Scaling Scientific Discovery Environments for Turn-Level Agentic RL

Scaling Scientific Discovery Environments for Turn-Level Agentic RL

Yucheng Xu † \dagger , Keyi Zhang † \dagger , Yuyang Yu , Min Zhang , Shiyuan Meng , Pei Chu , Zhongying Tu * Shanghai AI Lab

† \dagger Equal contribution

• Corresponding author

  Abstract Large language model agents have shown promising capabilities in data-driven scientific discovery tasks, where an agent interacts with an execution environment and produces a statistical claim. Long-horizon scientific analysis remains constrained by the lack of process-supervised environments over real-world scientific data. This paper introduces SciDisco , a scalable framework for training Sci entific Disco very agents in process-verifiable environments. SciThèque compiles hypotheses, datasets, hidden evidence graphs, and verifiers into task environments where analytical progress can be checked during interaction. DAG-grounded trajectory synthesis uses these environments to construct verifier-filtered multi-turn demonstrations. DiscoPO then uses the environment as the source of training signal, assigning turn-level credit to actions that produce verifiable analytical evidence. Experiments show that SciDisco-14B reaches state-of-the-art on hypothesis-driven scientific data analysis benchmarks.

  Scaling Scientific Discovery Environments for Turn-Level Agentic RL

  Yucheng Xu † \dagger , Keyi Zhang † \dagger , Yuyang Yu, Min Zhang, Shiyuan Meng , Pei Chu , Zhongying Tu * Shanghai AI Lab

  † \dagger Equal contribution

• Corresponding author

  1 Introduction

  Data-driven scientific discovery is a hypothesis-driven mode of empirical research over existing scientific datasets. In human practice, it requires researchers to understand scientific context and dataset structure, select appropriate methods, execute statistical analyses, and draw verifiable conclusions. Existing LLMs and LLM-based agents remain unreliable in this setting. Prior work shows that models struggle with statistical method applicability and data-based causal reasoning (Zhu et al. , 2024 ; Liu et al. , 2024a ) . Agentic code and reasoning scaffolds also achieve limited performance on data-driven discovery benchmarks (Majumder et al. , 2025 ; Gu et al. , 2024 ; Chen et al. , 2025 ) . These results suggest that data-driven scientific discovery should not be treated as one-shot question answering or single-turn code generation, but as an interactive process of executing, observing, revising, and verifying analyses over real data.

  The core challenge is to train agents that learn robust interaction policies, rather than merely encode scientific knowledge or generate code. Prompts and hand-written scaffolds can describe the steps of an analysis workflow, but benchmark evidence suggests that prompting and generic multi-turn training do not reliably produce goal-directed interaction. In contrast, reward-trained tool use can outperform supervised tool traces in unfamiliar tool settings (Abdulhai et al. , 2023 ; Qian et al. , 2025 ; Wang et al. , 2025b ) . Data-driven discovery agents must therefore learn from environmental feedback about when to inspect data, choose methods, test assumptions, revise analyses, and submit conclusions. This motivates training specialized discovery agents inside scientific data environments, rather than relying only on foundation models or inference-time workflow orchestration.

  Such training requires environments that are both executable and verifiable. A discovery RL environment should provide data-driven goals, stateful code execution, observations, and automatic verification. Open scientific datasets are abundant, but they usually remain static files or metadata without task goals or verifiers. Existing scientific-discovery benchmarks provide valuable evaluation tasks and ground-truth criteria, but they do not expose hidden scientific evidence states for large-scale process-supervised training with persistent cross-turn state and verifier-grounded turn credit (Majumder et al. , 2025 ; Gu et al. , 2024 ; Chen et al. , 2025 ) . General agent and ML engineering environments support executable feedback and, in some cases, training, but their objectives target web interaction or ML engineering rather than real-data hypothesis verification (Liu et al. , 2024b ; Chan et al. , 2024 ; Qiang et al. , 2025 ; Zhou et al. , 2026 ) .

  Executable environments alone are still not enough for effective RL. Without a behavioral prior, early rollouts can be dominated by invalid tool calls, variable misreadings, and incomplete analyses. Data-driven scientific discovery also poses a long-horizon credit-assignment problem: final conclusions depend on schema understanding, method selection, code execution, diagnostics, and evidence interpretation, while outcome-only rewards cannot identify which turns produced valid analytical evidence (Wang et al. , 2025b ; Wei et al. , 2025 ; Li et al. , 2026b ; Xie et al. , 2026 ) .

  These limitations point to a gap in agent training: data-driven scientific discovery needs an interface in which analytical progress can be executed, verified, and credited during interaction. SciDisco provides such a framework for data-driven scientific discovery agents. Its environment layer, SciThèque , compiles open scientific datasets into verifiable hypothesis-driven environments. Hidden evidence DAGs support verified multi-turn demonstrations for SFT and provide process feedback that DiscoPO uses for turn-level credit assignment in agentic RL. SciDisco therefore aligns task construction, behavior imitation, and policy optimization within realistic scientific data environments.

  The paper makes three contributions:

  1. SciThèque builds scalable discovery environments from open scientific datasets with hypotheses, sandboxed execution, and verifiers.

  2. DAG-grounded trajectory synthesis cold-starts agent behavior using verified multi-turn demonstrations.

  3. DiscoPO assigns turn-level credit to verified evidence-producing processes instead of only final answers.

  2 Related Work

  2.1 Discovery Agents and Environments

  Prior work on data-science and scientific-discovery agents follows two broad paradigms. The first improves data-specific capabilities, including tabular understanding, database operations, statistical reasoning, and quantitative reasoning over data (Zhu et al. , 2024 ; Liu et al. , 2024a ) . The second moves beyond single-turn QA by building agentic workflows with ReAct-style interaction, code interpreters, multi-agent scaffolds, or learned analysis trajectories. These systems often rely on strong closed-source models, leaving the underlying training environment difficult to reuse (Guo et al. , 2024 ; Zhang et al. , 2023 ; Hong et al. , 2024 ; Li et al. , 2024 ; Zhang et al. , 2025b ; Qiao et al. , 2025 ) . This makes it difficult to separate model capability from the executable interaction and verification substrate used to produce scientific conclusions.

  Discovery benchmarks evaluate agents on real data, open-ended questions, code execution, and conclusion checking (Majumder et al. , 2025 ; Gu et al. , 2024 ; Chen et al. , 2025 ; Egg et al. , 2025 ) . Interactive science environments support exploration in text-based or virtual worlds (Wang et al. , 2022 ; O’Sullivan et al. , 2024 ) . General agent and MLE environments cover tool use, web interaction, executable feedback, and training efficiency (Liu et al. , 2024b ; Zhou et al. , 2024 ; Chan et al. , 2024 ; Qiang et al. , 2025 ; Zhou et al. , 2026 ) . These settings are valuable, but they are not built around real-data hypothesis verification with persistent execution state and verifier-grounded process feedback. SciThèque constructs executable, verifiable discovery environments that connect data, hypotheses, code execution, and reusable multi-turn feedback, so the same environment can define tasks, filter trajectories, and provide RL signals. This shifts the role of an environment from a static evaluation artifact to a reusable training interface for scientific agent learning.

  2.2 Reinforcement Learning for LLM Agents

  Beyond prompting and SFT, recent work trains LLM agents with reinforcement learning. GRPO-style RLVR uses group-relative advantages to avoid explicit critics and train models on verifiable reasoning or coding tasks (Shao et al. , 2024 ; Guo et al. , 2025a ) . Later variants improve sampling, clipping, normalization, length bias, and self-distillation routing (Yu et al. , 2025 ; Liu et al. , 2025 ; Li et al. , 2026a ) . These methods provide a strong optimization backbone for agent training, but their rewards usually remain at the response or trajectory level.

  For multi-turn agents, recent work assigns credit at step, turn, or segment granularity, or calibrates intermediate decisions with hindsight, reflection, or trajectory correction (Feng et al. , 2025 ; Wei et al. , 2025 ; Li et al. , 2026b ; Zong et al. , 2026 ; Guo et al. , 2025b ; Lu et al. , 2026 ; Wang et al. , 2025a , c ) . These methods recognize that final rewards cannot identify which intermediate actions advanced the task. However, their signals often depend on repeated states, tree search, segmenters, hindsight models, reflection, or task-specific step rewards, rather than evidence produced by executed scientific analyses. DiscoPO instead uses scientific verifiers to credit turns that produce validated analytical evidence, aligning turn-level RL with hypothesis verification.

  3 Problem Formulation

  A scientific discovery task is formulated as a hypothesis-driven data analysis problem. Given a dataset, the task is to evaluate a scientific hypothesis, produce statistical findings, and support those findings with executable analysis.

  Let 𝒰 = { u j } j = 1 N \mathcal{U}={u{j}}{j=1}^{N} denote a collection of such tasks. Each task is represented as:

  u j = ( h j , d j , v j ) , u{j}=(h{j},d{j},v{j}), (1)

  where h j h{j} specifies the hypothesis or analysis goal, d j d{j} is the dataset associated with the task, and v j v_{j} denotes the task verifier.

  Each task u j u{j} is paired with an environment e j e{j} , and the corresponding environment collection is ℰ = { e j } j = 1 N \mathcal{E}={e{j}}{j=1}^{N} . The environment exposes h j h{j} and d j d{j} , executes the agent’s actions in a sandbox, and returns observations, while task-internal verification state may remain hidden. The interaction between policy π θ \pi{\theta} and environment e j e{j} can therefore be modeled as a partially observable Markov decision process (POMDP), producing a trajectory τ j \tau_{j} :

  τ j = ( o j , 0 , a j , 1 , o j , 1 , … , a j , T j , o j , T j ) , \tau{j}=(o{j,0},a{j,1},o{j,1},\ldots,a{j,T{j}},o{j,T{j}}), (2)

  where o j , 0 o{j,0} contains the initial task observation, actions a j , t a{j,t} can be executable analysis or terminal submission, and observations o j , t o_{j,t} contain environment feedback.

  A final reward provides only sparse credit assignment for a scientific discovery trajectory. It cannot distinguish an evidence-building trajectory from a one-shot shortcut that happens to produce the same submitted findings.

  Let r j , t r_{j,t} denote the process reward assigned to turn t t , measuring whether that turn contributes verifiable progress toward the task findings. The training goal is to maximize the accumulated process reward:

  R j ​ ( τ j ) = ∑ t = 1 T j r j , t , R{j}(\tau{j})=\sum{t=1}^{T{j}}r_{j,t}, (3)

  and the optimization objective is:

  J ​ ( θ ) = E ​ [ R j ​ ( τ j ) ] , J(\theta)=\mathrm{E}\left[R{j}(\tau{j})\right], (4)

  where the expectation is over tasks sampled from 𝒰 \mathcal{U} and trajectories generated by π θ \pi_{\theta} in their paired environments.

  4 Methodology

  SciDisco turns the process-reward formulation in Section 3 into a three-stage post-training pipeline for scientific discovery agents. SciThèque compiles scientific datasets into sandboxed task environments with task-local hidden evidence DAGs. DAG-grounded trajectory synthesis uses these environments to produce verifiable multi-turn demonstrations for SFT cold-start. DiscoPO then optimizes the policy with turn-level credit from environment-verified scientific progress. The task-local hidden evidence DAG is the shared interface across these stages: it defines what can be synthesized, what counts as an accepted transition during online interaction, and how progress is credited during RL.

  Figure 1: The pipeline of SciDisco : 1) SciThèque builds a scalable collection of data–hypothesis–verifier environments from scientific datasets; 2) Trajectory Synthesis traverses the evidence DAG through executable analysis primitives and keeps accepted state transitions as multi-turn SFT demonstrations; 3) DiscoPO defines turn-level rewards by increases in verified DAG progress, assigning credit to evidence-producing turns.

  4.1 SciThèque

  SciThèque instantiates the environment collection ℰ \mathcal{E} defined in Section 3 . It constructs tasks u j = ( h j , d j , v j ) u{j}=(h{j},d{j},v{j}) from source datasets and hypothesis templates, then mounts each task in a sandbox runtime to form an environment e j e{j} . The agent sees h j h{j} , data files d j d{j} , and observations o j , t o{j,t} , while the verifier logic, evidence DAG, and progress state remain hidden. Appendix A reports the data-source distribution and materialized environment contract.

  Catalog construction. SciThèque first builds candidate dataset entries c i = ( s i , d i , m i , ρ i ) c{i}=(s{i},d{i},m{i},\rho{i}) , where s i s{i} is the source, d i d{i} the dataset, m i m{i} the modality, and ρ i \rho_{i} a structural profile recording provenance, licence metadata, file structure, size, and analysis-relevant fields. Inexpensive gates remove unsuitable datasets before task creation, and each surviving dataset is classified as tabular, temporal, graph, spatial, or sequence data. Appendix A reports source citations and snapshot counts.

  Template-guided hypothesis generation. For each modality m m , SciThèque maintains a template library 𝒯 m \mathcal{T}{m} indexed by statistical pattern and outcome type. Each catalog entry is matched to templates allowed by ρ i \rho{i} , and an external LLM proposes constrained specifications ℋ i = { ( T i , k , η i , k ) } k = 1 K i \mathcal{H}{i}={(T{i,k},\eta{i,k})}{k=1}^{K{i}} , where T i , k ∈ 𝒯 m i T{i,k}\in\mathcal{T}{m{i}} is an applicable template and η i , k \eta{i,k} instantiates it on d i d{i} . The LLM proposes candidates but does not certify validity; retained specifications from all ℋ i \mathcal{H}_{i} form the materialization pool.

  Environment materialization. Materialization turns each retained specification into a training environment. SciThèque runs the reference analysis on d i d{i} under the selected template and accepts the specification only if the analysis executes and yields verifier-backed findings. The accepted specification becomes u j = ( h j , d j , v j ) u{j}=(h{j},d{j},v{j}) , then is paired with sandbox runtime 𝒮 j \mathcal{S}{j} to form e j = ( u j , 𝒮 j ) e{j}=(u{j},\mathcal{S}_{j}) .

  The verifier v j v_{j} is grounded in a hidden evidence DAG:

  g j = ( 𝒱 j , 𝒜 j ) , g{j}=(\mathcal{V}{j},\mathcal{A}_{j}), (5)

  Each node in 𝒱 j \mathcal{V}{j} defines a verifiable scientific state transition and is typed by a process primitive, such as data inspection, model fitting, diagnostic checking, or robustness analysis. The primitive type determines the admissible family of analysis operations for that node, so a turn is verified against the scientific role of the transition rather than arbitrary code execution. The directed edges 𝒜 j \mathcal{A}{j} encode prerequisite relations, so only frontier nodes are eligible at a given turn. The DAG is hidden from the agent and defines the progress state used by v j v_{j} to verify intermediate work. Appendix B details the visibility split, evidence-graph record, progress rule, and leakage controls.

  To assess whether the verifier design matches domain-level scientific practice, 10 domain experts covering the scientific domains represented in the environment corpus manually reviewed the evidence primitives, prerequisite structure, and acceptance criteria used by the hidden evidence graphs. Appendix B reports the review protocol and audit scope.

  The materialized output is therefore an environment specification, not a static question-answer pair. Each task stores its data files, instructions, verifier, task-local evidence DAG, final checker, and reproducibility metadata.

  4.2 DAG-Grounded Trajectory Synthesis

  DAG-grounded trajectory synthesis uses SciThèque environments to construct SFT demonstrations for model cold-start. Although the evidence DAG defines verifiable scientific state transitions, a base policy may not know how to realize them as executable analysis turns. SFT therefore places the policy in the right interaction regime before RL begins. Appendix C reports the synthesis stages, filtering criteria, and retained corpus snapshot.

  For each materialized task u j u{j} with hidden evidence DAG g j g{j} , the synthesizer constructs a private transition schedule σ j , ℓ \sigma{j,\ell} over the required DAG nodes. This schedule specifies the verified transition executed at each step while respecting the prerequisite structure of g j g{j} . For each scheduled node, the synthesizer selects an executable program variant consistent with the node’s process primitive. The schedule and program choices remain hidden. The transcript exposes only assistant reasoning , code actions , sandbox observations , and a final action.

  Each synthesized trajectory uses a fresh sandbox runtime, where the scheduled actions are executed. After each execution, the verifier v j v{j} checks whether the visible observation establishes the scheduled transition in g j g{j} . A trajectory is rejected if the code fails, produces no verifiable evidence, matches the wrong frontier node, violates the primitive process constraint, or accesses hidden verifier files. Each retained SFT turn therefore corresponds to an accepted scientific state transition rather than an arbitrary code fragment.

  After all scheduled transitions are accepted, synthesis closes the transcript with a action. The submission records findings already established through verified transitions rather than introducing a separate terminal reward. Let 𝒯 syn \mathcal{T}{\mathrm{syn}} be the pool of candidate trajectories produced by synthesis, and let τ j , ℓ \tau{j,\ell} denote the ℓ \ell -th candidate trajectory for task u j u{j} . A candidate is retained in 𝒟 SFT \mathcal{D}{\mathrm{SFT}} only when the required nodes in g j g_{j} have been completed and the submitted findings are consistent with the accumulated verifier state:

  𝒟 SFT = { τ j , ℓ ∈ 𝒯 syn ∣ A j ​ ( τ j , ℓ , g j ) = 1 } . \mathcal{D}{\mathrm{SFT}}={\tau{j,\ell}\in\mathcal{T}{\mathrm{syn}}\mid A{j}(\tau{j,\ell},g{j})=1}. (6)

  Here A j ​ ( τ j , ℓ , g j ) = 1 A{j}(\tau{j,\ell},g_{j})=1 indicates that the trajectory completes the required DAG transitions in a valid order and submits the findings verified during those transitions.

  For the same task, synthesis can retain several verifier-valid trajectories rather than a single canonical solution path. These trajectories may differ in transition schedule, executable program variant, finding-reporting order, and reasoning style, while sharing the same hidden DAG and verifier constraints. Training-data conversion removes scheduler labels, hidden node identifiers, reference values, verifier tolerances, and generator-only helper variables. This cold-start stage teaches the policy to reason through a scientific state-transition chain. DiscoPO can then optimize verified scientific progress from a policy that already follows the basic scientific behavior.

  4.3 DiscoPO

  DiscoPO trains the policy in SciThèque environments. At each turn within a trajectory, the agent emits either a cell or a action. The sandbox presents each task as an agent workspace, and its persistent kernel lets later turns reuse earlier variables, models, and files. Each observation provides evidence for verifying the next state transition.

  GRPO computes a group-normalized advantage from trajectory-level rewards and broadcasts that advantage to all generated tokens of the trajectory (Shao et al. , 2024 ) . DiscoPO keeps the same group-relative normalization but computes the advantage from verifier-gated progress at each interaction turn.

  Algorithm 1 DiscoPO

  1: task environments { e j } {e{j}} , hidden DAGs { g j } {g{j}} , old policy π old \pi{\mathrm{old}} , group size G G , clip parameters ϵ low \epsilon{\mathrm{low}} , ϵ high \epsilon_{\mathrm{high}}

  2: for each training iteration do

  3: sample a batch of task environments

  4: for each sampled environment e j e_{j} do

  5: sample G G rollouts { τ j , k } k = 1 G {\tau{j,k}}{k=1}^{G} with π old \pi_{\mathrm{old}}

  6: for each rollout k = 1 , … , G k=1,\ldots,G do

  7: initialize C j , k , 0 ← ∅ C_{j,k,0}\leftarrow\emptyset

  8: for each turn t t in τ j , k \tau_{j,k} do

  9: use v j v{j} to update C j , k , t + 1 C{j,k,t+1} from C j , k , t C_{j,k,t}

  10: compute r j , k , t r_{j,k,t} by Eq. 8

  11: end for

  12: end for

  13: compute b j , t b{j,t} and A j , k , t A{j,k,t} by Eqs. 11 and 12

  14: assign A j , k , t A_{j,k,t} to generated tokens in turn t t and mask observations

  15: end for

  16: update π θ \pi_{\theta} with the clipped surrogate over turn-level advantages

  17: set π old ← π θ \pi{\mathrm{old}}\leftarrow\pi{\theta}

  18: end for

  Verifier-gated state transitions. For task u j u{j} , let g j = ( 𝒱 j , 𝒜 j ) g{j}=(\mathcal{V}{j},\mathcal{A}{j}) be the hidden evidence DAG, and let 𝒱 j exec ⊆ 𝒱 j \mathcal{V}^{\mathrm{exec}}{j}\subseteq\mathcal{V}{j} denote the required DAG nodes verified through turns. For rollout k k of task j j , C j , k , t C_{j,k,t} records the DAG nodes credited before turn t t . Here j j follows Section 4.1 and indexes the materialized task, while k k indexes one trajectory in the rollout group used for group-relative advantage. Appendix F shows an example rollout trajectory.

  At each turn, the verifier v j v_{j} considers only frontier nodes, meaning uncredited nodes whose prerequisites have already been completed. The verifier then checks whether the current turn matches the process primitive assigned to one frontier node. A turn is accepted only when the observation provides evidence for exactly one such node and the executed code satisfies its node-local process check.

  Turns are rejected when they provide no observable evidence, match no frontier node, match multiple frontier nodes, repeat completed work, or collapse several analyses into one action. If a turn is accepted, the matched node is added to the hidden completed-node state; otherwise, the state is left unchanged. The action is checked against the verifier and can be accepted only after the required nodes have been credited. Appendix G gives concrete rejection categories and their training consequences.

  Credit assignment. For task u j u{j} , let N j = | 𝒱 j exec | N{j}=|\mathcal{V}^{\mathrm{exec}}{j}| be the number of required executable primitive transitions, and let s j sub s^{\mathrm{sub}}{j} denote the terminal submit node. The set of progress nodes is 𝒫 j = 𝒱 j exec ∪ { s j sub } \mathcal{P}{j}=\mathcal{V}^{\mathrm{exec}}{j}\cup{s^{\mathrm{sub}}_{j}} . DiscoPO defines the DAG-progress potential over credited-node sets as:

  Φ j ​ ( C ) = | C ∩ 𝒫 j | | 𝒫 j | . \Phi{j}(C)=\frac{|C\cap\mathcal{P}{j}|}{|\mathcal{P}_{j}|}. (7)

  The per-turn process reward is the potential difference:

  r j , k , t = Φ j ​ ( C j , k , t + 1 ) − Φ j ​ ( C j , k , t ) . r{j,k,t}=\Phi{j}(C{j,k,t+1})-\Phi{j}(C_{j,k,t}). (8)

  When a turn is accepted, it adds one previously uncredited primitive node. A valid action adds the terminal submit node once all required primitives have been completed. Turns that do not add a new progress node receive zero reward. Equivalently:

  r j , k , t = { 1 N j + 1 , accepted primitive/submit, 0 , otherwise. r{j,k,t}=\left{\begin{array}[]{ll}\frac{1}{N{j}+1},&\mbox{accepted primitive/submit,}\ 0,&\mbox{otherwise.}\end{array}\right. (9)

  For each task environment, a group of G G rollouts { τ j , k } k = 1 G {\tau{j,k}}{k=1}^{G} is sampled from the old policy π old \pi{\mathrm{old}} . Let T j , k T{j,k} be the length of rollout k k . Rewards from terminated rollouts are zero-padded when computing the same-turn baseline:

  r ~ j , k , t = { r j , k , t , t ≤ T j , k , 0 , t > T j , k . \widetilde{r}{j,k,t}=\left{\begin{array}[]{ll}r{j,k,t},&t\leq T{j,k},\ 0,&t>T{j,k}.\end{array}\right. (10)

  This same-turn baseline compares whether rollouts for the same task make verifier progress at interaction index t t ; it does not require different rollouts to complete the same DAG node at that index. DiscoPO then applies turn group normalization. The same-turn baseline is:

  b j , t = 1 G ​ ∑ k = 1 G r ~ j , k , t . b{j,t}=\frac{1}{G}\sum{k=1}^{G}\widetilde{r}_{j,k,t}. (11)

  The turn advantage is:

  A j , k , t = r ~ j , k , t − b j , t . A{j,k,t}=\widetilde{r}{j,k,t}-b_{j,t}. (12)

  Model DiscoveryBench DABStep DataSciBench

  Average HMS Easy Hard Average Success Rate Completion Rate

  Proprietary Models

  GPT-4o 13.2 66.7 6.1 15.8 66.3 68.4

  GPT-5 Mini 26.9 68.1 15.3 23.8 45.9 63.4

  Claude-Sonnet-4 23.2 81.9 19.8 29.7 67.6 76.6

  Open-source Models

  DeepSeek-V4-Flash 28.3 79.2 12.2 22.9 60.8 72.1

  Intern-S1-Pro 34.7 65.3 4.8 14.5 59.5 67.6

  Qwen3-8B 11.3 47.2 3.2 10.2 31.5 48.4

  Qwen3-14B 20.0 55.6 4.2 12.4 32.4 45.9

  Qwen3-32B 24.1 59.7 4.5 13.3 36.0 52.4

  Related Methods

  DeepAnalyze-8B 25.9 70.8 32.8 38.9 46.8 57.3

  AutoSDT-Coder-14B 7.3 56.9 2.7 11.3 1.0 1.3

  DataMind-14B 25.5 68.1 3.2 13.6 20.3 24.8

  SciDisco-14B 35.2 62.5 9.3 17.8 56.2 61.0

  Table 1: Main results on DiscoveryBench, DABStep, and DataSciBench. Performance is reported as percentages. The best results for each model group are highlighted in bold.

  Policy optimization. The policy update assigns A j , k , t A{j,k,t} to policy-generated tokens in turn t t , masks observation tokens, and applies a PPO clipped surrogate at turn granularity (Schulman et al. , 2017 ) . Let ℐ j , k , t \mathcal{I}{j,k,t} denote the generated-token positions in that turn, let ρ i ​ ( θ ) \rho{i}(\theta) be the old-policy likelihood ratio for token i i , and let ρ ¯ i ​ ( θ ) \bar{\rho}{i}(\theta) be its clipped value. DiscoPO optimizes:

  ℒ DiscoPO ​ ( θ ) \displaystyle\mathcal{L}{\mathrm{DiscoPO}}(\theta) = − E j , k , t ​ [ 1 | ℐ j , k , t | ​ ∑ i ∈ ℐ j , k , t s i ​ ( θ ) ] , \displaystyle=-\mathrm{E}{j,k,t}\left[\frac{1}{|\mathcal{I}{j,k,t}|}\sum{i\in\mathcal{I}{j,k,t}}s{i}(\theta)\right], (13)

  s i ​ ( θ ) \displaystyle s{i}(\theta) = min ⁡ ( ρ i ​ ( θ ) ​ A j , k , t , ρ ¯ i ​ ( θ ) ​ A j , k , t ) . \displaystyle=\min\left(\rho{i}(\theta)A{j,k,t},\bar{\rho}{i}(\theta)A_{j,k,t}\right).

  The optimizer is standard; the change is that verifier-gated DAG progress supplies turn-level token weights instead of a scalar trajectory reward. Groups whose turn advantages are identically zero are filtered before optimization because they provide no group-relative policy-gradient signal. Appendix E expands the token-level notation.

  5 Experiments

  The experiments evaluate whether SciDisco improves data-driven scientific discovery and executable scientific data analysis. The evaluation covers three benchmark regimes: hypothesis generation over scientific datasets, multi-step business data analysis, and programmatically checked data-science workflows.

  5.1 Benchmarks

  DiscoveryBench. DiscoveryBench tests whether an agent can produce data-driven scientific hypotheses from a dataset and a discovery goal (Majumder et al. , 2025 ) . Table 1 reports the average Hypothesis Match Score (HMS), which aligns predicted and gold sub-hypotheses along context, variable, and relationship dimensions. GPT-5 Mini was used as the judge model during evaluation.

  DABStep. DABStep contains 450 realistic multi-step data-analysis tasks over business datasets, split into 72 easy and 378 hard cases (Egg et al. , 2025 ) . Its tasks emphasise business data analysis, where an agent completes operational analyses over enterprise-like tables, reports, visualisations, or decision-support queries. Table 1 reports easy accuracy, hard accuracy, and the task-count-weighted average.

  DataSciBench. DataSciBench evaluates end-to-end data-science workflows with programmatic Task–Function–Code checks over data preparation, analysis, modelling, visualisation, and insight tasks (Zhang et al. , 2025a ) . Table 1 reports the benchmark-level success rate and completion rate, while Table 2 further uses the DataSciBench breakdown to report fine-grained F1–F5 scores for data preparation, plot validity, data exploration, data visualization, and data modeling.

  Setting Coarse-grained Metrics Fine-grained Metrics

  Success Rate

  Completion Rate

  VLM

  F1: Data Preparation

  F2: Plot Validity

  F3: Data Exploration

  F4: Data Visualization

  F5: Data Modeling

  Qwen3-14B 32.4 45.9 / 16.1 41.1 29.3 23.6 22.3

• SFT 46.8 53.3 / 37.9 51.8 32.8 50.9 23.4

• SFT + GRPO 49.5 55.4 / 39.8 53.6 34.5 52.7 26.3

• SFT + DiscoPO 56.3 61.0 / 43.0 57.1 41.4 56.4 31.6

  Table 2: Ablation study on SciDisco training objectives. Performance is evaluated on DataSciBench. Success Rate reports pass rate, Completion Rate reports completion accuracy, and F1–F5 measure fine-grained performance across data-science sub-tasks. VLM was left empty because the backbone model lacks vision capability.

  5.2 Models and Baselines

  We organize baselines into three groups to separate general model capability from data-science-specific training. The proprietary group includes GPT-4o (OpenAI, 2024 ) , GPT-5 Mini (OpenAI, 2025 ) , and Claude Sonnet 4 (Anthropic, 2025 ) , representing strong closed-source agentic models. The open-source group includes DeepSeek-V4-Flash (DeepSeek-AI, 2026 ) , Intern-S1-Pro (Intern-S1-Pro Team, 2026 ) , and Qwen3-8B/14B/32B (Yang et al. , 2025 ) , allowing comparison against publicly available general models at different scales. The related-method group includes DeepAnalyze-8B (Zhang et al. , 2025b ) , AutoSDT-Coder-14B (Li et al. , 2025 ) , and DataMind-14B (Qiao et al. , 2025 ) , which are specialized for data analysis or scientific discovery. SciDisco-14B is trained from Qwen3-14B using the SciDisco trajectory corpus and DiscoPO objective.

  5.3 Experimental Setup

  Starting from Qwen3-14B, we first apply cold-start SFT on SciDisco-produced verified trajectories and then perform DiscoPO agentic RL. Training uses the Slime post-training framework (THUDM, 2025 ) , with SGLang for rollout generation and inference (Zheng et al. , 2024 ) . For the RL stage, we use a rollout batch size of 16 and a group size of 8. The resulting SciDisco-14B checkpoint is evaluated on the three benchmarks above. All training and inference are conducted on NVIDIA A100 GPUs; detailed hyperparameters, source licenses, and corpus distribution are reported in Appendix A and Appendix D .

  5.4 Main Results

  Table 1 compares proprietary, open-source, and specialised data-science agents across three scientific data-analysis benchmarks. SciDisco-14B obtains the strongest DiscoveryBench result among all listed models, reaching 35.2% average HMS. This score exceeds the strongest proprietary baseline, GPT-5 Mini at 26.9%, the strongest open-source baseline, Intern-S1-Pro at 34.7%, and the strongest related-method baseline, DeepAnalyze-8B at 25.9%.

  SciDisco-14B also performs strongly on DataSciBench. It reaches 56.2% Success Rate and 61.0% Completion Rate, outperforming the related-method baselines on both metrics. The result remains below the strongest proprietary and open-source models on DataSciBench, but closes much of the gap while using a 14B open-weight backbone.

  DABStep shows a different pattern. SciDisco-14B reaches 17.8% average accuracy, below DeepAnalyze-8B at 38.9%, Claude-Sonnet-4 at 29.7%, and GPT-5 Mini at 23.8%. The hard split remains challenging for all Qwen3-family models, with Qwen3-8B, Qwen3-14B, and Qwen3-32B all below 5%. SciDisco improves over the Qwen3-14B backbone on DABStep, but does not match methods specialised for this benchmark regime.

  5.5 Ablation Studies

  Table 2 isolates the effect of the training objective on the 14B backbone using DataSciBench. SFT provides the main first-stage gain over Qwen3-14B, and adding GRPO gives a further but smaller improvement. The full SFT + DiscoPO setting obtains the strongest coarse-grained result, reaching 56.3% Success Rate and 61.0% Completion Rate.

  The fine-grained metrics follow the same ordering. SFT + DiscoPO obtains the highest score on every reported sub-task, with the largest incremental gains over SFT + GRPO appearing in Data Modeling and Data Exploration. These two categories are the clearest ablation signal for the turn-level objective, while the detailed mechanism is analysed in Section 5.6 .

  5.6 Analysis

  The results support a specific reading of SciDisco’s contribution. DiscoveryBench is closest to the paper’s data-driven scientific discovery setting because it evaluates hypothesis construction over scientific datasets. SciDisco-14B’s leading HMS score indicates that training on process-verifiable scientific environments improves the model’s ability to form benchmark-matched scientific hypotheses from data.

  DataSciBench provides a complementary executable analysis test. The gains over related methods and the ablation improvements in Table 2 indicate that SciDisco also improves code-mediated data analysis under programmatic checks. The strongest ablation gains appear in Data Exploration and Data Modeling, the two categories most directly tied to intermediate analytical decisions.

  DABStep exposes the boundary of the current training distribution. Its business data-analysis tasks often require completing operational analyses, reports, visualisations, or decision-support queries over enterprise-like tables. SciDisco instead trains on hypothesis-driven scientific environments where progress is defined by verifier-grounded evidence for statistical claims. The weaker DABStep result therefore suggests that the current environment distribution transfers better to scientific hypothesis analysis and executable scientific data analysis than to all forms of multi-step business analytics.

  The ablation results locate where the training signal enters. SFT gives the largest first-stage gain by exposing the model to verifier-filtered executable trajectories. GRPO adds a smaller improvement over SFT, indicating that trajectory-level outcome optimisation supplies useful reinforcement. SFT + DiscoPO gives the strongest result because the reward is assigned to turns that complete hidden evidence-state transitions, rather than only to trajectories that end with a correct answer.

  6 Conclusion

  We presented SciDisco, a framework for training data-driven scientific discovery agents in process-verifiable environments. Hidden evidence DAGs and verifier checks make intermediate analytical progress trainable, while SciTh‘eque compiles scientific datasets and hypothesis templates into sandboxed environments shared by DAG-grounded synthesis and DiscoPO. Evaluation results support SciDisco as a scalable pipeline for converting verified analytical evidence into process supervision.

  Limitations

  SciDisco focuses on executable data-analysis workflows over text-readable scientific data, rather than wet-lab discovery, multimodal scientific artifacts, or open-ended hypothesis generation beyond the template space. The coverage of the environment distribution is also limited by the available source adapters, modality handlers, and hypothesis templates. Finally, environment-accepted progress is still a proxy for scientific quality: it verifies whether an analysis step was valid for the task contract, but it does not by itself establish that the agent made a novel scientific discovery.