{"id":"5b1bd0e7-e224-4448-8e4c-7a2bf0538f81","arxiv_id":"2608.00432","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Training on synthetic search-open-write trajectories built from citation and hyperlink graphs improves the sample efficiency of downstream deep-research agent SFT.","lead":"This paper introduces a way to train deep-research AI agents on cheap offline data by converting citation links and Wikipedia hyperlinks into pretend search-and-write exercises. The method improved how well a 14-billion-parameter model performs after seeing only a quarter of the usual expert training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The offline oracle may leak through the queries themselves: §2.2 filters rationales that reveal hidden construction state, but every search query is synthesized from the target report and gold evidence partition, so the measured gains could depend on gold-targeting query patterns rather than…","rationale":"The reader's weakest assumption identifies the hidden-target oracle as the main risk; I agree that this is the right region of the argument. My concern sharpens it: the leak is not only in rationales that reveal construction state, which the paper filters, but in the queries themselves, which are generated with direct access to the target report and the gold evidence partition. The evidence-mismatch and raw-CPT controls are well designed but both keep the oracle-derived queries fixed, so they cannot rule out query-level leakage.\n\nI am not claiming the result is false. The paper has genuinely useful independent support: five matched subsets, paired intervals, a source-matched raw-CPT control, an evidence-mismatch control, transfer to three additional benchmarks, persistence through RL, and a scale transfer. Those are real evidence that the effect is not merely domain exposure or format imitation. The concern is a conditional one: the causal attribution to evidence-conditioned navigation would be much stronger if the query construction itself were de-oraclized. Because the paper's own reported evidence already supports transfer, and because the proposed test is an additional ablation rather than a demonstrated contradiction, I do not move the verdict. The reader's CONDITIONAL verdict already reflects that an independent check is needed; my analysis identifies which check is most decisive.","tokens_in":19943,"tokens_out":4559,"duration_ms":47049,"concrete_test":"Build a de-oraclized DRP-Web corpus by regenerating each query with the target report r and the gold partition P_i withheld: the query synthesizer sees only the proxy objective u, the causal history h_i, and the titles/leads of already opened pages. Render each search observation from the actual top-10 results returned by that query against a fixed offline snapshot of the source corpus (or a cached Serper index). Keep all rationales, open actions, and write targets unchanged. Then run the canonical 1B-token CPT and the 1/16- and 1/4-SFT comparisons on DeepResearch Bench. If the de-oraclized 1/4-SFT DRP-Web still exceeds the full-SFT Base checkpoint by the reported ~1.0 point with a 95% interval above zero, the oracle concern is not load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 concedes that the offline constructor \"uses the hidden target report as an oracle for creating the supervision.\" The stated filtering rule removes only rationales that reveal the construction process or unrevealed evidence; it does not remove the oracle information embedded in the queries. In Algorithm A1 and Appendix A.2, each query q_i is synthesized after seeing the target passage r and the gold partition P_i, so every query is engineered to retrieve exactly the evidence pages that will be opened. The model is therefore trained to produce search phrasings calibrated to a known answer set, not necessarily to discover unknown evidence in a live environment.\n\nThe load-bearing question is whether the low-data gain comes from a transferable search policy or from memorizing gold-targeting query patterns. The evidence-mismatch control in Section 4.2 keeps these oracle queries fixed and randomizes only the observations, so it cannot detect query-level oracle leakage. The raw-CPT control in Section 4.1 has the same blind spot. The positive transfer to ResearchQA, WebWalkerQA, and SimpleQA, and the persistence through RL, suggest that at least part of the effect is genuine. But the central causal claim—that DRP improves evidence-conditioned navigation rather than domain exposure or format imitation—depends on the queries transferring to a setting where the target report and gold evidence are unknown. This is the weakest link in the argument, and the paper does not directly test it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces Deep Research Pretraining (DRP), an offline method that converts citation networks and Wikipedia hyperlink structures into synthetic 'predictive navigation' trajectories (search-open-write) for continual pretraining, with the goal of improving the sample efficiency of subsequent supervised fine-tuning on deep-research agent trajectories. The authors continually pretrain Qwen3-14B-Base (and Qwen3-30B-A3B-Base) on 1B tokens of DRP data, then fine-tune on 1/16, 1/4, or full DR Tulu SFT data, evaluating on DeepResearch Bench (DRB) and transfer benchmarks. Across five matched SFT subsets, DRP-Paper and DRP-Web outperform direct fine-tuning at low data budgets; DRP-Web with 1/4 data also surpasses the fixed full-data Base checkpoint on DRB. The paper includes source-matched raw-CPT controls, an evidence-mismatch ablation, a repeated-evaluation stability check, and cross-scale results, and argues that the gains stem from evidence-conditioned navigation rather than domain exposure or format imitation.","tokens_in":20216,"tokens_out":5963,"duration_ms":52861,"significance":"The experimental design is a genuine strength: independent SFT subsets with paired t-intervals, paired-bootstrap question intervals, raw-domain CPT controls, an evidence-mismatch ablation, a repeated-evaluation stability check, and cross-scale transfer are all included. If the sample-efficiency gain is robust, DRP offers a cheap, scalable source of proxy navigation supervision for deep-research agents, and the paper's careful uncertainty reporting is exemplary. The main weakness is that the causal attribution to 'evidence-conditioned navigation' is not fully isolated from oracle-informed query synthesis: the ablations hold the queries fixed and vary only observations or domain text, so they cannot rule out the possibility that part of the gain comes from learning to imitate gold-targeting search patterns. This is a load-bearing concern for the paper's central claim, but it is addressable with an additional ablation or a suitably tempered claim.","major_comments":[{"comment":"The oracle leakage in query synthesis is not controlled. In Algorithm A1, each query q_i is generated by SynthesizeQuery(r, P_i, h), which receives the target report r and the gold evidence partition P_i; the filtering rule on line 21 removes only rationales that reveal the hidden construction process or unrevealed evidence, not the oracle signal embedded in the queries themselves. The raw-CPT control (Section 4.1) and the evidence-mismatch control (Section 4.2) both preserve these oracle-conditioned queries, so neither can distinguish a transferable evidence-conditioned navigation policy from a policy that has memorized gold-targeting query patterns. Because the abstract and conclusion attribute the gains to 'evidence-conditioned navigation,' this gap is load-bearing for the paper's main mechanistic claim. I request an ablation that re-synthesizes or perturbs queries without access to the gold report or evidence (e.g., conditioning query generation only on the history up to that point), or, failing that, a clear softening of the attribution claim to 'DRP improves trajectory-SFT sample efficiency' without the strong mechanistic conclusion.","section":"Section 2.2 and Algorithm A1"},{"comment":"Contamination is checked only for ResearchQA, not for the primary DeepResearch Bench evaluation. The DRP-Paper corpus is derived from arXiv source papers, and DRB is the headline benchmark; without an equivalent k-gram overlap analysis for the 50 DRB questions (and ideally their reference evidence pages), the possibility of benchmark contamination inflating the DRP-Paper DRB gains remains open. The transfer results and the DRP-Web improvements mitigate this concern, but the lack of a direct check for the main benchmark is a gap in an otherwise rigorous experimental report.","section":"Appendix A.7 / primary benchmark"}],"minor_comments":[{"comment":"The sentence 'These literature motivates both our pre-SFT intervention and our source-matched raw-CPT controls' contains a subject-verb agreement error; it should read 'This literature motivates'.","section":"Section 5 (Related Work)"},{"comment":"The claim that 'Longer context therefore provides no monotonic benefit in this experiment' is based on a single fixed SFT subset and a single pointing of the downstream budget; the sentence should explicitly say 'in this single-setting comparison' to avoid overgeneralization.","section":"Section 4.5"},{"comment":"The repeated-evaluation stability check is reported for only one full-SFT checkpoint, which is acknowledged; the text could add one sentence noting that a similar check at a low-data budget would strengthen the main sample-efficiency intervals, which currently mix subset variation with evaluation noise.","section":"Table A3 / Section B.3"},{"comment":"In Algorithm A1, the return of (τ, m) and the definition of m in step 22 are consistent, but the pseudocode for VerifyCite(y) = union of P_i is a step that is asserted rather than checked with a fallback; adding a brief note on how failures of this verification are handled (e.g., resampling or discarding) would improve reproducibility.","section":"Section A.1 / Algorithm A1"}],"recommendation":"major_revision","confidential_remarks":"The empirical sample-efficiency result seems solid, and the statistical care is commendable. My main concern is the oracle-query issue, which I believe is addressable; the authors should either add a query-level control or explicitly limit their attribution claim. I would not reject on this basis, but I would not accept as-is either, because the abstract and conclusion currently overstate the mechanistic conclusion. Also worth verifying: the paper cites a same-group prior work (WRAP++) for the hyperlink-motif construction; this is not a problem, but the novelty relative to that paper should be clearly scoped in the final version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know about this paper because it is the first, as far as I can tell, to turn citation and hyperlink graphs into offline search–open–write trajectories for pretraining deep-research agents, and the main sample-efficiency claim survives scrutiny. The experimental design is genuinely careful.\n\nWhat's new: DRP constructs proxy research objectives from naturally occurring evidence structures, renders them into history-conditioned navigation traces, and trains with a masked predictive-navigation objective. They then SFT on controlled fractions of 13K expert trajectories. The evaluation is unusually thorough for this area: five independently sampled matched subsets, paired t-intervals, paired-bootstrap question intervals, raw-domain CPT controls, an evidence-mismatch ablation, transfer to three external benchmarks, persistence through subsequent RL, general-capability preservation, and a standardized cost accounting. The central claim—that one quarter of the trajectories with DRP-Web beats full-data no-DRP SFT—is supported by the numbers and the intervals. This is real method work, not a repackaging.\n\nThe soft spot is load-bearing, and it is the one the stress-test flags. Section 2.2 concedes that the offline constructor uses the hidden target report as an oracle when creating supervision. The filtering rule removes rationales that reveal hidden construction state or unrevealed evidence, but every query is synthesized after seeing the target report and the gold evidence partition. So the training distribution contains search phrasings calibrated to a known answer set. The evidence-mismatch control in Section 4.2 keeps these oracle queries fixed and randomizes only the observations, so it cannot detect query-level leakage; the raw-CPT control has the same blind spot. The positive transfer to ResearchQA, WebWalkerQA, and SimpleQA, plus persistence through RL, strongly suggests at least part of the effect is genuine. But the paper's central causal claim—that the gain comes from evidence-conditioned navigation rather than gold-targeting query patterns—is not directly tested. That is the gap to close.\n\nTwo smaller issues: the primary benchmark has only 50 English questions, and the response-length budget was selected via an ablation on that same benchmark. Neither sinks the paper, but they temper how precisely the headline numbers should be read. I also wish the construction pipeline, corpus, and checkpoints were released; without them, independent reproduction is impossible.\n\nOverall, this is a solid, useful paper for anyone working on agent pretraining or synthetic data from document structure. It deserves serious peer review. The oracle concern does not kill it, but it should be addressed—ideally with an ablation that generates queries without access to the gold evidence, or an evaluation in a live search setting where the target report is unknown.","headline":"A genuinely careful paper on offline pretraining for deep-research agents; the sample-efficiency gain looks real, but the hidden-oracle queries leave the central causal claim not fully closed.","tokens_in":20744,"tokens_out":1691,"would_cite":true,"duration_ms":16113,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Paper citations and web links can pretrain deep-research navigation offline, so a model fine-tuned on a quarter of the expert trajectories beats the full-data model.","keywords":["deep research pretraining","predictive navigation","offline trajectory synthesis","citation graphs","hyperlink graphs","sample efficiency","continual pretraining","evidence-conditioned navigation"],"falsifier":"Generate the same DRP trajectories with every query and rationale produced by a generator that never sees the target report — the offline analogue of blind search — while keeping evidence partitioning, candidate sets, rendering budgets, and downstream fine-tuning identical. If the low-data gain over the base model collapses toward the level of the paper's raw-document control, then oracle knowledge of which documents are evidence, rather than an acquired navigation skill, is what the current numbers measure. A cheaper companion check: count how often DRP's synthesized queries reproduce the titles or key phrases of the target evidence documents; near-verbatim overlap would indicate the supervision leaks the answer into the question.","tokens_in":19744,"feed_emoji":"🧭","tokens_out":19370,"duration_ms":138759,"temperature":0.7,"pith_summary":"Deep research agents are usually trained on expensive, environment-grounded trajectories requiring repeated retrieval, document inspection, and report evaluation. This paper claims the same navigation skills can be acquired offline from evidence structures that already exist in document collections: citation edges in scholarly papers and hyperlinks in Wikipedia. The proposed method, Deep Research Pretraining (DRP), turns a citation-bearing or hyperlinked passage into a proxy research question, recovers the documents that support it, and renders a search–open–write trajectory without a live retrieval environment or any executed policy rollout. Continually pretraining a 14-billion-parameter base model on one billion tokens of such trajectories makes it consistently outperform its base model at one-sixteenth and one-quarter of the downstream expert-trajectory budget on the DeepResearch Bench evaluation, and with one quarter of the data the Wikipedia-based variant beats the no-DRP full-data model by 1.0 point. If the finding holds, expensive expert trajectories for research agents could be substantially replaced by cheap, static evidence graphs.","feed_headline":"Graph pretraining lets research agents beat full data with a quarter","feed_subtitle":"Practice from citation and hyperlink graphs lets a quarter of the expert data beat the full set.","key_machinery":"The central object is predictive navigation: conditioned on the research objective and the evidence observed so far, the model learns to predict what to search for, which returned documents to open or ignore, and how to synthesize the gathered evidence. The machinery that produces this supervision is offline reverse synthesis from an evidence graph — starting from a citation-bearing or hyperlinked passage as the target report, the constructor recovers the supporting evidence set, draws topologically nearby but unselected documents as candidate alternatives, and builds the trajectory backwards into interleaved search–open–write steps. Search observations are rendered as titles and snippets for the candidate set, open observations as source text, and the final write as the normalized report, all without a live retrieval environment or an executed policy rollout. Training uses a masked causal language-model loss that puts weight only on assistant reasoning, tool calls and arguments, and the final report, masking out user objectives and tool observations; the ablations show the masking is load-bearing, since removing it drops the evaluation score by 8.1 points and collapses downstream tool use.","core_discovery":"The paper's central claim is that the decisions a deep-research agent must make — what to search for, which documents to inspect, and how to synthesize evidence into a report — can be taught by a predictive-navigation objective trained entirely offline on naturally occurring evidence graphs. Given a target passage that cites or links to documents, DRP infers a proxy research objective, partitions the supporting evidence into rounds, fills each round's candidate set with the evidence plus topologically related but unselected alternatives, and renders the interleaved search–open–write trajectory with rationales conditioned only on the history so far. The model is continually pretrained on these trajectories with the language-model loss masked to assistant reasoning, tool calls, and the final report, so it learns evidence-conditioned decisions rather than imitation of environment responses. Across five independently sampled subsets at one-sixteenth and one-quarter of the 13,062 expert trajectories, both the citation-graph variant and the hyperlink variant beat matched no-DRP models on DeepResearch Bench; the hyperlink variant at one-quarter data exceeds the no-DRP full-data checkpoint by 1.0 point with its 95% interval above zero, and the low-data gains transfer to ResearchQA, WebWalkerQA, and SimpleQA while surviving 80 steps of downstream agentic reinforcement learning. Source-matched raw-text and evidence-mismatch controls are used to argue that the benefit comes from evidence-conditioned navigation rather than domain exposure or agent-format imitation.","pith_inferences":["A direct consequence of the paper's evidence-mismatch control that the authors do not spell out: DRP's value should degrade as the evidence graph becomes noisier or less topically structured, so rewiring hyperlink or citation edges into random co-occurrence graphs would be a cheap test of whether the stated mechanism rather than raw text exposure carries the gain.","The offline oracle leaves a deployment gap the current benchmarks cannot close: during construction the generator knows which documents are evidence, while at inference the agent must discover that itself, so a blind-search evaluation with queries generated without access to the target report would separate genuine navigation skill from question-side leakage.","The unmasked-observation failure (predicting tool outputs costs 8.1 points and cuts downstream tool calls from 5.7 to 0.6) suggests a general design rule for offline agent pretraining — supervise decisions and synthesis, never the environment's responses — that could transfer to other agent-pretraining pipelines the paper does not test."],"forward_implications":["With one quarter of the expert trajectories, a hyperlink-pretrained model beats the no-DRP full-data baseline by 1.0 point on DeepResearch Bench with its 95% interval above zero, so the same report quality can be reached with roughly 75% fewer expert demonstrations.","The low-data gains transfer beyond long-form report evaluation to ResearchQA coverage, WebWalkerQA accuracy, and SimpleQA accuracy, and the advantage over the base model persists through 80 steps of downstream agentic reinforcement learning from matched checkpoints.","Because construction renders supervision from static corpora and graph neighborhoods, no search engine, browser, or executed policy rollout is needed to generate training data; the paper's cost accounting estimates at least a 51.8% reduction in standardized data-construction cost for the one-quarter-SFT recipe.","The same 1B-token DRP stage helps on two structurally different evidence graphs and at two model scales (a 14B dense model and a 30B-A3B model), and replacing half of a general midtraining budget with DRP tokens preserves aggregate general capability while improving downstream agent initialization."],"supporting_citations":[{"why":"Supplies the 13,062 expert deep-research trajectories and the fine-tuning recipe on which the sample-efficiency comparisons are built.","marker":"Shao et al., 2025"},{"why":"Defines DeepResearch Bench, the primary evaluation whose RACE scores anchor all matched-SFT and full-data contrasts.","marker":"Du et al., 2025"},{"why":"Provides the unarXive 2024 scholarly corpus and citation graph from which DRP-Paper trajectories are constructed.","marker":"Besrour & Färber, 2026"},{"why":"Provides the English FineWiki corpus and hyperlink structure from which DRP-Web trajectories are constructed.","marker":"Penedo, 2025"},{"why":"Defines the Qwen3 base models (14B and 30B-A3B) that DRP continually pretrains and that anchor the model-scale results.","marker":"Yang et al., 2025"},{"why":"Provides the auxiliary generator used to infer proxy objectives, write rationales, and filter incidental links during offline construction.","marker":"Qwen Team, 2026"},{"why":"Defines the ARPO deep-search reinforcement-learning setting used to test whether the DRP advantage persists through downstream RL.","marker":"Dong et al., 2026"},{"why":"Supplies the dual-link and co-mention hyperlink motifs used to build DRP-Web candidate-alternative pools.","marker":"Zhou et al., 2026"}],"fun_headline_variants":["Graph pretraining lets a quarter of agent data beat the full set","Offline graph navigation pretraining beats full data with 25% of it","Predictive navigation from graphs outperforms full agent data at 1/4","Research agents learn from graphs and beat full data with a quarter","Graph-based pretraining: quarter data beats full, without live tools"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that supervision written by an oracle that already knows which documents are the evidence still teaches a policy that transfers to live search, where the model has to discover for itself which documents matter.","fun_headline_variants_meta":{"raw":{"variants":["Graph pretraining lets a quarter of agent data beat the full set","Offline graph navigation pretraining beats full data with 25% of it","Predictive navigation from graphs outperforms full agent data at 1/4","Research agents learn from graphs and beat full data with a quarter","Graph-based pretraining: quarter data beats full, without live tools"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1509,"prompt_tokens":1107,"completion_tokens":402,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":723,"completion_tokens_details":{"reasoning_tokens":308}},"tokens_in":723,"tokens_out":402,"duration_ms":3725,"temperature":1.0,"reasoning_tokens":308,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:19:42.365485+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate the same DRP trajectories with every query and rationale produced by a generator that never sees the target report — the offline analogue of blind search — while keeping evidence partitioning, candidate sets, rendering budgets, and downstream fine-tuning identical. If the low-data gain over the base model collapses toward the level of the paper's raw-document control, then oracle knowledge of which documents are evidence, rather than an acquired navigation skill, is what the current numbers measure. A cheaper companion check: count how often DRP's synthesized queries reproduce the titles or key phrases of the target evidence documents; near-verbatim overlap would indicate the supervision leaks the answer into the question.","supporting_citations":[{"cited_title":"2025 , eprint =","cited_arxiv_id":null,"evidence_quote":"Defines the Qwen3 base models (14B and 30B-A3B) that DRP continually pretrains and that anchor the model-scale results."},{"cited_title":"2026 , eprint =","cited_arxiv_id":null,"evidence_quote":"Defines the ARPO deep-search reinforcement-learning setting used to test whether the DRP advantage persists through downstream RL."}],"review_version":2}