word $i$ produced by an LLM, $B$ the word budget, $h(i)$ the dependency head of word $i$, and $Q$ the set of edges $(h, q)$ where dependent $q$ is a scope qualifier of predicate $h$. The program is
$$\max_x \sum_i s_i x_i \quad \text{s.t.} \quad \sum_i x_i \le B,\qquad x_i \le x_{h(i)}\ \forall i,\qquad x_q \ge x_h\ \forall (h,q)\in Q .$$
Together with head-presence, the last constraint makes a qualifier and its predicate stand or fall together. The budget may therefore force the solver to drop a whole claim, but it can never shorten a claim by stripping its condition. Every content word in the output comes from the source by construction, so each TLDR is traceable to its source span.
Relation to Prior Work
The proposal sits between two bodies of work that have not met. The scholarly side supplies the task and the motivating failure. TLDR: Extreme Summarization of Scientific Documents defines roughly 250:1 summarization and introduces CATTS, a BART model with title scaffolding over the abstract, introduction and conclusion (AIC). It generates freely, with no structural constraint. Introducing Spotlight: A Novel Approach for Generating Captivating Key Information from Documents trains reader-facing hooks with SFT and DPO, and its authors report failures in which engagement wins over grounding and commercial aspects are presented as research contributions. That is direct evidence that preference-tuned generation drifts toward overclaiming. We take both as motivation and as baselines. We do not try to match Spotlight's engagement goal; our target is faithful TLDRs.
The compression side supplies the method and a constraint on the design. We reuse the integer-programming formulation of Clarke and colleagues, and the extract-then-compress pattern of Neural Extractive Text Summarization with Syntactic Compression (JECS). JECS showed on news that compressing extracted sentences beats pure extraction, but its deletable-span inventory (appositives, relative clauses, adjectives) includes nothing scope-aware. Subtree Extractive Summarization via Submodular Maximization uses "must-link" constraints over dependency graphs, which is the closest structural relative of our scope-coupling rule. Its must-links protect obligatory arguments for readability, not conditions for truthfulness.
SSP and Improving Factual Consistency in Summarization with Compression-Based Post-Editing (CompEdit) appear to contradict us, since both fix errors by deleting. The two are compatible. Their deletions target content absent from the source, whereas our constraint protects content present in the source. Our system can still delete whole unsupported claims, because a dropped predicate takes its qualifiers with it. It cannot leave a supported claim standing without its conditions.
Parse Trees Guided LLM Prompt Compression (PartPrompt) shows that LLM-derived token values propagated over parse trees beat flat token dropping. We borrow the idea of combining LLM scores with tree structure, but our output is a reader-facing sentence rather than a prompt. Finally, Evaluating the Syntactic Transformations in Gold Standard Corpora for Statistical Sentence Compression warns that human compressions involve reorderings that pure deletion cannot express. This sets a ceiling on how fluent our outputs can be, and we measure that ceiling directly.
Research Program
The system has three parts. The first is candidate selection. An LLM ranks sentences in the AIC sections by how central they are to the paper's contribution, and the top few become candidates. This mirrors the AIC proxy that CATTS uses.
The second is word scoring and qualifier detection. The same LLM assigns salience to each word, using the same per-token probing approach that PartPrompt uses for entropy. A qualifier detector marks the edges in $Q$, built from Universal Dependencies relations (adverbial clauses with conditional or temporal markers, negation, modal auxiliaries and hedging adverbs, and oblique or nominal modifiers attached to the main predicate that restrict setting, data, or population). We validate the detector against human annotation and report its precision and recall, because a missed qualifier is a silent guarantee failure.
The third is solving. One integer program is solved per candidate sentence; the instances are small and solve in milliseconds. The highest-scoring feasible compression is emitted. If no candidate can fit the budget with its qualifiers attached, the system lengthens the output rather than violating the constraint, and we report how often this happens.
Two known weaknesses get dedicated analysis. Parser fragility on scholarly prose, flagged as an open problem for all tree-pruning methods, is measured by hand-checking parses on a sample and attributing each constraint failure to a parse error or a detector error. The deletion-only fluency ceiling is measured with CoLA-style acceptability scores and with human judges. We also test a constrained variant in which an LLM may change only inflections and function words, and a checker verifies that the set of content lemmas is unchanged.
The new resource we build is a scope-annotated subset of the SCITLDR test set. For roughly 300 papers, annotators mark, in the source abstract, each main claim and the qualifiers that limit it. Against this subset, any system's TLDR can be scored for "qualifier drop": a claim asserted in the TLDR whose source qualifiers are missing. No such annotation appears in the material we have. It is necessary because ROUGE is blind to this error.
Baselines
The naive reference point is the single highest-ranked AIC sentence, used uncompressed, alongside the same sentence truncated to the budget. The uncompressed sentence is perfectly faithful but usually too long. The truncated version shows what blind cutting does. Beating these means matching the uncompressed sentence's qualifier fidelity at TLDR length while keeping clearly higher acceptability than truncation.
The strongest published system on the benchmark is CATTS from TLDR: Extreme Summarization of Scientific Documents, which reports ROUGE-1 of 44.9 on AIC input and an average author-judged correctness of 2.5 out of 3. We will regenerate CATTS outputs on our annotated subset with the released configuration so that qualifier-drop rates are computed on identical papers. Beating it means a lower qualifier-drop rate and correctness at least equal to it, with ROUGE-1 no more than a few points below. We expect to trail it on ROUGE because deletion cannot paraphrase.
The strongest practical competitor is a prompted LLM: the same model we use for salience scoring, asked for a one-sentence TLDR from the AIC. We run it in two versions, a plain prompt and a prompt that explicitly instructs it to preserve all conditions and hedges. Both get the same input and a prompt-tuning budget equal to the effort spent tuning our detector on the development split.
A pipeline baseline applies SSP-style pruning from Improving Factual Consistency in Abstractive Summarization with Sentence Structure Pruning to the LLM's TLDRs. This tests whether post-hoc factual pruning addresses the problem; our hypothesis is that it worsens qualifier drop.
The key ablation removes only the scope-coupling constraint and leaves head-presence intact. This isolates whether the new constraint, rather than tree pruning in general, drives any improvement. A second ablation replaces LLM salience with uniform scores, to separate the contribution of the scorer from that of the structure.
Why not just…?
Why not just prompt a strong LLM to keep the qualifiers? This is the real threat. Instruction-following models may handle it well, and nothing in the material measures qualifier preservation for prompted LLMs. The Spotlight authors' report of engagement-driven drift under DPO is suggestive but concerns a different objective. The instructed-prompt baseline settles the question directly. If its qualifier-drop rate is statistically indistinguishable from ours, we will say so. The remaining contribution would then be the scope-annotated evaluation set, the finding that instruction suffices, and the traceability guarantee, which a prompted model cannot provide.
Why not just generate freely and then run a factuality pruner like SSP or CompEdit? Those pruners remove content unsupported by the source. A dropped condition is an omission, not an addition, so source-grounded checkers have nothing to catch in the TLDR itself. SSP's inventory of deletable units includes prepositional and adverbial modifiers, the very constructions that carry scope. The pipeline baseline tests whether post-editing leaves qualifier drop unchanged or makes it worse.
Why not just extract the best abstract sentence? Extraction is faithful, but TLDR: Extreme Summarization of Scientific Documents frames the task at about 250:1, and scientific sentences are long compound structures. On news, Neural Extractive Text Summarization with Syntactic Compression reports that compression lifts CNN ROUGE-L from 26.5 to 29.0 over its extractive baseline. The naive baseline shows how much length extraction costs on SCITLDR, and our length-overflow statistic shows how often even constrained compression must fall back toward it.
Proposed Experiments
The main evaluation runs on SCITLDR, scoring against both the author-written and peer-review-derived references, with AIC input. On the full test set we report ROUGE-1, ROUGE-2 and ROUGE-L. On the scope-annotated subset we report qualifier-drop rate, acceptability, and a human judgment in which annotators see the abstract and the TLDR and mark whether the TLDR overstates the paper. We use the same three-point correctness scale as Cachola and colleagues so that CATTS's reported 2.5 provides a reference anchor, while noting that our judges are not the original authors.
As a transfer check on reader-facing hooks, we apply the system to CSPubSum and compare against the DPO model of Introducing Spotlight: A Novel Approach for Generating Captivating Key Information from Documents, which reports BERT-F1 of 83.6 there. The question in this check is whether overclaiming drops and how much measured interest is lost. We expect some loss and will report it honestly.
The decisive check is fixed in advance. The proposal succeeds if the full system cuts the qualifier-drop rate substantially relative to both prompted LLMs and CATTS, with a target of at least halving it, while staying within three ROUGE-1 points of the instructed LLM and showing no loss in human-judged correctness. The scope-coupling ablation must account for most of that reduction; if it does not, the constraint is not the active ingredient.
The proposal fails informatively in two cases. If the instructed LLM matches us on qualifier drop, structure buys only traceability. If parse and detector errors account for most of the remaining failures, the bottleneck is analysis of scholarly syntax rather than the optimization.
References
- Constraint-Based Sentence Compression: An Integer Programming Approach
- Global Inference for Sentence Compression: An Integer Linear Programming Approach
- Improving Factual Consistency in Abstractive Summarization with Sentence Structure Pruning
- TLDR: Extreme Summarization of Scientific Documents
- Introducing Spotlight: A Novel Approach for Generating Captivating Key Information from Documents
- Neural Extractive Text Summarization with Syntactic Compression
- Subtree Extractive Summarization via Submodular Maximization
- Improving Factual Consistency in Summarization with Compression-Based Post-Editing
- Parse Trees Guided LLM Prompt Compression
- Evaluating the Syntactic Transformations in Gold Standard Corpora for Statistical Sentence Compression