Prospector: learning a compact graph traversal policy for scientific reasoning

1 Introduction

• Proprietary domain knowledge is often missing from off-the-shelf models, which may answer from general priors rather than grounded technical information.Semiconductor process data, for example, stays private to the company that owns it.Source: §1 · p. 2
• A knowledge graph stores sourced relations as edges, allowing an answer’s evidence chain to be checked and combined across documents.Engineering questions often depend on relations among materials, mechanisms, and process conditions scattered across sources.Source: §1 · p. 2
• Preselecting graph evidence commits to a question-centered subset before reasoning begins, potentially missing intermediate entities in multi-relation answers.Stepwise retrieval instead conditions each new search on what previous steps returned.Source: §1 · p. 2
• Prospector is a compact language model post-trained to choose what to examine next in a scientific knowledge graph and answer with a chain of cited edges.Its graph traversal uses three tools and conditions each step on what the policy has already read.Source: §1 · p. 2
• Prospector-4B approaches the strongest off-the-shelf model with the same graph tools, transfers zero-shot to three unseen fields, and runs faster and cheaper while keeping data on the owner’s machine.Its latent state separates following, backtracking, and answering, and can be modulated to steer behavior.Source: §1 · pp. 2–3
• The graph serves as the external encyclopedia, allowing the model to focus on traversing it rather than memorizing all its facts in its weights.Source: §1 · p. 2

2.1 Large language models with knowledge graphs

• Unlike prior graph-reasoning systems that use frozen models with fixed search procedures or prompts, Prospector trains the choice of what to read next into a compact model.Earlier approaches include pre-generation retrieval, prompted stepwise search, and graph traversal through dedicated interfaces.Source: §2.1 · p. 3

2.2 Training large language models on knowledge graphs

• Unlike prior language-model training on fixed-schema graphs such as Freebase and Wikidata, this work uses literature-extracted graphs with free-text relations and no schema.The trained policy is evaluated zero-shot on three graphs it did not encounter in training.Source: §2.2 · p. 3

2.3 Verifying answers against evidence

• Prospector checks support by requiring each cited edge to have been returned by the tools and the answer to appear in the cited edges.Source: §2.3 · p. 3

3 Results

• A three-edge evidence chain identifies HAR processing as the implicated stage, whereas all eight off-the-shelf models answer the example without the graph with a generic etching stage.The chain’s edges were extracted from three different documents.Source: §3 · p. 4

3.1 A compact policy approaches the strongest frontier models

• On 195 held-out items, Prospector-4B reaches 0.839 ± 0.015 accuracy, near gpt-6-astra’s 0.877 and above the other six off-the-shelf models’ 0.549 to 0.785.The models are compared while driving the same tools.Source: Figure 2 · p. 4, §3.1 · p. 4

3.3 Choosing the answer among retrieved candidates

• Verifier selection raises accuracy to 0.913, versus 0.839 for a single pass and 0.856 for majority vote over the same episode pool.The pool held a correct, grounded answer for 192 of 195 items.Source: Figure 4 · p. 6, §3.3 · p. 7

3.4 The graph-traversal capability transfers zero-shot to unseen domains

• With frozen weights, Prospector-4B reaches accuracy of 0.803 ± 0.017, 0.766 ± 0.022, and 0.809 ± 0.011 on the three unseen graphs, second to gpt-6-astra on each.The graphs cover bio-inspired materials, PFAS, and materials properties.Source: Table 1 · p. 7, §3.4 · p. 7

3.3 Choosing the answer among retrieved candidates

• Decision-level rounds reduced 4B decision misses from 20 to 11; the recipe raised Qwen3.5-0.8B/2B accuracy from 0.487/0.622 after supervised fine-tuning to 0.694/0.755, with own verifiers reaching 0.862/0.882.Source: §3.3 · p. 7

3.5 Prospector is cheap to serve and can run privately

• On H200, Prospector-4B answers 10.7 questions per minute versus 4.1 for gpt-6-astra; its pass costs $0.57 per thousand questions versus $10.72 for deepseek-v4-pro.It also runs on a MacBook M4 Pro without a network connection.Source: Figure 5 · p. 8, §3.5 · p. 8

3.6 Reading and steering the next move from the policy’s latent state

• At layer 20, linear readings predict next moves zero-shot on bio-inspired materials: AUC is 0.99 for answering versus continuing and 0.94 for following versus backtracking.The corresponding following-versus-backtracking AUC on the semiconductor graph is 0.91.Source: §3.6 · p. 8, Figure 6 · p. 9
• Adding the backtrack-to-follow direction lowers mean tool calls without accuracy loss at moderate dose; answer-direction steering advances or delays final answers on seen and unseen graphs.The same trends are also reported on the PFAS and materials-properties graphs.Source: §3.6 · pp. 8–9, Figure 6 · p. 9

Separation of knowledge from reasoning

• Separating knowledge from reasoning lets a compact policy carry its graph-search capability to graphs it has never read.The external graph can also be corrected or extended without retraining the model.Source: §4 · pp. 9–10
• The policy’s latent state indicates its next move on seen and unseen graphs, and modulation can shift its traversal behavior mode.The discussed modes are following, backtracking, and answering.Source: §4 · p. 10

Serving on private data

• Prospector can run on a laptop without a network connection, keeping questions, graph edges, and answers on the owner’s machine.The discussion links this local serving to frozen-policy transfer on graphs the model was never trained on.Source: §4 · p. 10

Extension to hypothesis generation

• Open-ended hypothesis generation lacks the graph-derived answer set used here, so judging whether a claim merits an experiment requires a second criterion.The authors suggest explicit rubrics for plausibility, novelty against literature, and experimental cost as one possible direction.Source: §4 · p. 10

5 Conclusion

• Prospector prospectively gathers evidence turn by turn through tool calls on free-text knowledge graphs and approaches frontier models while running on a laptop.The paper describes it as faster and cheaper than the best frontier models.Source: §5 · p. 10
• Trained only on semiconductor manufacturing, Prospector approaches frontier performance zero-shot on unseen bio-inspired-materials, PFAS, and materials-properties graphs.The model’s weights remain frozen during this transfer.Source: §5 · p. 10
• The policy’s latent state clusters by traversal mode, predicts its next move on unseen graphs, and can steer behavior from one mode to another.The paper presents this as evidence about the learned traversal representation.Source: §5 · p. 10

6.1 Graph datasets

• Four graphs support the study: semiconductor manufacturing for training and main evaluation, plus bio-inspired materials, PFAS, and materials-properties graphs.The semiconductor graph was built for SAGE-QA; the other graphs were released with SciAgents or GraphAgents.Source: §6.1 · p. 10

6.2 Benchmarks

• Questions are generated from sampled self-avoiding gold paths, with intermediate-node and answer leakage filtered out; held-out sets contain 195, 210, 228, and 223 items across the four graphs.The procedure also creates 3,971 training items, none sharing a gold path with an evaluation item.Source: §6.2 · p. 11

6.3.1 Tools

• The graph remains outside the context window; find locates nodes, expand returns their edges, and connect returns up to three shortest paths of at most three hops.The harness refuses a withheld tool if the policy calls it.Source: §6.3.1 · p. 11

6.3.2 Episode

• Each episode alternates reasoning with exactly one tool call or a final answer, which includes a cited relation chain and an asserted answer.Episodes begin by locating the question’s named entity with find and then expand outward.Source: §6.3.2 · p. 11

6.3.3 Gate

• An answer passes the gate only when every cited edge was seen during the episode and the answer is supported within those edges; correctness requires an answer-set member.Source: §6.3.3 · p. 11

6.4.1 Closed book

• The closed-book baseline receives only the question, measuring how much graph content the model already holds.Source: §6.4.1 · p. 11

6.4.2 Retrieved subgraph

• The retrieved-subgraph baseline answers in one pass from edges selected around the question’s entity, using the better of k-hop and Personalized PageRank for each graph.Both selection rules are seeded with the named entity, unlike Prospector, which must locate it with find.Source: §6.4.2 · p. 11

6.5 Training recipe

• Training uses fixed base weights with rank-16 adapter updates, followed by demonstration fine-tuning, two reinforcement-learning phases, rejection-sampling rounds, and verifier training.Source: §6.5 · p. 12

6.5.1 Supervised fine-tuning

• Supervised fine-tuning learns reasoning and tool-call text from teacher episodes, while tool replies are masked from the next-token loss.Rejection-sampling rounds blend up to two policy episodes per question, shortest first, with teacher demonstrations.Source: §6.5.1 · p. 12

6.5.2 Reinforcement learning

• Reinforcement learning uses GRPO with 16 rollouts per group, rewarding gated success and otherwise a quarter of gold-path recall.The two phases train full episodes and decision-state continuations, respectively.Source: §6.5.2 · p. 12

6.5.3 Verifier

• A separate yes/no verifier scores sampled episodes on whether the asserted answer is correct and supported by their cited connected edge chain.It ranks candidates using the margin between yes and no reply-token log probabilities.Source: §6.5.3 · p. 12

6.6 Reading and steering the latent state

• Latent-state analysis labels post-expand turns as follow, backtrack, or answer, with 705, 668, and 104 semiconductor turns and 696, 741, and 135 bio-inspired-materials turns, respectively.A follow expands a node on the current page; a backtrack expands one named earlier or in the question; an answer ends the episode.Source: §6.6 · p. 13
• Figure 6A projects semiconductor and zero-shot bio-inspired-materials states onto Fisher discriminant axes; panel B evaluates two move-reading directions using ROC AUC.The readings distinguish follow from backtrack and answer from either continuing move, with semiconductor holdout and bio-inspired-materials zero-shot evaluation.Source: §6.6 · p. 13
• Panel C tests steering with layer-20 unit vectors for follow and answer, added from the header token through generation, against a random-vector baseline.The follow vector points from the backtrack mean to the follow mean; the answer vector points from the other-move mean to the answer mean.Source: §6.6 · p. 13

6.7 Configuration

• Evaluations cap runs at 20 tool calls and 400 seen edges, while Prospector uses a 16,384-token context except for PFAS, which uses 32,768 tokens.Prospector-4B is the stage-4 Qwen3.5-4B configuration with rank-16 adapters and two decision-level rounds; turns use greedy decoding with at most 1,024 new tokens.Source: §6.7 · p. 13

• Claude Code assisted with software development, including the evaluation harness, training pipeline, figure scripts, and drafting and editing parts of the manuscript.The authors say they designed the study, reviewed the code and text, checked results against run records, and take full responsibility.Source: Use of Generative AI · p. 13

• Taiwan Semiconductor Manufacturing Company supported the work.Source: Funding Acknowledgment · p. 13
• Additional support came from MIT’s Generative AI Impact Consortium.Source: Funding Acknowledgment · p. 13
• MIT Office of Research Computing and Data and a Thinking Machines Lab research grant provided partial compute support; Yu-Chuan Hsu was thanked for helpful discussions.Source: Funding Acknowledgment · p. 14

• The prompt offers find for matching nodes, expand for connected edges, and connect for shortest paths between known endpoints.Find returns no edges; expand can filter or page results, while connect returns up to three shortest paths.Source: A System prompt · p. 20
• For nodes with more than 150 connections, expand returns a relation summary rather than edges, so the prompt advises narrowing first.The graph has 5,289 distinct free-text relation names, which must be copied exactly from observations when filtering or citing.Source: A System prompt · p. 20
• The policy must call exactly one tool per turn or provide its final answer, and it can see only what tools return.Tool calls use a specified block; final answers include a graph-edge chain and a concise answer.Source: A System prompt · p. 20
• Final answers must cite only observed edges that support the answer chain, never invent graph elements, and state what is missing if the tool budget runs out.The prompt recommends the fewest expansions that reach the answer.Source: A System prompt · p. 20

B Example episode

• Prospector begins by searching for the named processing method, and the first tool call finds six matching nodes.The episode has a budget of 20 tool calls.Source: B Example episode · p. 20
• An attempted expansion of “Ion” returns no edges, while another observation reports 18 new edges and 40 edges seen so far.The episode’s observations track newly returned edges, accumulated edges, and remaining tool calls.Source: B Example episode · pp. 21–22
• An expansion returns edges including that the wafer-holding system transfers particles to the wafer back side and optimizes charge build-up.The retrieved edges also link the system to a transfer chamber, cluster tools, and photolithography.Source: Table 2.3 · p. 21, B Example episode · p. 22
• A three-hop path connects the processing method to the wafer-holding system, which transfers particles to the wafer back side.The path runs through fabrication and the wafer back side.Source: B Example episode · p. 22