Angular Sweep Outlines for Coherent Literature Surveys

Automated survey writers have two separate problems. They must choose which few hundred of many thousands of relevant papers to discuss, and they must put those papers in an order a reader can follow. Current systems handle the first by citation counts and the second by asking a language model to write an outline, then filling it in section by section. This note proposes a single geometric rule that does both. Place the candidate papers on a local two-dimensional map around an anchor paper. Sweep a ray around the anchor, the way a lighthouse beam turns. At each angle, keep only the first paper the ray hits. The papers hit in sweep order become the outline. Papers hidden behind a nearer, broader paper in the same direction are dropped as redundant. A narrow paper in an otherwise empty direction survives no matter how rarely it is cited.

The rule comes from computational geometry. A "ray-rotating query" turns a ray around a fixed point and reports the next obstacle that changes what the ray sees first. That is exactly the event sequence an angular sweep produces. The idea matters because it makes selection and ordering one explicit, inspectable object. Unlike a language model's outline, that object can be edited, re-run when new papers arrive, and checked against how human experts actually structure surveys.

Core Claim

The claim is empirical and falsifiable. On a suitably built local map, human-written surveys discuss their cited work in roughly angular order around their central contribution. Adjacent sections cover angularly adjacent directions, and each direction is represented by its nearest, least-redundant paper.

If this holds, an outline produced by an angular visibility sweep should give surveys that are more coherent across sections and less redundant than outlines written by a language model. It should also keep more minority findings than outlines built from citation-count backbones, with the same writer model and the same retrieved pool.

If human surveys are not angularly ordered on such maps, the idea is wrong, and a cheap pre-test (below) will show that before anything expensive is built.

To make "visible" precise, give each paper $j$ a position $(r_j, \theta_j)$ on the anchor's map, where $r_j$ is its semantic distance from the anchor and $\theta_j$ its direction. Also give it an angular half-width $w_j$ that reflects how broad the paper is. Paper $j$ is visible from the anchor when no nearer paper covers its direction:

$$ \text{vis}(j) \iff \neg\,\exists\, i:\; r_i < r_j \ \text{and}\ |\theta_i - \theta_j| < w_i . $$

Sweeping $\theta$ from $0$ to $2\pi$ and recording each change in the first-hit paper produces the ordered outline.

A section that bridges two anchors, $a$ and $b$, is treated as the straight segment between them. Its evidence set is the set of papers visible from at least one point on that segment, which geometers call weak visibility. Consecutive sections share an anchor endpoint, so every transition is tied to a paper both sections can see.

Relation to Prior Work

Two strands of geometry supply the method.

Weak Visibility Queries of Line Segments in Simple Polygons introduced the ray-rotating query. It finds, in logarithmic time after linear-space preprocessing, the next vertex hit by a turning ray. It uses this to compute what is visible from a query segment by moving a viewpoint along it. Our section-as-segment evidence rule borrows exactly that moving-viewpoint construction.

Visibility and Intersection Problems in Plane Geometry gave the classic efficient structure for firing a single ray. We do not claim an algorithmic contribution. Per-review candidate pools are a few thousand papers, and papers-as-obstacles form a polygon with holes, where the file notes the simple-polygon guarantees break down. We therefore run a plain sort-and-sweep per anchor. The geometry supplies the semantics of "first hit" and "occluded," not the speed.

A Practical Algorithm with Performance Guarantees for the Art Gallery Problem plays a motivating role. It found that visibility computation, not the optimization around it, dominates cost in practice, and that caching weak-visibility structure saves up to 87.3% of queries. That matters to us only for living reviews that re-sweep many anchors after each corpus update.

Three systems motivate the problem and set the bar.

SurveyMaster anchors section-by-section drafting on a "core-paper backbone" of under 5% of papers, chosen by citation volume and venue prestige. We replace that selection with a visibility rule that ignores citation counts. We also replace its heuristic topic-tree order with sweep order.

DiffResearch attacks outline drift by refining all sections at once with a diffusion language model, which generates by repeatedly denoising a whole draft. Its global structure is implicit inside that model. Ours is an explicit object that can be edited: moving an anchor re-derives the affected sections' evidence sets.

GLIMPSE showed that consensus-seeking summarization hides distinctive critiques and that uniqueness scoring recovers them. That motivates our expectation that angular occlusion keeps divergent work. GLIMPSE, however, only extracts sentences and imposes no document order.

On the coherence side, Re3 showed that what the writer keeps in view matters more than how much text it is fed. It reports a 14% absolute gain in plot coherence over a rolling-window baseline. We apply the same principle: each section's prompt contains only its visible evidence set plus the shared anchors. Prove Your Point! showed that planning claims before evidence before prose improves reasoning validity, from 80.26 to 84.64. We reuse this claim-then-grounds order inside each section.

Research Program

The work proceeds in four steps.

The first step builds the local map. For an anchor paper, take its retrieved neighborhood and embed each paper with a scientific text encoder. Radius is one minus cosine similarity to the anchor. Angle comes from projecting each paper's difference vector (its embedding minus the anchor's) onto the top two principal directions of all such difference vectors. Map quality is reported as the share of each paper's ten nearest neighbors that survive the projection. Low-fidelity anchors are split into several maps rather than forced into one.

A paper's angular half-width is the spread, on the same map, of the papers it cites and the papers that cite it. Broad papers therefore occlude more. Foundational syntheses naturally hide the incremental work they subsume, and isolated narrow papers stay visible.

The second step runs the pre-test on human surveys, before any generation. The goal is to measure how angularly ordered real surveys are.

The third step is generation. Anchors are chosen by a first sweep from the query's centroid. Each consecutive anchor pair defines a section segment whose weakly visible papers form its evidence. A fixed writer model drafts each section using claim-then-grounds planning, seeing only its evidence and the neighboring anchors.

The fourth step is revision. A judge flags incoherent transitions. The system responds by moving or inserting an anchor and re-sweeping, instead of rewriting prose blindly. This gives an explicit, traceable analogue of the global revision DiffResearch performs implicitly.

Baselines

The naive reference is a single frontier model given the same retrieved pool and asked to outline and write the survey. It uses the same writer model, the same token budget, and no selection rule. On SciSurveyBench, SurveyMaster reports a baseline without a backbone at 3.63 Fact Correctness and 3.62 Citation Precision. That is the floor we expect to reproduce.

The strongest competitor on SciSurveyBench is SurveyMaster itself, which reports 4.84 Fact Correctness and 4.49 Citation Precision on a 5-point scale, and 4.06 Critical Analysis. It is the strongest because it already solves drift with an explicit anchor set, which is precisely the ingredient we contest. If its code is available, we run it on the same retrieved pool with the same writer model. Otherwise we compare against its reported numbers, clearly labeled as not matched on writer or retrieval. Beating it means matching its factual scores within noise while winning on cross-section coherence and minority-finding retention. We do not expect large factual gains, since selection does not change fact checking.

On DeepResearch Bench, DiffResearch reports 48.03 overall and 48.20 Insight, against 46.45 overall for OpenAI Deep Research. Separately, DeepResearch Bench reports 49.7 RACE for Gemini-2.5-Pro Deep Research. These come from different papers and possibly different benchmark versions, so we compare only within a single scoring protocol. Beating DiffResearch's Insight score with an autoregressive writer would show that explicit geometric structure can replace diffusion's implicit global refinement.

The key ablations hold everything fixed except one part. The selection rule is swapped for maximal marginal relevance (a greedy pick that trades relevance against similarity to already-picked papers) or for citation-count top-k. The ordering is swapped for model-chosen, chronological, or shortest-tour order in the full embedding space. The map is swapped for a random two-dimensional projection. If random projections perform as well as the learned map, the geometry is decorative.

Why not just…?

Why not just use maximal marginal relevance for selection and let the model order the outline? This is the real threat. MMR already gives diversity, and language models are decent at ordering topics. The ablation grid above settles it directly: same pool, same writer, swapping only selection and only ordering. Our specific bet is that angular occlusion differs from MMR in keeping narrow isolated papers that MMR's relevance term suppresses. We measure this as retention of reference-survey papers with below-median citations. If MMR plus model ordering ties on coherence and on minority retention, we will report that the geometric rule adds interpretability but not quality.

Why not just order papers by a shortest tour through the full embedding space, with no 2D projection? A tour gives smooth adjacent transitions without losing information to projection. It does not, however, define occlusion, so it must be paired with a separate redundancy filter. It also does not give each section a shared anchor. The pre-test compares angular order against tour order as predictors of human section order, so the data decide which structure humans actually follow.

Why not just use a citation-count backbone or a knowledge graph, as SurveyMaster and Scaling Scientific Knowledge Discovery with Neuro-Symbolic AI and Large Language Models do? These are strong for factual grounding. The neuro-symbolic work eliminated hallucinations on a 49-paper corpus. But citation thresholds are biased toward consensus. The knowledge-graph pipeline also stitches sub-graph summaries together by concatenation, which leaves cross-section flow unconstrained. Our method is compatible with both, since graph facts can ground each section's evidence. The comparison isolates only the selection-and-ordering step.

Why not just give a million-token model everything? SurveyMaster's backbone-free baseline and Re3's rolling-window result both indicate that dumping context degrades structure. We include this as the naive reference rather than assume it fails.

Proposed Experiments and Decisive Checks

The decisive first experiment uses SciSurveyBench's human reference reviews. For each review, build maps around its most central cited paper. Then measure the circular rank correlation between each cited paper's sweep position and the position of the section that cites it. Compare that correlation against the same statistic for chronological order and for shortest-tour order, and against a null in which cited papers are shuffled across sections.

If angular order does not beat tour order and the shuffled null by a clear margin, we stop, or we retreat to using the sweep only as a redundancy filter. This check costs no generation.

The main experiment runs on SciSurveyBench's 100 tasks and on DeepResearch Bench. The two groups of metrics are:

  • Reported metrics: Fact Correctness, Citation Precision, and Critical Analysis on SciSurveyBench; overall and Insight scores on DeepResearch Bench.
  • Metrics we add:
  • pairwise human judgments of cross-section coherence;
  • redundancy, measured as the rate of near-duplicate claims across sections;
  • transition grounding, the share of adjacent sections that share cited evidence;
  • minority retention, defined above.

Human pairwise judgments are required because LLM judges suffer from the same position biases discussed in the long-form coherence literature.

Retrieval quality is held fixed across all conditions, and we check it on LitSearch, for which no reported result was found on Lacuna in the supplied excerpts. That check ensures differences come from selection and order, not recall.

A final experiment tests the one place where speed matters: a living review. We replay a field's papers month by month and measure how many anchor re-sweeps are needed per update. We then check whether caching weak-visibility structure yields savings comparable to the art-gallery result.