Skip to content
Modelwire
Subscribe

Foundation models excel at forecasting but fail at causal inference

Source published ·Modelwire updated

Original coverage: arXiv cs.LG ↗·How Modelwire adds context

The development

Foundation models trained to minimize forecasting error may systematically fail at causal inference, a critical gap for real-world decision-making in retail and beyond. Researchers evaluated tabular and time-series foundation models, including one specifically designed for causal estimation, and found that prediction accuracy provides no guarantee of accurate causal effect estimation. Instrumental variables, the econometric standard for handling confounding, paradoxically worsened performance across all tested models when applied as raw instruments. This finding exposes a fundamental misalignment between how foundation models are optimized and what practitioners need to drive business outcomes.

Modelwire’s AI-generated summary of coverage from arXiv cs.LG.

Modelwire analysis

Explainer

Our AI-generated reading of the wider context and the next developments to watch.

The paper's core finding isn't just that foundation models fail at causal inference, but that instrumental variables, the econometric gold standard for handling confounding, actively degrades performance across all tested architectures. This inversion of expected behavior suggests the problem runs deeper than model capacity.

This connects directly to the causal graph fusion work from November 6th, which solved how to inject causal reasoning into frozen tabular foundation models at inference time. That approach bypassed the training-optimization problem by compiling causal structure into attention mechanisms. Today's paper documents why that workaround was necessary: foundation models optimized for prediction loss have no built-in mechanism to recover causal effects, and naive application of econometric tools backfires. The gap between these two papers defines the current frontier: inference-time causal steering works, but only because training-time optimization actively works against causal reasoning.

If the researchers test whether the causal-specific foundation model mentioned in the summary (trained explicitly for causal estimation) maintains its advantage when evaluated on out-of-distribution confounding patterns not seen during training, that determines whether the problem is solvable through architecture alone or requires fundamentally different training objectives. Watch for follow-up work in Q1 2027 testing robustness across held-out causal graphs.

This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error

Coverage behind this analysis

These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.

  1. ·arXiv cs.LG

    Causal graphs now steer frozen tabular models without retraining

    Researchers have solved a longstanding friction point in clinical AI: how to apply causal reasoning across fragmented medical datasets without retraining foundation models. The work introduces a two-part system that first fuses causal graphs learned from incomplete, overlapping cohorts into a unified structure, then compiles that graph directly into a frozen tabular foundation model's attention…

    Read Modelwire coverage →Original source ↗

MentionsFoundation models · Tabular foundation models · Time-series foundation models · Instrumental variables

MW

How this coverage is produced

Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.

Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “From Forecasting to World Modeling: Can Temporal Foundation Models Recover Causal Response?”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.

More to explore

Causal graphs now steer frozen tabular models without retraining

Synthetic training breaks tabular model scaling law predictions

arXiv cs.LG·

Foundation models for power grids fail generalization tests against classical baselines

arXiv cs.LG·
Foundation models excel at forecasting but fail at causal inference · Modelwire