interventions

A preregistered causal audit finds that covariate-aware time series foundation models (Chronos-2, TiRex-2, TabPFN-TS, TimesFM-2.5) forecast associations rather than intervention effects, with a derived confounding gap that can be reduced up to 83% via brief interventional post-training with LoRA adapters.

The paper explores K/V-cache interventions for persona control in decoder-only language models, finding dissociations between representation alignment and behavioral expression, with mid-layer interventions being effective for persona expression while preserving lexical diversity.

The paper evaluates how large language model agents overtrust unreliable tool returns, measuring adoption rates of corrupted information and testing interventions to mitigate this issue.

This paper establishes a precise mathematical correspondence between graph surgery and the do-operator in acyclic structural causal models, proving their equivalence in terms of dependency graphs.

DoTime is a synthetic benchmark generator for interventional and counterfactual time series, providing scalable TSCM-based data generation with exact ground truth, released as a PyPI package with evaluation suites. It enables training and benchmarking causal foundation models on non-observational time series data.

This paper introduces a matched evaluation protocol for sparse feature interventions in language models, showing that the claimed efficiency advantage of SAE-based safety control disappears or reverses when properly comparing against fair dense baselines.

Anthropic's J-space paper demonstrates that they can perform 'brain surgery' interventions into reasoning to change topics midstream, and the model is able to detect what intervention was done, indicating a form of eval awareness.