A researcher whose objective is not strongly convex (unregularized logistic regression, LASSO, low-rank matrix recovery, a bilinear game) faces a recurring question: has someone already proved that my algorithm converges fast on my kind of problem? The answer is spread across hundreds of papers. Each one states its guarantee under a different weakening of strong convexity. The Polyak-Łojasiewicz inequality (a condition requiring that the gradient stays large whenever the loss is far from optimal) is one such weakening. Others are error bounds, quadratic growth, sharpness, and inverse-Lipschitz saddle operators. These conditions imply one another in a known but intricate pattern. Search tools match words, not logic. A query about "linear convergence without strong convexity" returns papers that say "strong convexity" often, which are exactly the wrong ones. It can also miss a result stated under quadratic growth that does apply.
We propose a retrieval system that extracts every convergence theorem in a corpus as a structured record. Each record holds the algorithm, the problem class, the assumptions, and the kind of rate. The system answers queries by logical entailment over a graph of known implications between assumptions, conditioned on side hypotheses such as smoothness and convexity. A theorem is returned only if the user's problem provably satisfies its hypotheses. Cells of the algorithm-by-assumption grid that contain no theorem are reported as candidate open problems. The core claim is that, for theorem-level questions in this field, assumption entailment is the missing relevance signal. It should beat strong dense retrieval with LLM reranking, especially on "trap" queries where lexical similarity points to a theorem whose hypotheses the user's problem fails.
Why this field, and what exists
The optimization side supplies both the motivation and the logic the system reasons with. Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak-Łojasiewicz Condition shows two things. First, gradient descent, coordinate descent and SGD converge linearly under PL. Second, for smooth functions PL sits in a chain of implications with strong convexity, error bounds and quadratic growth, and proximal-PL is equivalent to the Kurdyka-Łojasiewicz condition with exponent one half. That chain is the seed of our implication graph. Error bounds, PL condition, and quadratic growth for weakly convex functions extends these links to weakly convex functions, and it is a direct test of whether edges carry the right side conditions. Several other papers add results stated under heterogeneous assumptions:
- Xu et al. on variance reduction under error bounds.
- Roulet et al. on sharpness with restarts.
- Contractivity and linear convergence in bilinear saddle-point problems on inverse-Lipschitz operators for linearly coupled games.
- On Linear Convergence in Smooth Convex-Concave Bilinearly-Coupled Saddle-Point Optimization on quadratic growth for coupled games.
- Exponential convergence rates for momentum SGD in the overparametrized setting.
These papers are not merged for us anywhere. The open problems named in this literature are themselves "empty cells" of exactly the grid we propose to build:
- direct acceleration under PL without restart wrappers,
- nonlinear coupling in saddle problems,
- stochastic minimax under inverse-Lipschitz conditions.
On the literature-synthesis side, the closest systems structure papers at a coarser grain.
- AgenticScholar builds a problem-method taxonomy and gap matrices over 237 papers. Its cells are "method applied to problem", not "theorem holds under hypotheses", and it has no notion of one cell implying another.
- AgentExpt recommends baselines and datasets from citation contexts. Its authors note the popularity bias this induces, which matters here because new assumption variants are rarely cited yet.
- The LLMEVALDB work in Can LLMs Help Uncover Insights about LLMs? extracts empirical numbers, not formal statements.
- OpenScholar shows that grounded retrieval removes citation fabrication. It is our strongest general competitor, but it retrieves passages by semantic similarity rather than checking hypotheses. - Scaling Scientific Knowledge Discovery with Neuro-Symbolic AI motivates anchoring generation to verified graph triples. We go a step further by reasoning over the triples, not just citing them.
The proposal therefore differs from all of these in its unit of retrieval (a theorem's hypothesis set) and its relevance function (entailment, not similarity).
Research Program
The system has three parts.
Extraction. An LLM-based extractor reads theorem statements and their surrounding assumption paragraphs and outputs a record:
- the algorithm and its variant (stochastic, proximal, accelerated, restarted),
- the problem class (smooth, composite, finite-sum, bilinear saddle),
- each assumption, normalized to a controlled vocabulary with its constants,
- the conclusion type (linear in function value, linear in distance, sublinear with exponent, linear to a noise neighborhood).
Unrecognized assumptions are kept verbatim and flagged instead of being forced into the vocabulary. This matters because the field keeps inventing variants, such as the proximal-map-residual PL variant in Linear Convergence of the Proximal Gradient Method under the PL Inequality and Its Variant.
The implication graph. This is hand-seeded from the relations stated in the Karimi-style PL paper and the weakly convex paper. Each edge is labeled with the side hypotheses under which it holds; for example, "error bound implies PL" requires smoothness, and some reverse directions require convexity. Each edge also stores a pointer to the source theorem, so every inference the system makes is auditable. We then test whether the extractor can propose new candidate edges from papers that prove equivalences. Each proposed edge is checked by a human before it is admitted, because a single wrong edge silently corrupts every downstream answer.
Querying. A user describes their problem's properties, either in plain language or by choosing tags. The system computes the closure of those properties under the graph. It returns every theorem whose hypotheses lie inside that closure, with the derivation path shown. It separately reports near-misses: theorems that need one hypothesis the user's problem lacks. A gap query lists grid cells reachable by no theorem.
The scientific questions are whether entailment-based retrieval outperforms similarity-based retrieval for theorem questions, and which component carries the gain: extraction, the graph, or the side conditions on its edges.
Evaluation corpus and benchmark
No existing benchmark tests hypothesis-aware theorem retrieval. LitSearch (LitSearch) and ScholarQABench test paper- or passage-level relevance, so we must build one.
The corpus will be a few thousand optimization papers on first-order convergence, built outward from the optimization papers in this file. Optimization-trained annotators write about 300 queries of three kinds:
- Entailment queries, where the correct theorems are stated under weaker or differently named assumptions than the query uses.
- Trap queries, where the most lexically similar theorems require strong convexity, or quadratic growth in the bilinear setting, that the queried problem lacks. LASSO with a non-unique solution set is the canonical case.
- Gap queries, whose gold answers include the open problems this literature explicitly names.
Two annotators label each query, and disagreements are adjudicated. As a sanity check that the system does not degrade ordinary search, we also report its retrieval stage on LitSearch.
Baselines
The naive reference point is BM25 over theorem-bearing passages. LitSearch reports that dense retrievers beat BM25 by nearly 25% in recall on general literature search. We expect BM25 to look worst on trap queries, where keyword overlap is actively misleading.
The strongest existing retrieval stack is a dense retriever with LLM reranking. For this we use GritLM-7B, which LitSearch reports at 74.8% recall@5 on its specific questions, with GPT-4o reranking adding a further 4.4%. The strongest full synthesis system is OpenScholar (Synthesizing scientific literature with retrieval-augmented language models). It reports about a 0% citation hallucination rate and outperforms GPT-4o by 6.1% and PaperQA2 by 5.5% in rubric correctness on Scholar-CS. There is no reported result found on Lacuna for any system on hypothesis-entailment theorem retrieval, so all numbers on our benchmark will be new.
For fairness, every system sees the same corpus and the same passage chunking. Every system that uses an LLM gets the same model and the same per-query token budget, and the rerankers get the same number of candidates. Prompt tuning for the baselines receives the same development-set budget as our extractor.
Beating the baselines means two things on held-out queries:
- higher recall@10 of gold theorems at matched precision, and
- a substantially lower unsound-return rate, meaning the fraction of returned theorems whose hypotheses the queried problem does not satisfy.
The unsound-return rate is the headline metric because returning an inapplicable theorem is the failure that actually misleads a researcher.
Two ablations isolate the key ingredient:
- extraction plus exact tag matching with no implication graph, and
- the full graph with its side-condition labels deleted.
If the first ablation matches the full system, the graph adds nothing. If the second matches it, the careful conditioning is unnecessary.
Why not just…?
Why not just use OpenScholar-style retrieval-augmented generation and let the LLM judge applicability? This is the real threat. A strong LLM reading a PL theorem may well know that PL is implied by strong convexity, so the question is empirical. What OpenScholar's results establish is citation faithfulness, meaning the cited passage exists. They do not establish hypothesis soundness, meaning the cited theorem applies to the user's problem. The trap-query split measures exactly this difference. If retrieval-augmented generation matches our unsound-return rate there, we will report that the structured layer is unnecessary and reduce the contribution to the benchmark itself.
Why not just put the implication diagram in the prompt? This is cheaper than a symbolic graph and could close much of the gap. We include it as a baseline: dense retrieval followed by an LLM reranker given the full edge list with side conditions. Our hypothesis is that the LLM will apply edges inconsistently once the derivation needs two or more steps, or depends on a side condition such as convexity. We will stratify results by derivation length. If prompting matches the symbolic graph at every length, we will adopt the simpler design.
Why not just maintain a hand-curated survey table? The Karimi-style PL paper already contains a careful implication picture, and a static table is reliable. But it covers the results its authors chose and freezes at publication. Later contributions sit outside it, for example the bilinear inverse-Lipschitz results and the PEP-based proximal-PL variants. We test this directly with a temporal split. The system is built on papers up to a cutoff, and we then measure how many post-cutoff theorems it correctly places in the grid without human edits. Since the survey table covers none of them by construction, this quantifies the update advantage.
Why not just ask a long-context LLM directly? The file reports that GPT-4o fabricates citations 78–90% of the time on scientific queries (OpenScholar), so ungrounded answers are not a credible competitor for theorem lookup. We include this condition only as a lower anchor for the LLM-based systems.
Decisive checks
Three results would make or break the proposal.
- Extraction accuracy. Measured against annotator-labeled hypothesis sets on 200 theorems, normalized-assumption F1 must be high enough that downstream errors are not dominated by extraction. If it is not, the project becomes an extraction problem first.
- Trap-query advantage. The full system must cut the unsound-return rate on trap queries well below the prompted-diagram baseline. The gain should also be largest for multi-step derivations; if it is flat across lengths, the graph is redundant.
- Gap detection. On gap queries, the empty cells reported by the system should recover the open problems this literature names: direct PL acceleration without restarts, nonlinear-coupling contractivity, and stochastic inverse-Lipschitz minimax. They should not be swamped by spurious gaps caused by extraction misses. Each reported gap will be expert-verified as genuine or an artifact.
References
- Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak-Łojasiewicz Condition
- Error bounds, PL condition, and quadratic growth for weakly convex functions, and linear convergences of proximal point methods
- Adaptive SVRG Methods under Error Bound Conditions with Unknown Growth Parameter
- Sharpness, Restart and Acceleration
- Contractivity and linear convergence in bilinear saddle-point problems: An operator-theoretic approach
- On Linear Convergence in Smooth Convex-Concave Bilinearly-Coupled Saddle-Point Optimization: Lower Bounds and Optimal Algorithms
- Exponential convergence rates for momentum stochastic gradient descent in the overparametrized setting
- AgenticScholar: Agentic Data Management with Pipeline Orchestration for Scholarly Corpora
- AgentExpt: Automating AI Experiment Design with LLM-based Resource Retrieval Agent
- Can LLMs Help Uncover Insights about LLMs? A Large-Scale, Evolving Literature Analysis of Frontier LLMs
- Synthesizing scientific literature with retrieval-augmented language models
- Scaling Scientific Knowledge Discovery with Neuro-Symbolic AI and Large Language Models
- Linear convergence of the proximal gradient method for composite optimization under the Polyak-Łojasiewicz inequality and its variant
- LitSearch: A Retrieval Benchmark for Scientific Literature Search