Loss-Scaled, Deficit-Aligned Seed Injection for LLM Research Ideation

Systems that use large language models (LLMs) to propose research ideas often improve their output by showing the model a "seed": a problem, method, or domain taken from a paper in another field. A recent diagnostic study found a puzzle. Seeds retrieved deliberately from other fields beat seeds from the same field, yet they only tie with seeds picked at random. Under random seeds, the model also tends to name the borrowed concept without actually using its mechanism. We propose to treat this as a question about noise, drawing on what stochastic gradient descent (SGD) theory says about when noise helps a search. That theory shows the useful noise in SGD is not uniform. Its size scales with the current loss, so it fades as the optimizer nears a good solution. Its direction lines up with the directions in which the landscape is most constrained. Random seeds match neither property.

The core idea fits in two sentences. First, run an iterative idea-refinement loop in which a critic scores each part of the current proposal. Second, inject cross-domain seeds whose number and domain distance grow with the overall score deficit and which are aimed at the specific component that is failing, rather than at the problem as a whole. We predict this recovers the novelty benefit of random seeds without the hollow, name-only borrowing. A component-targeted query should also retrieve ground-truth inspirations more often than a whole-problem query. That second prediction can be checked against labelled data, not only against LLM judges.

Core Claim

In SGD, noise helps a search when it is large where the search is failing, small where it is succeeding, and concentrated along the directions that are actually blocked. We claim the same two properties, scaling with the deficit and alignment with the failing component, explain when cross-domain seeds help an LLM ideation loop. We also claim that alignment is the property current systems lack. The random-seed tie reported by Read, Grep, and Synthesize is consistent with this account. Random seeds supply the escape that same-field retrieval lacks. Because they are not aimed at what is broken, the model cites them but does not use them.

We are borrowing a design principle and a testable signature, not a theorem. An idea space has no gradient and no smooth Hessian, and we do not pretend otherwise. The optimization papers play two roles here. They supply the design, and they supply a prediction about how long a search stays stuck in an incremental idea before it escapes.

Why the Optimization Literature Is the Right Lens

Three results carry over as design constraints. Mori et al. show that SGD noise variance is proportional to the loss. This turns escape from a basin into a power law in the ratio of losses, rather than the exponentially slow escape produced by constant "thermal" noise. It also means the noise vanishes at a solution with zero loss. Hodgkinson and Mahoney show that multiplicative noise hops between distant basins while keeping local resolution, and that additive noise "smears" the landscape instead. They demonstrate this on the Musk dataset and on the one-dimensional objective $f(x)=\tfrac{1}{10}x^2+1-\cos(x^2)$. Wu, Wang, and Su show that SGD noise covariance aligns with the Fisher information and Hessian, so noise is strong along sharp directions and quiet along flat ones. Wang and Wu formalize this as "noise geometry."

Mapped to ideation, the first two results say seed injection should shrink as the proposal improves, so it does not disrupt a good idea. The third says seeds should be aimed at the proposal's most constrained part. None of these papers touches ideation. As the file records, no work connects the two literatures.

On the ideation side, Chain of Ideas arranges retrieved papers into chronological citation chains. These beat an unorganized "pile" of papers by 65 Elo points, which shows that how retrieved material is structured matters. The Chain of Ideas agent uses no critic-driven, state-dependent control of what gets retrieved. ResearchTown simulates communities of research agents and notes that interdisciplinary ideas often come out as shallow mashups. It has no mechanism for deciding when or where to import foreign ideas. Both papers motivate our work. Chain of Ideas also supplies a strong baseline.

Research Program

We build a refinement loop. At each round $t$, a critic grades the current proposal on $K$ components: problem framing, core mechanism, a technical sub-step that is currently blocked, evaluation design, and feasibility. Each component $k$ gets a deficit $d_{t,k}\in[0,1]$, where a higher value means a worse score. The overall loss is $L_t=\tfrac{1}{K}\sum_k d_{t,k}$.

Seeds come from the PaperGym library of 1,167 seeds, drawn from 446 conference papers. Seeds are sampled with a probability that depends on two things: the component they target and their domain distance $\delta$ from the problem's home field. The rule is

$$ p_t(\text{seed}\mid k,\delta)\;\propto\;\frac{d_{t,k}}{\sum_j d_{t,j}}\cdot \exp\!\left(-\frac{\delta}{c\,L_t}\right), \qquad n_t=\lceil m\,L_t\rceil . $$

Here $n_t$ is the number of seeds injected in round $t$, and $c$ and $m$ are constants that set the scale. The first factor is the alignment term, the analogue of noise covariance following curvature. The exponential lets seeds come from farther fields when the loss is high, which is the analogue of noise variance being proportional to the loss. The retrieval query for component $k$ is built from the critic's written complaint about that component, not from the problem statement.

We will also build a mechanism-integration score. A second extraction pass, using the tool-augmented read, grep, and bash pipeline from PaperGym, checks whether the operative step of each seed appears in the final proposal's method section. A blind human sample of 20% of cases will validate this check. The score measures "citation abandonment" directly instead of inferring it.

Baselines

The naive reference point has two versions: the same refinement loop with no retrieval, and the same loop with seeds retrieved only from the problem's own field. PaperGym reports that cross-domain retrieval wins 60% of novelty comparisons against no retrieval and 67% against same-field retrieval (Read, Grep, and Synthesize). If our reproduction does not show similar margins, we will know our setup is off before we interpret anything else.

The strongest direct competitor on this question is the random-seed control. In PaperGym, targeted cross-domain retrieval won only 47% of comparisons against it, which is a tie. This is the result our claim must overturn. We will run random seeds inside the same loop with the same number of seeds per round as our method's average $n_t$, the same generator LLM, the same critic, and the same token budget.

The strongest existing structured method is the Chain of Ideas agent, which beat its runner-up by 65 Elo in human evaluation. We will run it on the same 30 PaperGym problems, matching its per-idea cost (reported at $0.50) and giving it the same number of generation calls. PaperGym's multi-domain query paraphrasing will serve as the breadth baseline, since it reaches 7 of 7 domains for every problem. Beating a baseline means one of two things. Our method either wins more than 55% of blind pairwise comparisons on novelty while also scoring higher on mechanism integration, or it ties on novelty and wins clearly on integration and feasibility.

Two ablations isolate the key ingredients. The first keeps the loss scaling but retrieves against the whole problem, removing the alignment term. The second keeps the alignment but fixes $n_t$ and the distance temperature, removing the loss scaling. The ablation most important for honesty is a threshold rule: inject one far-domain seed aimed at the worst component only when the critic rejects. If the threshold rule matches the continuous rule, then the physics-inspired scaling adds nothing beyond "aim at what is broken," and we will report that.

Why not just…?

Why not just use random seeds, since they already tie with targeted retrieval? This is the real threat. Our account says random seeds work because they escape the incremental basin, and fail because they are not aimed at anything. The prediction is falsifiable. If random seeds match our method on both novelty and mechanism integration at equal seed counts, alignment does not matter and the claim fails. The integration metric exists because PaperGym's novelty judges could not see this difference.

Why not just raise sampling temperature or use best-of-N sampling? Higher temperature adds noise that is the same everywhere in the proposal, the ideation analogue of the additive noise that Hodgkinson and Mahoney found smears resolution. The analogy does not prove this for LLMs, so we will run temperature-swept best-of-N generation with the same number of calls and report where it falls against our method.

Why not just retrieve a new seed whenever the critic rejects the proposal? This is the threshold ablation above, and it may well work as well as our method. If it does, the contribution reduces to deficit-aligned retrieval with a simple trigger. That is still a useful and cheaper method, and we would recommend it over the continuous rule.

Why not just use better whole-problem retrieval, such as Chain of Ideas chains? Chains give retrieved material a direction, but they are anchored to the problem as a whole. The ResearchBench experiment below tests whether aiming the query at the failing component beats a strong whole-problem query on ground-truth hits.

Proposed Experiments

The main experiment runs every method on the 30 PaperGym problem statements with three random seeds each. Pairwise novelty is judged by Claude 3.5 Sonnet with Semantic Scholar verification, following the PaperGym protocol, together with the mechanism-integration score and a feasibility rating. Chain of Ideas reports 74.9% agreement between GPT-4o judges and human experts. Because judge agreement is imperfect, we will use human experts to adjudicate all contrasts that are close to a tie.

The objective check uses the hypothesis-discovery ResearchBench of Liu et al., 1,386 papers across 12 disciplines with labelled inspirations. We will compare whole-problem queries against queries built from a critic's complaint about a first-draft hypothesis. The reference number is GPT-4o's 45.65% hit ratio at retrieving the true inspiration within the top 4% of candidates. The test is decisive: if component-targeted queries do not raise this hit ratio for the same model, the alignment mechanism is not doing what we claim. ResearchTown's PaperBench (resource) reports 0.68 semantic similarity for paper writing. It is a secondary sanity check that injecting seeds does not break reconstruction on in-distribution tasks.

The escape-time signature is our test of the physics analogy. For each problem, we count the rounds until the proposal leaves its initial incremental basin. We define that as the proposal's embedding moving beyond a fixed distance from its same-field nearest neighbours while the critic's loss stays below its starting value. The loss-scaled rule should escape quickly from low-scoring starts and rarely disturb high-scoring ones. A constant-rate random schedule should do neither. We calibrate this measurement pipeline first on the one-dimensional synthetic objective, where the multiplicative-versus-additive contrast is already known. If the ideation loop shows no dependence of escape time on the starting loss, we drop the dynamical interpretation and keep only the retrieval result.