Qiushi Discovery Engine
- Qiushi Discovery Engine is an LLM-based system that autonomously conducts scientific discovery on a real optical platform by integrating hypothesis generation, measurement, and revision.
- It employs a dual-layer architecture with specialized agents and nonlinear research phases to maintain long-horizon coherence and adaptive experimental trajectories.
- Demonstrations include reproducing established experiments, translating theory into observables, and discovering optical bilinear interactions analogous to Transformer attention.
Searching arXiv for the primary paper and closely related "discovery engine" work to ground the article in current research. Qiushi Discovery Engine is an LLM-based agentic system for end-to-end autonomous scientific discovery on a real optical platform. It is presented as a system that moves beyond assisting predefined research workflows by maintaining adaptive and stable research trajectories across long-horizon investigations involving thousands of LLM-mediated reasoning, measurement and revision actions. In the reported demonstrations, it autonomously reproduces a published transmission-matrix experiment on a non-original platform, converts an abstract coherence-order theory into experimental observables, and, in an open-ended optical-computing study, proposes and experimentally validates optical bilinear interaction, a physical mechanism described as structurally analogous to a core operation in Transformer attention (Yang et al., 29 Apr 2026).
1. Conceptual scope and research problem
Qiushi Discovery Engine is framed around a specific research problem: how to sustain a coherent scientific investigation when the system must read literature, design observables, write code, control instruments, revise hypotheses, and narrow claims under experimental constraints. The paper treats this as a long-horizon coherence problem rather than a short-horizon tool-use problem. Its central claim is that the system does not merely propose hypotheses or automate isolated steps, but can start from a broad scientific theme, interact with a physical apparatus, revise its own claims, and produce experimentally supported results (Yang et al., 29 Apr 2026).
The reported domain is a real free-space optical platform. This matters because the platform is described as combining high-dimensional control, strong mode mixing, and direct measurement, thereby forcing the agent to manage calibration, representation, observability, and evidential sufficiency in a setting that is neither purely simulated nor purely textual. The paper therefore positions Qiushi as a proof of principle for research-level autonomous agents operating under real measurement constraints rather than as a generic literature-only or benchmark-only system (Yang et al., 29 Apr 2026).
A recurring misconception is that the term discovery engine necessarily denotes a search or recommendation system. In the Qiushi case, the term instead refers to a system that performs reasoning, measurement, and revision actions within an experimental loop. This distinguishes it from discovery engines that operate over papers, structured corpora, or static datasets.
2. Dual-layer architecture, nonlinear phases, and Meta-Trace memory
The architecture is explicitly dual-layer. The core layer contains four role-specialized agents: Lead Investigator, Method Builder, Experimentalist, and Critical Reviewer. The support layer contains context-isolated sub-agents for history review, knowledge retrieval, hypothesis exploration, trajectory tracking, and evidence verification. The paper emphasizes that these are not arranged as a rigid pipeline; rather, they are complementary reasoning modes that can be activated in different orders as the investigation evolves (Yang et al., 29 Apr 2026).
The system’s research flow is organized into three nonlinear phases: Explore, Execute, and Express. Explore covers literature reading, hypothesis generation, theory mapping, and observable design. Execute covers code, simulation, measurement, and data analysis. Express covers figure making, evidence synthesis, manuscript drafting, and critical review. These phases are decoupled from the agent roles, so any core agent can act in any phase. This produces what the paper describes as a combinatorially rich trajectory space rather than a fixed workflow. A failed measurement can send the system from execution back to exploration; manuscript drafting can send it from expression back to experiment if a claim is not sufficiently supported (Yang et al., 29 Apr 2026).
Long-horizon stability is handled through Meta-Trace memory and a two-level information flow. At the boundary of each Agent Step, the acting agent distills the current state into a structured scientific memory specifying what was attempted, what was found, what evidence supports the current view, what limitations remain, what artifacts were produced, and what the next agent should do. At the next step, the system receives only updated, condensed context—short-term memory, Meta-Trace, and progressively disclosed knowledge/skills—rather than the full raw history. Within a step, support sub-agents and tools can be consulted, but their outputs are curated into task-relevant evidence rather than appended as uncontrolled context. The paper presents this mechanism as its answer to the long-horizon coherence problem (Yang et al., 29 Apr 2026).
This design suggests that Qiushi’s central contribution is not only experimental control, but also trajectory management: the preservation of scientific continuity while still permitting revision.
3. Optical platform and experimental interface
The physical substrate is a real free-space optical platform containing a laser diode, a spatial light modulator (SLM), beam splitters, optical power monitoring, diffraction and filtering optics, a diffuser to create strong scattering, and two CMOS cameras for reconstructed-field and scattering-output detection. The platform is deliberately complex. The diffuser produces distributed speckle and nonlocal input-output relations, so the system must reason about a difficult medium rather than a trivially interpretable one (Yang et al., 29 Apr 2026).
The SLM provides more than two million 10-bit-addressed pixels, corresponding to a control space on the order of configurations, while the cameras observe tens of millions of output pixels. The reported implication is not that the engine searches this space exhaustively, but that the apparatus exposes an extreme control-and-observation regime in which experiment design, observable selection, and calibration become central (Yang et al., 29 Apr 2026).
The paper’s experimental logic repeatedly depends on translating abstract theory into platform-specific observables. In the coherence-order study, the system determines that raw camera intensity in the self-referenced optical setup is not itself the correct observable because it contains background and interference terms. It also determines that the platform launches coherent pure fields, so mixed-state coherence spectra must be represented through deterministic weighted reconstruction rather than random-state preparation. This is characteristic of the system’s methodology: theory is not applied directly, but reformulated into an experimentally tractable representation compatible with the hardware (Yang et al., 29 Apr 2026).
A plausible implication is that the platform is serving two roles simultaneously: it is both the target of control and the medium that forces the agent to make scientifically disciplined choices about observables, limitations, and claim scope.
4. Demonstration studies
The paper reports three studies of increasing difficulty: a reproduction study, a theory-to-observable translation study, and an open-ended discovery study. Together they are intended to show that the same architecture can transfer a known protocol, validate a nontrivial theoretical prediction, and generate an experimentally supported new mechanism (Yang et al., 29 Apr 2026).
| Study | Main reported outcome | Reported scale |
|---|---|---|
| Transmission-matrix reproduction | Reproduced a published transmission-matrix experiment on a non-original platform; achieved phase-conjugate focusing; bounded unsupported reconstruction claims | Over 50 Agent Steps; 366.4 minutes; 27.6 million tokens; 482 LLM calls; 439 tool calls; 1,025 calibrated measurements |
| Coherence-order validation | Converted majorization-order theory into transport observables and validated the predicted interval structure on a real optical platform | 38 Agent Steps; 175.8 minutes; 22.06 million tokens; 337 LLM calls; 182 tool calls |
| Open-ended optical-computing study | Proposed and experimentally validated optical bilinear interaction | 206 Agent Steps; 1,288.1 minutes; 145.9 million tokens; 3,242 LLM calls; 1,242 tool calls; 163 research notes; 44 scripts |
In the first study, Qiushi starts from Popoff et al. (PRL 2010) and a basic platform description, then translates the protocol into the local hardware, repairs the software-hardware interface, designs calibrated phase-stepped measurements, performs pilot runs, and scales to a full transmission-matrix acquisition. The resulting matrix supports phase-conjugate focusing. A notable internal control occurs when the Critical Reviewer rejects stronger image or pattern reconstruction claims, after which a targeted follow-up experiment confirms the limitation. The paper also reports that focusing enhancement improves the best-case focusing enhancement from 25.59 to 46.1 when an annular reference-field geometry is optimized (Yang et al., 29 Apr 2026).
In the second study, the target is the majorization-order theory of wave coherence, according to which a less coherent spectrum should have a more restricted set of achievable transport responses than a more coherent one. Qiushi reformulates this in terms of transmission-matrix-derived transport operators, using a self-referenced $16$-port transmission matrix. Across all tested comparable pairs, the response interval of the less coherent spectrum is nested inside that of the more coherent spectrum, exactly as the theory predicts. For incomparable pairs, there is no universal nesting order; instead, partial overlap appears in at least one readout system. The paper presents this as the first experimental observation of this coherence-order structure in optics and the first validation of this transport prediction on a real optical platform (Yang et al., 29 Apr 2026).
In the third study, the initial prompt is deliberately broad—optical computing for AI—with no target mechanism specified. The system explores four candidate directions: deterministic physical token embedding in complex scattering media, bilinear interaction engines for compact optical pairwise computation, family-controlled physical interaction geometry, and scale-dependent specialization and degeneracy boundaries in family-programmable scattering engines. The final discovery emerges only after a recorded shift in framing, captured in Meta-Trace at Agent Step 39, away from the single-token embedding formulation (Yang et al., 29 Apr 2026).
5. Optical bilinear interaction and the attention analogy
The final open-ended result is optical bilinear interaction. The mechanism is derived by combining coherent superposition of independently encoded optical fields, high-dimensional scattering-induced mixing, and square-law detection by the camera. From these ingredients, the system infers that two encoded inputs should generate a measurable cross-term rather than merely two separate embeddings. The paper describes this as a conceptual shift from treating the scattering medium as a static encoder to treating it as a physical interaction engine (Yang et al., 29 Apr 2026).
The analogy to Transformer attention is drawn at the level of bilinear compatibility. The paper gives the schematic form
and then presents an optical analogue in which two independently encoded optical fields are coherently superposed with controlled relative phase, passed through a scattering medium, and then detected with square-law intensity measurement. Using four-phase interferometric demodulation and blank subtraction, the system isolates a complex channel-wise bilinear term at each detector channel. The resulting pair-dependent feature is described as a complex interaction field. The key claim is that this is not a sum of two independent features; it is a cross-term generated by the physics of coherent interference plus nonlinear detection (Yang et al., 29 Apr 2026).
The evidence is experimental rather than purely formal. The system validates the mechanism in a four-token XOR experiment and an eight-token semantic benchmark. The paper emphasizes that XOR is diagnostically useful because it cannot be solved by a purely linear representation of individual inputs; it requires an interaction term. The extracted Complex-B field is reported to resolve pair identity and XOR parity. In the semantic benchmark, Complex-B responses cluster distinctly by ordered pair, and under matched linear evaluation they preserve pair identity, same-category relation, and category-pair structure simultaneously. By contrast, simple token concatenation and an intensity-only digital bilinear baseline each fail on at least one of these axes (Yang et al., 29 Apr 2026).
The paper interprets this as evidence for a physically generated pair-dependent representation, not merely a post-processed intensity image. It further states that the discovered mechanism suggests a route towards high-speed, energy-efficient optical hardware for pairwise computation (Yang et al., 29 Apr 2026).
6. Position within the broader discovery-engine literature, limitations, and implications
Within the broader literature, discovery engine is a heterogeneous term. Etymo is a discovery engine for AI research built around an adaptive similarity-based network of research papers for ranking, recommendation, and visualization (Zhang et al., 2018). A distinct line of work builds a search engine for scientific challenges and directions, using expert-annotated full-text sentences and a dedicated search interface (Lahav et al., 2021). Another framework uses LLM-driven distillation of publications into structured “knowledge artifacts” and a Conceptual Tensor for AI-driven synthesis and navigation of scientific knowledge landscapes (Baulin et al., 23 May 2025). A further system benchmarked as The Discovery Engine combines machine learning with state-of-the-art ML interpretability for automated modeling across scientific datasets (Foxabbott et al., 1 Jul 2025). Qiushi differs in that its reported core is a real experimental loop involving measurement, tool calls, physical apparatus control, and revision of scientific claims under evidence (Yang et al., 29 Apr 2026).
The paper is also explicit about limitations. Its success depends on a carefully designed physical interface and a highly structured optical platform. It notes that raw traces must be carefully managed to avoid destabilizing the main reasoning context, which is why Meta-Trace and the dual-layer design are necessary. It further states that the experiments remain within a domain where the platform can be iteratively probed, calibrated, and modeled, and that this is not a claim that any arbitrary scientific field is immediately automatable. The open-ended discovery is likewise bounded by the chosen domain of optical computing and by a platform with rich but controlled physics (Yang et al., 29 Apr 2026).
The principal scientific significance claimed for the work is twofold. First, the paper presents the first experimental observation of this class of coherence-order structure on a real optical platform. Second, it presents what it describes as the first demonstration of an AI agentic system autonomously identifying and experimentally validating a nontrivial, previously unreported physical mechanism in a real experimental environment (Yang et al., 29 Apr 2026). A cautious reading is that Qiushi is best understood as a proof of principle for long-horizon, evidence-constrained autonomous research in instrumented laboratory settings, rather than as a universal template for all scientific automation.