When a language model synthesizes scientific evidence, it has to decide how much of each study to read: the title, the abstract, the results tables, or the full text. Current pipelines barely make this choice. They pour everything into the prompt, or they keep or drop whole papers. Both options fail in documented ways. On MedEvidence, a benchmark built from Cochrane systematic reviews, accuracy falls as more studies and tokens are added. It collapses from about 92% when all source studies agree to about 41% when they conflict, and models treat small, biased trials as seriously as large rigorous ones (Can Large Language Models Match the Conclusions of Systematic Reviews?). We propose to treat reading depth as an integer budget per study. A whole number $x_i \in \{0,\dots,B\}$ says how deeply study $i$ is represented in the prompt. The vector is chosen to maximize a diminishing-returns objective that weights evidence by study quality and keeps both sides of a disagreement visible, all under a fixed token budget.
The bet is specific. The model's skepticism and context failures are partly failures of input allocation, not only of reasoning. A large, well-blinded trial should get its full methods and results. Ten tiny unblinded trials pointing the other way should be shown, but compressed, so their combined weight grows slowly. Optimization on the integer lattice (the grid of whole-number vectors) supplies the formal tool: functions with coordinate-wise diminishing returns, which can be maximized greedily with guarantees. In this proposal the lattice view is the modeling language and the evaluation lens. We do not claim it as a speed trick, for reasons given below.
Core Claim
Fix a token budget. A synthesizing model given a quality-weighted, depth-allocated view of the evidence will match Cochrane conclusions more often than the same model given all studies at full depth, all studies at abstract depth, or a binary selection of studies. The gain should concentrate in the conflicting-evidence questions where current models fail. The ingredient we expect to matter is the combination of two things: depth chosen per study, and concave, quality-weighted aggregation within each direction of effect. Either one alone should help less.
The Objective
Each candidate study $i$ is first read cheaply at the abstract level. A small model extracts from it a set of outcome claims $u$, each paired with an effect direction (for example, "mortality, favors treatment"). It also extracts design features: sample size, randomization, and blinding. From these features we form a quality weight $w_i$.
Each claim–direction pair $u$ accumulates evidence as follows:
$$c_u(x)=\sum_i w_i\,a_{iu}(x_i), \qquad f(x)=\sum_u g\big(c_u(x)\big)$$
In this formula, $a_{iu}(x_i)$ is the estimated fraction of study $i$'s information about $u$ revealed at depth $x_i$. It is nondecreasing and concave in depth, and is estimated from section structure and the abstract-level extraction. The function $g$ is a concave, increasing aggregator, such as $\log(1+c)$. We maximize $f(x)$ subject to $\sum_i t_i(x_i)\le T$, where $t_i$ is the token cost of study $i$ at a given depth and $T$ is the total budget.
Because $g$ is concave and each $a_{iu}$ is concave, the marginal gain of deepening one study shrinks in two ways. It shrinks as that study is read more deeply, and it shrinks as other studies already cover the same claim. This makes $f$ DR-submodular in the sense formalized by On Multiset Selection With Size Constraints, so the budgeted greedy algorithm of Optimal Budget Allocation: Theoretical Guarantee and Efficient Algorithm applies. That algorithm repeatedly gives one more unit to the coordinate with the best gain per unit cost. Its $(1-1/e)$ guarantee transfers directly to our setting.
Two design choices carry most of the load. First, opposing directions are separate claims $u$, so conflict is never optimized away. Each side receives representation that grows with its quality-weighted evidence. Second, the concave $g$ borrows the outlier-damping role that concave transforms play in Subspace Selection via DR-Submodular Maximization on Lattices. That paper studies geometric objectives on subspace lattices. We reuse its idea that concavity limits the influence of many weak contributors, and apply it to clinical evidence instead.
We deliberately do not use the non-monotone formulation suggested by the conflict collapse. That formulation would subtract a penalty for contradicting studies and solve with the double-greedy methods of Non-Monotone DR-Submodular Function Maximization or Profit Maximization in Social Networks and Non-monotone DR-submodular Maximization. A systematic review must report disagreement, not hide it. An objective that rewards dropping inconvenient studies would trade correctness for apparent agreement. We will run the penalized variant once, as a negative control. We expect it to raise accuracy on agreeing questions and to misreport the conclusion on genuinely mixed ones.
Relation to Prior Work
Set-based summarization treats sentences or documents as in-or-out items. Examples include Subtree Extractive Summarization via Submodular Maximization, Fast Constrained Submodular Maximization: Personalized Data Summarization, and Deep Submodular Networks for Extractive Data Summarization. These methods neither weight by study design nor model graded depth. We keep their coverage idea and change the decision variable.
The lattice algorithms themselves come from work on advertising and sensor budgets. The relevant papers are Optimal Budget Allocation: Theoretical Guarantee and Efficient Algorithm and the decreasing-threshold method of A Generalization of Submodular Cover via the Diminishing Return Property on the Integer Lattice, which uses a binary search over step sizes so runtime scales with $\log B$. Neither touches text. We supply the method; they supplied the mathematics.
On the literature side, three systems allocate effort, but only heuristically. Screening with dynamic few-shot examples escalates cases with confidence below 0.9 to a stronger model (Leveraging LLMs for Title and Abstract Screening for Systematic Review: A Cost-Effective Dynamic Few-Shot Learning Approach). SurveyMaster anchors drafting on a backbone of fewer than 5% "core papers," chosen by citations and venue (Automating Cognitive Distillation for Expert-Level Scientific Literature Synthesis). OpenScholar iteratively retrieves more passages when its self-critique asks for them (Synthesizing scientific literature with retrieval-augmented language models). These systems motivate the problem. We differ by making the allocation an explicit, quality-aware optimization over depth that can be audited after the fact.
The MedEvidence paper plays two roles. It motivates the work through its findings of a skepticism gap, a context tax, and conflict collapse. It also constrains the design: its finding that medically fine-tuned models underperform suggests that the fix should operate on inputs rather than on model weights.
Research Program
We first build a depth ladder for biomedical studies. Level 0 is title only, level 1 the abstract, level 2 abstract plus results tables, level 3 adds the methods, and level 4 is full text. For open-domain synthesis, depth is instead the number of retrieved passages taken from a paper, ranging from 0 to roughly 30. That larger $B$ is the one setting where the decreasing-threshold method is worth running.
We then build the cheap extraction step that produces claims, directions, and design features from abstracts. We measure its agreement with the risk-of-bias judgments in the Cochrane reviews, but use those judgments only as an oracle upper bound and never as input.
Finally, we test the assumption the whole approach rests on. We sample pairs of allocations $x \le y$ and add one unit of depth to the same study in each. If the true downstream quantity, the model's accuracy against the Cochrane conclusion, has diminishing returns, the gain should be smaller at $y$ than at $x$. We check how often it is. From the same samples we estimate a submodularity ratio, a number measuring how far the function is from exact diminishing returns, in the spirit of Fast Maximization of Non-Submodular, Monotonic Functions on the Integer Lattice. If real accuracy is far from diminishing-returns behavior, the surrogate's guarantee says little about what users care about, and we will report that.
Baselines
The naive reference is to give every study at full text, truncated to the budget. This is what long-context practice does. Under the context tax we expect it to degrade as the number of studies grows. The second naive reference is every study at abstract depth, the cheapest uniform option.
The strongest existing reference on MedEvidence is simply the best frontier model on the benchmark's own inputs. The paper reports that DeepSeek V3 and GPT-4.1 reach an accuracy ceiling of roughly 60–62%, about 41% on conflicting-evidence questions, and about 92% when all studies agree. Time-constrained human experts exceed 75%. We will rerun GPT-4.1 or the strongest model we can access with the published prompt. Beating it means higher accuracy at equal or fewer input tokens, with a significant gain on the conflict subset in a paired test across the 284 questions. Because the questions are few, we will report bootstrap intervals and not claim small differences.
The strongest heuristic competitor is a two-level policy. It ranks studies by the same automatically extracted quality weight, gives the top $k$ full text and the rest their abstracts, and tunes $k$ per budget on a held-out split. This uses exactly our information. It is the baseline most likely to tie us.
For open-domain synthesis on ScholarQABench, the reference is OpenScholar with its self-feedback retrieval loop, run over the OpenScholar DataStore. The file reports only relative figures: OpenScholar-8B exceeds GPT-4o by 6.1% and PaperQA2 by 5.5% in correctness on multi-paper synthesis, with roughly 0% citation hallucination. No absolute correctness score for it was found on Lacuna. Beating it means higher rubric correctness at matched retrieved tokens with no rise in hallucinated citations.
Two ablations isolate the key ingredient. The first keeps depth allocation but sets every $w_i=1$, which removes quality weighting. The second keeps quality weighting but restricts each study to levels 0 or 4, which removes graded depth. A third check runs exhaustive search over the surrogate on the smaller MedEvidence instances, where $5^n$ cheap surrogate evaluations are affordable. This tells us whether greedy loses anything relative to the true optimum.
All methods share the same synthesizing model, prompt template, token budgets, extraction outputs, and an equal tuning budget of held-out questions.
Why not just…?
Why not just rank by quality and read the top $k$ in full? This is the real threat, and we take it seriously. If $k$ is tuned well, it may capture most of the benefit, because depth allocation with nearly separable gains reduces to ranking. Our objective differs in two ways. It accounts for redundancy, since a second large trial on an outcome that is already covered earns little. It also reserves budget for the minority direction of effect. Both should matter mainly on conflicting questions. If the tuned two-level baseline comes within the bootstrap interval of our method on the conflict subset, we will conclude that a well-weighted ranking suffices and report that as the finding.
Why not just use classical set submodularity with one item per reading level? With five levels, there is no computational reason not to. The set greedy over "upgrade study $i$ by one level" items is the lattice greedy, and the $O(nB)$ cost that A Generalization of Submodular Cover via the Diminishing Return Property on the Integer Lattice avoids is trivial at $B=5$. We claim no speed advantage there. The lattice framing earns its place in two ways. It clarifies which objectives keep the greedy guarantee, since coordinate-wise diminishing returns is stricter than lattice submodularity. It also pays off computationally only for passage counts in ScholarQABench, where we will report the measured runtime gap honestly.
Why not just put everything in a long context and prompt the model to appraise risk of bias first? MedEvidence reports that accuracy falls with more studies and that frontier models plateau. Still, the file does not show that explicit appraisal prompting fails. We therefore include a "full context plus structured appraisal instructions" arm at matched model and prompt length. If it matches our allocation, the skepticism gap is a prompting problem and not an allocation problem.
Why not let an agent decide what to read next, as OpenScholar does? Iterative self-critique is a reasonable adaptive policy. The results of Adaptive Budget Allocation for Maximizing Influence of Advertisements warn that policies reacting to every noisy signal can underperform. Those results come from advertising cascades, however, and need not carry over. The OpenScholar baseline at matched tokens settles the comparison directly. A batch variant that re-runs our allocation once after level-1 extraction tests whether one round of adaptation is enough.
Decisive Checks
Three outcomes would confirm the proposal. Our method must beat full-context GPT-4.1 on the MedEvidence conflict subset at equal or lower token cost. It must beat both ablations, which shows that weighting and depth each contribute. The empirical submodularity ratio of real accuracy must be well above zero, which shows the surrogate tracks what matters.
Three outcomes would falsify it. The first is a tuned two-level heuristic matching us. The second is the appraisal-prompted long-context arm matching us. The third is accuracy so far from diminishing-returns behavior that the surrogate's optimum is uncorrelated with downstream correctness. Secondary checks follow the same design. We measure ScholarQABench correctness and citation accuracy against OpenScholar at matched tokens, and we sweep accuracy against budget on MedEvidence to see whether depth allocation flattens the context-tax curve.
References
- Can Large Language Models Match the Conclusions of Systematic Reviews?
- On Multiset Selection With Size Constraints
- Optimal Budget Allocation: Theoretical Guarantee and Efficient Algorithm
- Subspace Selection via DR-Submodular Maximization on Lattices
- Non-Monotone DR-Submodular Function Maximization
- Profit maximization in social networks and non-monotone DR-submodular maximization
- Subtree Extractive Summarization via Submodular Maximization
- Fast Constrained Submodular Maximization: Personalized Data Summarization
- Deep Submodular Networks for Extractive Data Summarization
- A Generalization of Submodular Cover via the Diminishing Return Property on the Integer Lattice
- Leveraging LLMs for Title and Abstract Screening for Systematic Review: A Cost-Effective Dynamic Few-Shot Learning Approach
- Automating Cognitive Distillation for Expert-Level Scientific Literature Synthesis
- Synthesizing scientific literature with retrieval-augmented language models
- Fast Maximization of Non-Submodular, Monotonic Functions on the Integer Lattice
- Adaptive Budget Allocation for Maximizing Influence of Advertisements