InternAgent-1.5: Unified Scientific Discovery

• InternAgent-1.5 is a unified agentic framework that orchestrates hypothesis generation, verification, and evolution to support long-horizon scientific discovery.
• It leverages Graph-Augmented Monte Carlo Search and structured cognitive memory to optimize both algorithmic and empirical research processes.
• The system overcomes traditional limitations by unifying computational modeling with laboratory experimentation to enable continuous, end-to-end discovery across diverse scientific domains.

InternAgent-1.5 is a unified agentic framework for long-horizon autonomous scientific discovery designed to operate across both computational and empirical domains. It is presented as an end-to-end system for the full discovery cycle rather than a benchmark-only reasoner: it generates hypotheses and plans, verifies them through computation, simulation, or experiment, and evolves its behavior and internal knowledge over extended research programs. The system is organized around three coordinated subsystems—Generation, Verification, and Evolution—and three corresponding foundational capabilities—deep research, solution refinement, and long-horizon memory. In the reported evaluation, it achieves leading performance on GAIA, HLE, GPQA-diamond, and FrontierScience, while also demonstrating autonomous algorithm discovery and empirical workflows in earth, life, biological, and physical sciences (Feng et al., 9 Feb 2026).

1. Scope and problem formulation

InternAgent-1.5 is positioned against what the paper characterizes as four recurrent limitations of existing autonomous science systems: domain-specific architectures, partial foundational capabilities, linear optimization pipelines, and limited long-horizon operation. The proposed target setting is end-to-end, long-horizon, autonomous scientific discovery that can operate continuously across computation, simulation, and laboratory experimentation while improving over time. In this framing, many earlier “AI scientist” systems are treated as vertical systems specialized for one domain or one workflow, whereas InternAgent-1.5 is explicitly intended to be horizontal and unified (Feng et al., 9 Feb 2026).

The system distinguishes between two broad discovery regimes. In algorithm discovery, it transforms objectives into solutions in formal systems, including learning algorithms, RL methods, forecasting systems, and molecular modeling pipelines. In empirical discovery, it transforms observations into generalizations about the physical world, including target discovery in life science, climate diagnostics, chemical reaction reasoning, and wet-lab protein engineering. This bifurcation is central to the paper’s claim that a single agentic framework can span dry-lab and wet-lab research without adopting separate architectural abstractions for each domain (Feng et al., 9 Feb 2026).

Relative to InternAgent 1.0, the paper emphasizes a shift from generation/reflection toward a fuller generation–verification–evolution loop, stronger long-horizon memory, broader support for end-to-end algorithm optimization on original codebases, more unified support across computational and empirical science, and stronger benchmark performance. The underlying claim is not merely broader task coverage, but the replacement of isolated prompt-based cycles with an explicit recurring discovery process (Feng et al., 9 Feb 2026).

2. Unified architecture

InternAgent-1.5 is built around three coordinated subsystems: Generation, Verification, and Evolution. Together they instantiate a recurring scientific loop in which Generation proposes ideas, hypotheses, plans, and candidate methods; Verification tests them through computation, simulation, or physical experimentation; and Evolution interprets outcomes, updates memory and strategy, and shapes future Generation and Verification. The paper treats this as a common abstraction for scientific discovery: hypothesis formulation, methodological evaluation, and evidence-driven refinement (Feng et al., 9 Feb 2026).

The Generation subsystem is responsible for constructing hypotheses, designing methods, building research plans, and gathering and integrating literature. Its core enabling capability is deep research, which extends InternAgent 1.0’s generation-and-reflection paradigm and produces structured hypotheses, methodological plans, and reasoning traces. It can invoke scientific tools during planning and knowledge extraction.

The Verification subsystem tests generated hypotheses and methods, executes code, runs simulations, coordinates laboratory-style assessments, and compares variants in parallel. Its core enabling capability is solution refinement, and it supports computational analyses, simulations, physical experiment execution, and parallel evaluation of alternatives. Verification is therefore not a single-pass checker; it is the system’s experimental optimization substrate.

The Evolution subsystem absorbs outcomes from Generation and Verification, interprets success and failure, updates strategies and long-term knowledge, and maintains persistent coherence over long horizons. Its core enabling capability is long-horizon memory, implemented as a structured memory system with procedural, episodic, and semantic components. In the paper’s formulation, Evolution is what turns a sequence of experiments into a research program rather than a collection of disconnected runs (Feng et al., 9 Feb 2026).

3. Deep research and structured reasoning

Deep research is the foundational capability that supports the Generation subsystem. It retrieves scientific information, integrates cross-disciplinary evidence, organizes reasoning into structured subproblems, and synthesizes grounded outputs. Its implementation has three principal components: a cross-disciplinary knowledge graph, a dynamic structured knowledge flow graph, and graph-guided output synthesis with multi-path reasoning (Feng et al., 9 Feb 2026).

The cross-disciplinary knowledge graph is explicitly richer than a simple triple-based KG. It stores documents, concepts, methods, datasets, empirical settings, and problem statements, with typed edges such as citation, usage, dependence, and product/by-product-type relations. It is constructed from papers, surveys, technical reports, and domain notes using parsing, domain-specific scientific tools, schema-guided extraction, named entity recognition, noun-phrase mining, co-citation and co-usage analysis, and consolidation of textual and citation evidence. Retrieval combines graph search, dense vector retrieval, and ranking/merging, yielding path-structured evidence chains rather than flat snippets.

The dynamic structured knowledge flow is a DAG-based representation of the research process itself:

$G=(V,E),$

with nodes

$v_i=(t_i,d_i,s_i,c_i),$

where $$t_i \in \{\text{search}, \text{solve}, \text{answer}\}$$, and edges

$e_{ij}=(v_i,v_j,r_{ij})$

encode dependency relations such as “requires result from,” “provides evidence for,” and “constrains reasoning of.” This representation turns inquiry into an explicit task-and-reasoning graph. The planner starts from a root query, identifies parts needing decomposition or more evidence, generates successor nodes, updates dependencies, and continues until coverage is sufficient. The stated consequences are adaptive decomposition, multi-agent collaboration, revisiting and refining branches, and dependency-aware propagation of results.

Once the flow graph is built, executable nodes whose dependencies are satisfied are activated, assigned to agents, executed through sequential reasoning and retrieval, and propagated downstream with updated context. A cross-disciplinary knowledge collector gathers information from scientific tools, remote resources, and SCP-connected resources. For each query, the system generates three responses—direct answer, search-augmented answer, and self-driven answer—and aggregates them into a final output. The intended effect is reduced overreliance on any single reasoning route, increased completeness, and improved factual reliability (Feng et al., 9 Feb 2026).

4. Verification, optimization, and long-horizon memory

The Verification subsystem’s central optimization problem is to search for the best solution $s$ in a search space $\mathcal{S}$, where a solution may be executable code, a parameterization, or an experimental protocol, and performance is measured by a task-dependent score $h(\mathcal{T}, s)$. This formulation is deliberately shared across algorithmic proposals and empirical experiment plans (Feng et al., 9 Feb 2026).

To optimize over this space, InternAgent-1.5 uses a Graph-Augmented Monte Carlo Search framework. It retains the classical MCTS loop—selection, expansion, simulation, and backpropagation—but replaces the rigid tree with a dynamic solution graph that aggregates information from prior experiments. The goal is cross-trajectory reuse rather than isolated branch search. The framework defines four expansion operators: Primary Expansion for local improvements using the immediate parent; Intra-branch Evolution using ancestor history within the same branch; Cross-branch Reference to import useful design elements from another branch when the current one stagnates; and Multi-branch Aggregation to combine complementary strengths from multiple top-performing nodes. Proposals are executed in a simulator, code environment, or physical experiment setup, scored, and then backpropagated through their ancestry (Feng et al., 9 Feb 2026).

In algorithm discovery, proposals are executable code specifying model components, preprocessing, hyperparameters, and evaluation settings. Candidates are compiled and run in controlled environments, and metrics such as accuracy, runtime, and resource use are collected. In empirical discovery, proposals are complete experimental procedures whose optimization may alter parameters, step ordering, operational protocols, or sub-protocol combinations. When simulated, domain models return outcomes such as yield or protein stability; when physical, SCP coordinates lab devices and measurements such as fluorescence intensity or assay signal quality are returned to the optimization loop.

Long-horizon coherence is implemented through Structured Cognitive Memory, which has three layers. Strategy-Procedural Memory (SPM) stores distilled reusable procedures, including successful reasoning structures, decision pivots, organizational patterns, lessons from failed strategies, and root-cause insights. Task-Episodic Memory (TEM) stores fine-grained within-trajectory experiment records, including attempted methods, extracted metrics, and improvement judgments. Semantic-Knowledge Memory (SKM) maintains long-term conceptual continuity through a Long-Term Experience Library and an Idea Graph, and uses pairwise combinations of generated methods and contrastive learning across method/result pairs to distill high-level principles and low-level heuristics. Its novelty scoring is defined as

$$\mathrm{nov}(c)=1-\max_{x \in \mathcal{G}} \mathrm{sim}\!\left(f_{\text{enc}}(c),\, f_{\text{enc}}(x)\right),$$

which rewards objectives far from already explored conceptual regions. The paper’s synthetic characterization of the improvement loop is: propose $\rightarrow$ evaluate $\rightarrow$ learn $v_i=(t_i,d_i,s_i,c_i),$0 re-propose (Feng et al., 9 Feb 2026).

5. End-to-end workflows, tools, and scientific domains

The end-to-end workflow begins with a scientific goal, such as an open-ended question, an optimization task, a discovery challenge, or a benchmark problem. The Generation subsystem then queries the cross-disciplinary KG, performs dense retrieval, constructs a DAG-based knowledge flow graph, decomposes the problem into search/solve/answer nodes, and records dependencies. Executable nodes are assigned to agents that search literature, query databases, invoke tools, reason over scientific evidence, and synthesize hypotheses and candidate methods. Verification then executes code, compiles programs, performs benchmark evaluations, runs simulators, or executes SCP-coordinated lab procedures. Proposals are scored with task-specific metrics—including accuracy, RMSE, MAE, F1, $v_i=(t_i,d_i,s_i,c_i),$1, HK-PCC, product prediction top-1, fluorescence intensity, assay signals, and climate diagnostic measures—and Evolution updates SPM, TEM, and SKM before the next cycle (Feng et al., 9 Feb 2026).

A major architectural claim is that InternAgent-1.5 unifies computational modeling and laboratory experimentation within the same loop. On the computational side, the system supports code execution for algorithmic proposals, simulation environments, benchmark evaluation pipelines, data analysis tools, RDKit, ESMFold, ProSST, and retrieval from scientific corpora and databases. On the laboratory side, SCP (Science Context Protocol) is the coordination layer for remote scientific resources, automated devices, wet-lab protocol execution, and measurement collection. In biological workflows, SCP coordinates DNA assembly, expression, purification, and fluorescence intensity measurement. The interaction pattern is consistently described as planning tool use, invoking the tool, collecting structured outputs, and integrating them into the reasoning graph and memory (Feng et al., 9 Feb 2026).

The reported empirical discovery cases cover four domains. In earth science, the system performs automated climate diagnostics by comparing 20 CMIP6 models against ERA5 and autonomously identifying warming trends and regional biases, and it performs climate downscaling from NCEP-NCAR-R1 coarse fields ($v_i=(t_i,d_i,s_i,c_i),$2) to ERA5-like fine resolution ($v_i=(t_i,d_i,s_i,c_i),$3). In the downscaling task, the reported RMSE is 0.8488, compared with 3.1658 for Kriging and 0.9049 for BCSD. In life science, it reproduces GPR160 as a therapeutic target in HCC by reducing an initial pool of 125 genes through multi-round evidence compression and reflection, and it prioritizes ARG2 in CRC through a multimodal evidence graph and iterative reflection, producing assay and organoid validation suggestions. In biological sciences, it executes an end-to-end fluorescent protein engineering workflow in which literature search identifies parent variants, sfGFP is selected, ESMFold and ProSST support dry-lab analysis, and SCP translates the design into DNA assembly, expression, purification, and fluorescence measurement. In physical science, it reports major-product prediction on ChemCoTBench with Top-1 $v_i=(t_i,d_i,s_i,c_i),$4 and FTS $v_i=(t_i,d_i,s_i,c_i),$5, by-product prediction with Top-1 $v_i=(t_i,d_i,s_i,c_i),$6 and FTS $v_i=(t_i,d_i,s_i,c_i),$7, and a scaffold-hopping workflow that reasons over shape/electrostatic alignment, synthetic accessibility, Tanimoto similarity, and LogP (Feng et al., 9 Feb 2026).

6. Evaluation, lineage, and limitations

InternAgent-1.5 is evaluated on scientific reasoning benchmarks, algorithm discovery tasks, and empirical discovery tasks. The benchmark results are used to support the claim that the system has strong foundational capabilities, while the discovery tasks are intended to show end-to-end scientific operation beyond question answering (Feng et al., 9 Feb 2026).

Evaluation	Best reported InternAgent-1.5 result
GAIA validation	Level 1: 92.45; Level 2: 89.53; Level 3: 61.54; Average: 86.06
HLE	Text-only: 40.87%; All/full benchmark: 40.00%
GPQA-diamond	Bio: 84.21; Chem: 79.57; Phys: 96.51; Avg: 87.37
FrontierScience Olympiad	Bio: 46.00 ± 8.00; Chem: 85.50 ± 3.67; Physics: 76.80 ± 2.99; All: 77.20 ± 3.06
FrontierScience Research	Bio: 10.33 ± 4.64; Chem: 22.00 ± 6.00; Physics: 3.67 ± 2.87; All: 12.00 ± 2.49
SGI-Bench	Deep Research: 37.74; Idea Generation: 58.11

In algorithm discovery, the paper divides results into scientific algorithm discovery and AI algorithm discovery. On the six scientific tasks, the reported InternAgent-1.5 scores are: AutoRYP $v_i=(t_i,d_i,s_i,c_i),$8 $v_i=(t_i,d_i,s_i,c_i),$9, AutoTPPR $$t_i \in \{\text{search}, \text{solve}, \text{answer}\}$$0 MSE, AutoPower $$t_i \in \{\text{search}, \text{solve}, \text{answer}\}$$1 RMSE, AutoTSF $$t_i \in \{\text{search}, \text{solve}, \text{answer}\}$$2 MAE, AutoMD $$t_i \in \{\text{search}, \text{solve}, \text{answer}\}$$3 Energy-MAE, and AutoEAP $$t_i \in \{\text{search}, \text{solve}, \text{answer}\}$$4 HK-PCC. On the four AI algorithm tasks, the reported results are: AutoTTS $$t_i \in \{\text{search}, \text{solve}, \text{answer}\}$$5 accuracy, AutoMem $$t_i \in \{\text{search}, \text{solve}, \text{answer}\}$$6 F1, AutoTTRL $$t_i \in \{\text{search}, \text{solve}, \text{answer}\}$$7 accuracy, and AutoLM $$t_i \in \{\text{search}, \text{solve}, \text{answer}\}$$8 accuracy. The strongest gains are described for AutoEAP, AutoMD, AutoPower, and AutoTPPR (Feng et al., 9 Feb 2026).

The system’s lineage is explicit. The original InternAgent is described as a unified closed-loop multi-agent framework for Autonomous Scientific Research across 12 tasks, centered on self-evolving idea generation, idea-to-methodology construction, and evolutionary experimental planning and execution (Team et al., 22 May 2025). InternAgent-1.5 extends that base toward a more explicit generation–verification–evolution decomposition, stronger long-horizon memory, broader end-to-end algorithm optimization on original codebases, and more unified support across computational and empirical science (Feng et al., 9 Feb 2026). The clearest ablation concerns SPM on GAIA: without SPM, average score is 82.42 and average tool calls are 22.69; with SPM, average score is 86.06 and average tool calls fall to 18.52, with the largest gains on harder questions (Feng et al., 9 Feb 2026).

The paper also leaves several limitations visible. It does not provide a full end-to-end training recipe, explicit RL or IL objectives for the overall agent, a compute budget for training or inference, exact loss functions beyond retrieval similarity and novelty formulas, or full low-level specifications for orchestration internals, planning policies, memory update criteria, MCTS parameterization, and laboratory control protocols. Reproducibility is partial: an open-source repository is provided at https://github.com/InternScience/InternAgent, and some complete reports are released, but run budgets, hyperparameters, infrastructure details, total compute usage, wall-clock duration, search iterations, lab throughput, token budgets, and hardware accounting are sparse. The safety discussion is also limited: the paper describes verification before hypothesis adoption, structured evidence chains, laboratory coordination through SCP, and explicit measurement loops, but does not provide a detailed formal oversight framework, biosafety protocol section, or explicit risk taxonomy (Feng et al., 9 Feb 2026).

A related systems-security paper is relevant to deployment context rather than to InternAgent-1.5’s core scientific architecture. “AIP: Agent Identity Protocol for Verifiable Delegation Across MCP and A2A” describes a protocol centered on Invocation-Bound Capability Tokens and is presented as highly relevant to an InternAgent-1.5-style multi-agent system, best understood as a foundational security layer beneath the agent framework, while also serving as a delegation protocol between sub-agents, a tool authorization framework, and a provenance/audit mechanism (Prakash, 25 Mar 2026). This suggests that as InternAgent-style systems move toward persistent multi-agent execution across tools and laboratories, delegation semantics, identity, and auditability become architectural concerns adjacent to planning, optimization, and memory.