Research

Proof, benchmarks, and
research output

Research is where Abel publishes benchmark results, structural findings, and the validity boundaries behind the product and platform claims.

Read about our approach or start with the Product, Platform, or Docs pages for onboarding.

0+

Publications

0K+

Causal Variables

0+

Causal Methods

0

Benchmark Categories

Research Directions

Four pillars of causality-empowered AI

Each direction strengthens the foundations of intelligence systems — from autonomous reasoning to world-scale causal computation.

Autonomous Reasoning

Expert Causal Agent

Autonomous causal reasoning agents that discover, intervene, and validate without human-in-the-loop — making rigorous causal analysis accessible through natural language.

Explore research
Interpretability

White-box LLM / VLM

Interpretable causal foundation models that make their reasoning chains transparent and auditable — transforming black boxes into steerable, intervention-ready systems.

Explore research
Simulation

Causal Driven World Model

Causal world simulation and planning that goes beyond next-token prediction to structural consequence computation — observe, intervene, learn.

Explore research
Structure Learning

Causal Discovery Model

Large-scale transformer-based causal structure learning at scale — reframing causal discovery as a general-purpose inference capability embedded in foundation models.

Explore research

Publications

Selected research

Peer-reviewed papers and technical reports from Abel Research.

Causal-Copilot: An Autonomous Causal Analysis Agent

Xinyue Wang, Kun Zhou, Wenyi Wu, Har Simrat Singh, Fang Nan, Songyao Jin, Aryan Philip, Saloni Patnaik, Hou Zhu, Shivam Singh, Parjanya Prashant, Qian Shen, Biwei Huang

arXiv preprint arXiv:2504.13263 (2025)
Expert Causal AgentCausal Discovery

GPU-Accelerated PCMCI for Large-Scale Causal Discovery

Abel Research

NeurIPS 2026 Workshop
Causal DiscoveryScalability

Schema-as-API: Decoupling Language from Causal Computation

Abel Research

Technical Report
Expert Causal AgentArchitecture

Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models

Zekai Zhao, Qi Liu, Kun Zhou, Zihan Liu, Yifei Shao, Zhiting Hu, Biwei Huang

arXiv preprint arXiv:2505.17697 (2025)
White-box LLM/VLM

Towards Generalizable Reinforcement Learning via Causality-guided Self-adaptive Representations

Yupei Yang, Biwei Huang, Fan Feng, Xinyue Wang, Shikui Tu, Lei Xu

arXiv preprint arXiv:2407.20651 (2024)
Causal World Model

Modeling Unseen Environments with Language-guided Composable Causal Components in Reinforcement Learning

Xinyue Wang, Biwei Huang

arXiv preprint arXiv:2505.08361 (2025)
Causal World ModelWhite-box LLM/VLM

Learning World Models with Identifiable Factorization

Yu-ren Liu, Biwei Huang, Zhengmao Zhu, Honglong Tian, Mingming Gong, Yang Yu, Kun Zhang

NeurIPS 2023
Causal World ModelLatent Causal Discovery

Generalized Independent Noise Condition for Estimating Latent Variable Causal Graphs

Feng Xie, Ruichu Cai, Biwei Huang, Clark Glymour, Zhifeng Hao, Kun Zhang

NeurIPS 2020
Latent Causal Discovery

Learning Discrete Concepts in Latent Hierarchical Models

Lingjing Kong, Guangyi Chen, Biwei Huang, Eric Xing, Yuejie Chi, Kun Zhang

NeurIPS 2024
Latent Causal Discovery

Differentiable Causal Discovery for Latent Hierarchical Causal Models

Parjanya Prajakta Prashant, Ignavier Ng, Kun Zhang, Biwei Huang

arXiv preprint arXiv:2411.19556 (2024)
Latent Causal DiscoveryCausal Discovery

10 publications

Structural Findings

Real discoveries from Abel's live graph

Statistically significant directed causal edges with p-values, β coefficients, and time lags.

Preferred Share / mREIT Reflexive Cluster

Reflexive Loop

A tightly coupled feedback loop between preferred shares and mortgage REITs, invisible to correlation analysis. Yield-seeking behavior creates self-reinforcing price dynamics.

BTC and SOL as Causal Islands

Causal Isolation

Bitcoin and Solana are causally isolated from the broader macro graph. The "crypto as diversifier" thesis is structurally supported — not just statistically observed.

USDC → Regional Bank Chain

Cross-Domain

A directed causal path from USDC stablecoin dynamics to regional bank behavior. No correlation screen would surface this cross-domain link.

Epistemic Honesty

Validity boundaries

Observational, not interventional ground truth

Abel discovers causal structure from observational time-series data. Scale (200K+ variables) and freshness (daily re-discovery) mitigate limitations — but Abel does not claim interventional ground truth from observational data alone.

Coverage is non-uniform

Structural estimates are strongest where data is dense, high-frequency, and temporally structured. Coverage weakens in domains with sparse, low-frequency, or heavily censored data.

Structural computation, not metaphysical certainty

When Abel says "A drives B," the claim is: the best structural estimate under PCMCI is a directed edge with these parameters and this confidence level. A statistical structural claim, not a philosophical one.

See our Glossary for definitions of key terms.

Benchmark

CausalBench

The first comprehensive benchmark for causal reasoning in AI. Comparing LLMs vs Abel on Layer 2 and Layer 3 tasks.

CausalBench v1.0

March 2026

Model accuracy across Pearl's Causal Hierarchy

Intervention

P(Y | do(X)) — causal manipulation

L2
1Abel
84%
2Claude 3.5
26%
3GPT-4o
23%
4Gemini 1.5
20%

LLMs score near-random on intervention tasks. P(Y|do(X)) cannot be computed from observational patterns alone — this is a mathematical impossibility.

Results from CausalBench v1.0 (March 2026). Accuracy is exact-match for L1/L2/L3; path fidelity for cross-domain.

Explore CausalBench