Sibyl: Anticipatory Decision Systems

• Sibyl is a term used for diverse anticipatory decision systems that filter and predict uncertain evidence using tailored machine learning methods across multiple domains.
• In gamma-ray astrophysics, Sibyl employs a Random Forest classifier achieving over 97% accuracy to distinguish AGNs, pulsars, and potential dark matter subhalo candidates.
• Other implementations include reinforcement learning for hybrid storage, LLM-based reasoning for dialogue and planning, workload forecasting in databases, and SHAP-driven decision support in high-stakes applications.

Searching arXiv for papers titled or containing “Sibyl” to ground the article and verify the set of relevant works. {"query":"ti:Sibyl OR abs:Sibyl", "max_results": 10} I’ll use the arXiv search tool directly to retrieve relevant “Sibyl” papers. {"query":"Sibyl", "max_results": 10} Sibyl, sometimes capitalized as SIBYL, is a recurrent name for computational systems that perform prediction, triage, explanation, or adaptive control in technically distinct domains. In the arXiv literature, the name denotes a Random Forest classifier for unassociated gamma-ray sources in the Fermi-LAT catalogue, an online reinforcement-learning policy for data placement in hybrid storage systems, an LLM-based agent framework for complex real-world reasoning, a future-aware commonsense inference framework for empathetic dialogue, a workload-forecasting framework for database queries, a visual analytics tool for high-stakes child welfare screening, and a file-backed autonomous research system built around self-evolving trial-and-error harnesses (Mirabal et al., 2012).

1. Scope, naming, and major uses

Capitalization varies between Sibyl and SIBYL, and the term does not denote a single method or software lineage. Rather, it is reused for systems whose functions differ substantially in problem setting, model class, and evaluation protocol.

Domain	Sibyl denotes	Representative paper
Gamma-ray astrophysics	Random Forest classifier for Fermi-LAT sources	(Mirabal et al., 2012)
Hybrid storage systems	Online RL-based data placement policy	(Singh et al., 2022)
LLM agents	Reentrant reasoning framework with global workspace and debate jury	(Wang et al., 2024)
Empathetic dialogue	Prophetic commonsense inference framework	(Wang et al., 2023)
Database systems	Future-query workload forecasting framework	(Huang et al., 2024)
High-stakes decision support	Visual analytics tool for interpretable ML	(Zytek et al., 2021)
Autonomous research	File-backed self-evolving research system	(Wang et al., 21 May 2026)

This distribution suggests a shared naming preference for systems that are expected to anticipate latent structure, future events, or decision-relevant signals. That interpretation is inferential; the individual papers are methodologically unrelated except where later works explicitly build on earlier ones, as in the hybrid-storage lineage (Singh, 2022).

2. Gamma-ray source classification and dark-matter triage

The earliest Sibyl in this set is introduced in Fermi gamma-ray astronomy as a Random Forest-based gamma-ray source classifier for unassociated Fermi-LAT sources at high Galactic latitudes (Mirabal et al., 2012). The motivating problem is that the Second Fermi-LAT catalogue (2FGL) contains many sources with no obvious lower-energy counterpart, while the dominant known high-latitude classes are active galactic nuclei (AGN) and pulsars. The system is therefore framed as a binary classifier whose broader purpose is not only source classification but also the isolation of unusual objects that might merit consideration as dark matter subhalo candidates.

The training sample comprises 800 labeled AGNs and 108 labeled pulsars, for a total of 908 identified/associated sources. Because the classes are highly imbalanced, the pulsar sample is replicated to give more effective weight to the minority class. From the 68 reported features in 2FGL, the authors additionally construct four flux-ratio features from the five energy bands. Random Forest feature importance, measured with MeanDecreaseAccuracy, identifies Index, Curve, Variability, and the flux ratios $FR_{12}, FR_{23}, FR_{34}, FR_{45}$ as the most discriminating variables (Mirabal et al., 2012).

Sibyl follows Breiman’s Random Forest procedure: a large ensemble of classification trees votes on whether a source is an AGN or a pulsar, with the forest prediction given by majority vote. Although the formal class decision is associated with $P > 0.5$, the catalogue application uses a stricter threshold of $P > 0.7$, so that at least 70% of the trees must agree. Across 100 alternate training/testing splits, each using 2/3 of the sample for training, 1/3 for testing, and 500 trees, the classifier reaches 97.1% overall accuracy, with 97.7% for AGNs and 96.5% for pulsars. Including Galactic latitude $|b|$ changes the classwise performance to 97.4% for AGNs and 95.5% for pulsars, and the authors note that latitude may introduce a small bias because pulsars are concentrated near the Galactic plane (Mirabal et al., 2012).

Applied to the 269 unassociated 2FGL sources with $|b| \ge 10^\circ$, Sibyl predicts 216 AGN candidates, 16 pulsar candidates, and 37 inconclusive sources under the $P > 0.7$ cutoff. The inconclusive set is then examined with Random Forest proximity and outlyingness. The paper notes that outlyingness values much greater than 10 are usually suggestive of novel cases; among the 37 inconclusive sources, the largest values are around 9.0, 7.2, 7.1, 6.6, 6.2, with 2FGL J0953.6−1504 at 9.0. The authors therefore report no compelling dark matter subhalo candidate in this sample (Mirabal et al., 2012).

3. Online reinforcement learning for hybrid storage systems

A second major use of the name appears in computer architecture and storage systems, where Sibyl is an online, reinforcement learning-based data-placement system for hybrid storage systems (HSS) (Singh et al., 2022). Here the problem is page or request placement between a fast but small device and a slower but larger device, under workloads and device characteristics that vary over time. The 2026 review article situates this Sibyl alongside Pythia and Hermes as one of three ML-guided architectural policies spanning the memory hierarchy (Bera et al., 15 Mar 2026).

The RL formulation uses a compact 6-dimensional state vector

$$O_t = (size_t, type_t, intr_t, cnt_t, cap_t, curr_t),$$

where the components encode request size, read/write type, access interval, access count, remaining fast-tier capacity, and current placement. In the dual-device setting, the action space consists of placing the page in fast storage or slow storage. The reward is based on request latency, with an explicit eviction penalty: $$R = \begin{cases} \frac{1}{L_t} & \text{if no eviction from fast storage to slow storage} \ \max\left(0,\frac{1}{L_t} - R_p\right) & \text{in case of eviction} \end{cases}$$ with $R_p = 0.001 \times L_e$, where $L_e$ is the time spent evicting pages from fast to slow storage (Singh et al., 2022).

The implementation uses a deep Q-learning-style design with Categorical DQN (C51), a lightweight feed-forward network with 6 input neurons, hidden layers of 20 and 30 neurons, and one output neuron per action. It is deliberately split into two parallel threads: an RL decision thread on the critical path and an RL training thread that asynchronously consumes experiences $P > 0.5$0 from an experience buffer and periodically copies weights into the inference network. The 2026 article emphasizes the same architectural decoupling, noting that Sibyl is implemented in the storage management layer of the host operating system and uses two identical neural networks, one for inference and one for background training (Bera et al., 15 Mar 2026).

Quantitatively, Sibyl improves average request latency by 21.6% in a performance-oriented HSS and 19.9% in a cost-oriented HSS compared to the best previous data-placement technique (Singh et al., 2022). In more detailed comparisons, the performance-oriented configuration shows 28.1% over CDE, 23.2% over HPS, 36.1% over Archivist, and 21.6% over RNN-HSS, while the cost-oriented configuration shows 19.9%, 45.9%, 68.8%, and 34.1%, respectively (Bera et al., 15 Mar 2026). In a Tri-HSS with three heterogeneous devices, Sibyl improves average performance by 23.9%–48.2% over the best heuristic-based policy and achieves about 80% of oracle performance (Singh et al., 2022). The reported total DRAM overhead is 124.4 KiB, with roughly 10 ns inference latency and about 2 µs training latency per step (Singh et al., 2022).

4. LLM agents and future-aware dialogue reasoning

In LLM systems, Sibyl names two distinct frameworks. The first is a simple yet effective LLM-based agent framework for complex real-world reasoning (Wang et al., 2024). It is organized around four modules: a tool planner, an external information acquisition channel, a global workspace, and a multi-agent debate-based jury. The design emphasizes reentrancy, treating internal LLM calls as stateless QA functions rather than persistent dialogues. It relies primarily on two general-purpose tools—a web browser and a Python terminal—and uses selective compression so that only new, unique, necessary information enters the shared workspace (Wang et al., 2024).

Evaluated on GAIA with GPT-4o API (text only) and a maximum of 20 reasoning steps, the system reaches 47.31% on Level 1, 32.70% on Level 2, 16.33% on Level 3, and 34.55% overall on the test set. The paper reports this as the best result among GPT-4-based agents in that setup, exceeding AutoGen at 32.33%, FRIDAY at 24.25%, AutoGPT-4 at 5.00%, GPT-4 with plugins at 14.6%, and GPT-4 at 4.00%. On the validation set, full Sibyl reaches 40.00, compared with 36.97 without the debate jury and 36.77 without majority vote (Wang et al., 2024).

The second LLM-related Sibyl is the dialogue framework also called Prophetic Commonsense Inference (PCI) (Wang et al., 2023). Its premise is that dialogue commonsense should be both sensible and visionary, meaning aligned with the immediately succeeding turn(s) rather than inferred only from past context. The framework operationalizes this with four categories of prophetic commonsense: cause, subsequent event, emotion reaction, and intention. Training proceeds in three stages: ChatGPT first generates future-aware commonsense using dialogue context and the ground-truth response; LLaMA2-7B chat is then fine-tuned with SFT + LoRA to predict that commonsense from context only; finally, the downstream response generator conditions on both dialogue input and predicted commonsense (Wang et al., 2023).

The evaluation uses EmpatheticDialogues (ED) and Emotion Support Conversation (ESConv), along with automatic metrics, human evaluation, and G-Eval with GPT-3.5. The paper reports that LLaMA2 + PCI improves response quality across most metrics, and that human judges prefer PCI over CASE, Vanilla, COMET, and CICERO on empathy, coherence, informativeness, and supportiveness. Ablations show that removing any commonsense category hurts performance, that intent is especially important, and that removing the future-aware acquisition stage causes a sharp decline (Wang et al., 2023).

5. Forecasting, explainability, and decision support

In database systems, SIBYL is an end-to-end ML framework for forecasting future query workloads when production workloads are time-evolving (Huang et al., 2024). The framework predicts entire future query statements, including query templates, parameter values, and arrival times, under both next-$P > 0.5$1 and next-$P > 0.5$2 forecasting regimes. Its core technical choices are template-based featurization, a stacked-LSTM encoder-decoder architecture, cut-and-pack mechanisms for scalability, and incremental fine-tuning for workload drift. On four real workloads, it achieves an 87.3% median F1 score and yields $P > 0.5$3 and $P > 0.5$4 performance improvement when applied to materialized view selection and index selection, respectively (Huang et al., 2024).

A different Sibyl addresses explainability and usability in high-stakes human decision making, specifically child welfare screening (Zytek et al., 2021). The tool is built around local factor contributions computed using SHAP, presented through a visual analytics interface that aligns with domain language such as risk factors and protective factors. The underlying predictive model is a LASSO regression trained on 461 features and mapped to a 1–20 risk score representing the likelihood of removal from home within the next two years. Through field observations, interviews, iterative prototyping, and two formal user studies—12 non-expert participants and 13 expert participants—the paper identifies the Case-Specific Details view as the most useful component: experts marked it helpful in 91.8% of case analyses and non-experts in 90.7%. The broader conclusion is that, in this setting, case-specific local explanations were more useful than global summaries during live decision making (Zytek et al., 2021).

6. Autonomous research, auditability, and recurring motifs

The most recent Sibyl in this set is Sibyl-AutoResearch, instantiated in SIBYL, a file-backed autonomous research system built around Scientific Trial-and-Error Harnesses (Wang et al., 21 May 2026). The paper’s central claim is that executable research workflows do not by themselves produce research judgment; what matters is whether trial outcomes change later behavior and whether recurring failures cause the harness itself to evolve. Two auditable conversion units formalize this requirement: trial-to-behavior conversion, which asks whether a trial signal changes later research actions, and trial-to-harness-behavior conversion, which asks whether recurring process failures change the harness through new gates, overlays, telemetry requirements, scheduler policies, repair tasks, artifact contracts, or protected constraints (Wang et al., 21 May 2026).

The system preserves research state, plans, role outputs, experiment artifacts, reviews, reflections, writing products, and event logs as files, enabling retrospective inspection. The paper reports 8 high-confidence conversion events, with median latency of 1 iteration and maximum visible latency of 3 iterations, and a recovered-failure registry containing five naturally occurring failure classes: duplicate result files, confidence-interval inversion, stale headline number, feature-count mismatch, and unsupported statistics. The authors explicitly state that these traces do not establish a comparative performance claim; rather, they show that the proposed conversion paths are recoverable from realistic autonomous-research workspaces (Wang et al., 21 May 2026).

Across these disparate uses, Sibyl usually names a system that filters, predicts, or routes uncertain evidence into a more actionable form. This suggests a common editorial characterization of Sibyl as an anticipatory decision system—an Editor's term—but that characterization is interpretive. In the source papers themselves, Sibyl is not a unified framework; it is a reused name attached to distinct technical artifacts spanning Random Forest classification, deep reinforcement learning, stacked-LSTM forecasting, SHAP-based interpretability, LLM-agent orchestration, future-aware dialogue commonsense, and auditable autonomous-research infrastructure (Mirabal et al., 2012).