The Answer Depends on the Next Search

TL;DR for operators

An internal research assistant may need to discover several technical relationships before it can answer a question. Selecting documents or graph records only once, before reasoning begins, can exclude relationships that become relevant later.

Prospector¹ demonstrates that this search behavior can be learned by a compact 4B-parameter language model. On 195 semiconductor questions, it achieves 83.9% single-pass accuracy, compared with 87.7% for the strongest evaluated frontier model using the same graph tools. Selecting among 32 sampled episodes with a trained verifier raises Prospector’s score to 91.3%, but adds substantial inference work.

For enterprise R&D teams, the finding favors testing iterative evidence retrieval on questions involving multiple relationships. Local deployment could reduce data exposure and serving expenditure. Neither advantage eliminates the need to validate extracted knowledge or measure performance on real organizational questions.

The first retrieval can exclude the relationships that matter

Consider an engineering team asking how a manufacturing process influences a downstream material property. The answer may depend on intermediate mechanisms documented in separate technical publications. Those intermediate entities are absent from the original question, leaving a retrieval system without the information needed to identify every relevant relationship upfront.

One conventional approach selects a collection of documents or graph records before asking a language model to reason over them. Increasing the model’s reasoning capability cannot recover a relationship that the retrieval stage has excluded.

The Prospector study provides a controlled demonstration of this distinction. On its semiconductor-manufacturing benchmark, gpt-6-astra answers 47.7% of questions correctly with a preselected retrieved subgraph, but 87.7% when allowed to navigate the graph iteratively. Across eight evaluated off-the-shelf models, iterative graph access improves accuracy by 16–40 percentage points over the paper’s fixed-retrieval baselines (Section 3.1, Figure 2).

The comparison holds the model fixed while changing how evidence becomes available. It therefore isolates a consequential architectural choice: retrieval can finish before reasoning, or each new relationship can inform the next retrieval action.

The static baselines use specified k-hop and Personalized PageRank procedures, with selection based on gold-chain coverage for each graph. The result does not establish that every static or hybrid retrieval design must perform worse. It shows that, for these graph-derived multihop questions, deciding what to retrieve before discovering intermediate entities can be a substantial constraint.

A compact model can learn where to look next

Prospector separates scientific knowledge from the policy used to search it. Domain information remains in an external knowledge graph, while a Qwen3.5-4B-based model learns how to navigate that graph.

Its search interface contains three tools:

• Find locates entities by name, without retrieving their relationships.
• Expand retrieves incoming and outgoing relationships around a selected entity.
• Connect retrieves short paths between two located entities.

A question initiates a sequence of tool calls. Each returned relationship becomes available when the model decides which entity to inspect next. Episodes are limited to 20 tool calls and 400 observed edges.

The authors train this behavior progressively using supervised demonstrations, full-episode reinforcement learning, rejection-sampling fine-tuning, and targeted decision-level reinforcement learning.

The training-stage comparison is substantial (Section 3.3, Figure 4):

These stages support the interpretation that graph navigation is a trainable specialization, although the reported progression does not independently isolate every component of the training recipe.

The tool-withholding experiments provide a more specific mechanism test. Removing expand reduces accuracy from approximately 84.1% to 39.8%. Removing find lowers it to 74.0%, while removing connect leaves accuracy near 83.8% (Section 3.2, Figure 3D).

Prospector averages approximately 9.1 expansion calls per episode, compared with 2.4 find calls and 0.8 connect calls. The policy primarily succeeds by repeatedly inspecting local relationships, rather than relying on explicit shortest-path computation.

For teams designing graph-based agents, this ablation suggests that dependable neighborhood retrieval deserves priority over more elaborate graph-navigation features. Whether the same priority holds depends on the graph structure and question distribution.

Retrieving the correct answer does not mean selecting it

Once the model can navigate a graph, another failure becomes visible.

An unsuccessful episode may never retrieve the correct answer, retrieve it but select another entity, or produce a citation that fails the evidence-grounding requirements. These are distinct problems, and improving search breadth does not necessarily resolve all three.

Prospector’s decision-level reinforcement learning targets consequential intermediate states, particularly decisions about expansion and final-answer selection. In the reported comparison, decision misses fall from 20 to 11 benchmark items, while accuracy increases from 79.0% to 83.9%.

The authors also train a separate verifier to evaluate candidate answers and their cited graph relationships.

When the system samples 32 episodes for each semiconductor question, at least one correct, grounded candidate appears for 192 of 195 questions. The challenge is selecting that candidate consistently.

Source: Section 3.3, Figure 4C.

The 91.3% result exceeds the strongest evaluated frontier model’s 87.7% single-pass score, but these are different inference configurations. Prospector’s ordinary single-pass score remains below gpt-6-astra’s 87.7% and claude-opus-5.5’s 84.6%.

This distinction affects system design. A team encountering frequent retrieval misses should investigate graph coverage and navigation. When correct answers are retrieved but subsequently rejected or confused, targeted decision training or candidate verification may offer a better intervention.

The verifier’s improvement also has a cost: generating and evaluating 32 candidate episodes is materially different from executing one.

Transfer is promising; serving economics depend on configuration

A policy trained to navigate semiconductor literature might simply learn domain-specific search patterns. The authors therefore freeze its weights and evaluate it on three previously unseen scientific graphs.

Source: Section 3.4, Table 1. Results use the same tool interface and a 20-call budget; the leading model is gpt-6-astra.

These zero-shot evaluations are evidence of reusable navigation behavior across the tested domains. They are not evidence that the policy can immediately operate on arbitrary enterprise databases. The unseen graphs come from related scientific knowledge-extraction pipelines and retain the same navigation interface.

Deployment measurements offer another reason to investigate compact policies. With sequential execution on an H200, Prospector processes 10.7 questions per minute, compared with 4.1 for gpt-6-astra. The model also runs offline on a MacBook M4 Pro at 1.2 questions per minute (Section 3.5, Figure 5).

Under the paper’s concurrent GH200 rental assumptions, single-pass Prospector serving is estimated at approximately $0.57 per 1,000 questions, against $10.72–$209.76 for evaluated hosted-model alternatives.

Those estimates depend on hardware pricing, utilization, and workload. They do not include full graph-construction and maintenance costs, enterprise integration, or the additional computation required for 32-episode verifier selection.

For organizations handling confidential technical literature, local inference offers a potential data-locality advantage, provided graph storage, execution, and logging remain within the controlled environment.

A cited graph relationship is not independently verified science

Prospector’s evaluation uses questions generated from paths through existing knowledge graphs. Gold answers are defined using graph relationships, and the system’s grounding check requires cited edges to have appeared in retrieved observations and the asserted answer to appear in its cited evidence.

This provides machine-checkable consistency between the answer and the graph records retrieved during the episode.

It does not independently establish that those records correctly represent the original publications.

An extraction error, ambiguous entity name, or incomplete relationship can therefore support a graph-grounded answer that is scientifically misleading. The 856 held-out questions across four graphs measure performance against constructed graph answer sets, not independently validated scientific discoveries.

The paper’s latent-state experiments add a narrower insight into model behavior. Linear probes trained on semiconductor states predict whether the policy will answer or continue searching, reaching an AUC of 0.99 on both the semiconductor and unseen bio-inspired-materials graphs. Interventions on those representations also alter navigation tendencies (Section 3.6, Figure 6).

These are informative exploratory mechanism experiments. They demonstrate prediction and limited steering of defined navigation behaviors, not general scientific reasoning competence or assured improvements in decision quality.

Likewise, extending Prospector to new hypothesis generation remains a proposed research direction rather than a demonstrated capability.

What an enterprise pilot needs to establish

Cognaptus interprets these findings as a reason for R&D knowledge-management teams to compare retrieval architectures before committing to a larger general-purpose model.

A meaningful pilot would use actual technical questions whose answers may involve relationships across several documents. It would compare fixed retrieval, frontier-model iterative retrieval, and compact learned traversal against the same underlying evidence and evaluation criteria.

The evaluation should distinguish whether a failure originated in graph extraction, missing evidence, navigation, answer selection, or unsupported interpretation. Consequential answers need checks against original technical sources rather than graph-edge membership alone.

Enterprise AI architects should also compare systems at matched accuracy targets, accounting for inference, verifier sampling, hardware, graph maintenance, and operational overhead.

Prospector demonstrates that a relatively small model can learn to search structured scientific knowledge effectively, including across several unfamiliar graphs. Its strongest implication is architectural: when the next piece of evidence depends on what has just been discovered, the retrieval policy becomes part of the reasoning system.

Whether that architecture delivers reliable, economical answers in an enterprise depends on the quality of the underlying knowledge and the decisions those answers must support.

Cognaptus: Automate the Present, Incubate the Future.