Human interventions are a common source of data in autonomous systems during testing. These interventions provide an important signal about where the current policy needs improvement, but are often noisy and incomplete. We define Robust Intervention Learning (RIL) as the problem of learning from intervention data while remaining robust to the quality and informativeness of the intervention signal. In the best case, interventions are precise and avoiding them is sufficient to solve the task, but in many realistic settings avoiding interventions is necessary but not sufficient for achieving good performance. We study robust intervention learning in the context of emergency stop interventions and propose Residual Intervention Fine-Tuning (RIFT), a residual fine-tuning algorithm that treats intervention feedback as an incomplete learning signal and explicitly combines it with a prior policy. By framing intervention learning as a fine-tuning problem, our approach leverages structure encoded in the prior policy to resolve ambiguity when intervention signals under-specify the task. We provide theoretical analysis characterizing conditions under which this formulation yields principled pol
It is recognised that treatment-related clustering should be allowed for in the sample size and analyses of individually-randomised parallel-group trials that evaluate therapist-delivered interventions such as psychotherapy. Here, interventions are a treatment factor, but therapists are not. If the aim of a trial is to separate effects of therapists from those of interventions, we propose that interventions and therapists should be regarded as two potentially interacting treatment factors (one fixed, one random) with a factorial structure. We consider the specific design where each therapist delivers each intervention (crossed therapist-intervention design), and the resulting therapist-intervention combinations are randomised to patients. We adopt a classical Design of Experiments (DoE) approach to propose a family of orthogonal factorial designs and their associated data analyses, which allow for therapist learning and centre too. We set out the associated data analyses using ANOVA and regression and report the results of a small simulation study conducted to explore the performance of the proposed randomisation methods in estimating the intervention effect and its standard error,
Sparse Autoencoders (SAEs) decompose residual-stream activations into interpretable features. Recent latent-space defenses increasingly rely on these decompositions, assuming that identified "unsafe" SAE features serve as actionable handles for monitoring and intervention. In this paradigm, clamping a specific harmful feature is expected to reliably prevent model misbehavior. However, we show that this success may hide a recoverable failure mode: the clamp may block one visible route to a behavior without eliminating the behavior itself. We formulate this vulnerability as post-intervention recovery, a constrained residual-space optimization problem. Starting from the post-intervention residual state, we optimize residual perturbations to recover the pre-intervention behavior while preserving the post-intervention values of the targeted SAE features. Even under a strong threat model where the intervention remains active throughout optimization and generation, recovery remains possible. To rule out that recovery simply undoes the intervention, we use encoder-orthogonal updates for single-layer interventions and the corresponding feature-map Jacobian in the cross-layer setting. Across
Cyberbullying is a pervasive problem in online environments, causing substantial psychological harm to victims. Although bystander intervention has proven effective in mitigating its impact, motivating bystanders to engage in direct intervention remains a persistent challenge. Studies have suggested that difficulties in intervention skills and defending self-efficacy hinder bystanders from initiating direct intervention. To address this challenge, we introduced EmojiGen, an AI intervention tool designed to empower bystanders for direct intervention. EmojiGen enabled users to simply select an emoji as an intention clue, which subsequently combined the cyberbullying context to generate responses. In a between-subjects experiment involving 90 participants on a custom-built social media platform, we found that EmojiGen significantly increased the frequency of direct bystander interventions, both in supporting victims and in confronting perpetrators, driven by different factors. EmojiGen also increased the sense of knowing how to help and defending self-efficacy, while reducing perceived workload and anxiety associated with initiating intervention. The study contributed to the CSCW comm
Representation intervention has emerged as a promising paradigm for aligning large language models toward desired behaviors without modifying model weights. Existing methods typically apply a fixed intervention uniformly across all inputs. However, we find that the appropriate intervention direction and strength vary substantially across samples, and such indiscriminate intervention leads to degradation of general capabilities on benign inputs. To address these challenges, we propose Multi-Adapter Representation Interventions via Energy Calibration (MARI). Specifically, we introduce a competitive multi-adapter mechanism in which specialized experts capture non-linear correction patterns and adaptively determine the appropriate intervention direction and strength for different samples. Furthermore, we design an energy-based gating module that leverages internal propagation dynamics to distinguish inputs that are applicable for intervention. Extensive experiments across diverse model families and parameter scales demonstrate that MARI achieves state-of-the-art alignment performance. Our method significantly improves performance on TruthfulQA, BBQ, and safety benchmarks, while maintai
Estimating the causal effect of a time-varying public health intervention on the course of an infectious disease epidemic is an important methodological challenge. During the COVID-19 pandemic, researchers attempted to estimate the effects of social distancing policies, stay-at-home orders, school closures, mask mandates, vaccination programs, and many other interventions on population-level infection outcomes. However, measuring the effect of these interventions is complicated by time-varying confounding: public health interventions are causal consequences of prior outcomes and interventions, as well as causes of future outcomes and interventions. Researchers have shown repeatedly that neglecting time-varying confounding for individual-level longitudinal interventions can result in profoundly biased estimates of causal effects. However, the issue with time-varying confounding bias has often been overlooked in population-level epidemic intervention evaluations. In this paper, we explain why associational modeling to estimate the effects of interventions on epidemic outcomes based on observations can be prone to time-varying confounding bias. Using causal reasoning and model-based s
Rainfall is associated with the outbreak of certain waterborne faecal-oral diseases, driving the implementation of various human interventions for their control and prevention. Taking into account human intervention and temporal variation in rainfall, this paper develops a two-season switching faecal-oral model with impulsive intervention and free boundaries. In this model, the infection fronts are represented by fixed boundaries during the dry season and by moving boundaries during the wet season, with impulsive intervention occurring at the end of each wet season. The simultaneous introduction of impulsive intervention and seasonal switching creates new difficulties for mathematical analysis. We overcome these challenges through novel analytical techniques, resulting in a spreading-vanishing dichotomy and a sharp criteria governing this dichotomy. Finally, numerical simulations are presented to validate the theoretical results and to visually illustrate the influence of seasonal switching and impulsive intervention. Our results mathematically explain that two factors, the duration of the dry season and the intensity of impulsive intervention are both positively correlated with ef
Prescriptive Process Monitoring (PresPM) recommends interventions during running business processes to optimize key performance indicators (KPIs). In realistic settings, interventions are rarely isolated: organizations need to align sequences of interventions to jointly steer the outcome of a case. Existing PresPM approaches only partially address this challenge. Many focus on a single intervention decision, while others treat multiple interventions independently, ignoring how they interact over time. Methods that do address these dependencies depend either on simulation or data augmentation to approximate the process to train a Reinforcement Learning (RL) agent, which may create a reality gap and introduce bias. We introduce SCOPE (Sequential Causal Optimization of Process Interventions), a PresPM approach that learns aligned sequential intervention recommendations. SCOPE employs backward induction to estimate the effect of each candidate intervention action, propagating its impact from the final decision point back to the first. By leveraging causal learners, our method can utilize observational data directly, unlike methods that require constructing process approximations for RL
Estimating the causal effect of a time-dependent treatment on time to death is challenging. In this paper, we formulate the problem using the illness-death model and focus on a stochastic intervention that modifies the hazard governing the transition from no treatment to treatment initiation. Such an intervention can only be implemented at the level of the observed data, whereas the causally valid intervention is defined at the level of the true data-generating process. We provide conditions under which the practically feasible intervention corresponds to the desired causal intervention in the specific setting. We first consider an intervention in which treatment is initiated at a fixed time point, which may subsequently be varied across the relevant time span. However, the resulting estimand is not pathwise differentiable, preventing the development of assumption-lean inference. To address this, we instead consider a smoothed intervention that assigns treatment within a time window around the target time point, thereby yielding a parameter amenable to semiparametric analysis. We derive the corresponding efficient influence function and propose a debiased one-step estimator with de
Neural networks are growing more capable on their own, but we do not understand their neural mechanisms. Understanding these mechanisms' decision-making processes, or mechanistic interpretability, enables (1) accountability and control in high-stakes domains, (2) the study of digital brains and the emergence of cognition, and (3) discovery of new knowledge when AI systems outperform humans. This paper traces how attention head intervention emerged as a key method for causal interpretability of transformers. The evolution from visualization to intervention represents a paradigm shift from observing correlations to causally validating mechanistic hypotheses through direct intervention. Head intervention studies revealed robust empirical findings while also highlighting limitations that complicate interpretation. Recent work demonstrates that mechanistic understanding now enables targeted control of model behaviour, successfully suppressing toxic outputs and manipulating semantic content through selective attention head intervention, validating the practical utility of interpretability research for AI safety.
Intervention is one of the most representative and widely used methods for understanding the internal representations of large language models (LLMs). However, existing intervention methods are confined to linear interventions grounded in the Linear Representation Hypothesis, leaving features encoded along non-linear manifolds beyond their reach. In this work, we introduce a general formulation of intervention that extends naturally to non-linearly represented features, together with a learning procedure that further enables intervention on implicit features lacking a direct output signature. We validate our framework on refusal bypass steering, where it steers the model more precisely than linear baselines by intervening on a non-linear feature governing refusal.
Evolutionary accumulation models (EvAMs), also known as cancer progression models (CPMs), infer dependencies in the order of accumulation of mutations during tumor progression from cross-sectional data. It has been suggested that EvAMs could be used to identify therapeutic targets, but there is no procedure in the literature for how to extract predictions under intervention from these models. A simple approach of conditioning on the absence of a mutation gives incorrect predictions. We address this gap by formalizing what "intervene" means for all currently available EvAM methods (OT, OncoBN, CBN, H-ESBCN, MHN, HyperHMM, HyperTraPS), using Pearl's do operator and conditional interventions. For each model, we show how to implement the intervention (in most cases as specific parameter modifications), identify equivalent implementation procedures, and analyze whether the modularity assumption -- required for the intervention to be well-defined -- is justified. Drawing on individual-level causal DAGs that make fitness an explicit variable, we distinguish two types of intervention (killing and inactivating) that are conflated in standard EvAM representations. Since the goal is to priori
This article provides resources to design effective diversity, equity and inclusion (DEI) interventions in physics settings. It summarizes critiques of DEI interventions and reviews research findings suggesting that DEI interventions show modest but positive benefits. It covers physics-specific concerns around DEI (resistance from skeptical participants, unintended negative consequences for members of minoritized groups). The article then explores six design questions: 1) What are the intended outcomes of the intervention? (with a list of possible outcomes) 2) How will the intervention be justified to participants? 3) Will the intervention be voluntary or required? 4) Who will the participants be? 5) What activities will the intervention consist of? (with suggested activities paired to outcomes) 6) How will you know whether the intervention worked? (with tools used by other researchers to measure particular outcomes, and bonus suggestions for how to carry out publication-worthy evaluations). Each question is explored using research results and recommendations, and illustrated using examples from physics and other STEM settings.
Modern language models (LMs) have significantly advanced generative modeling in natural language processing (NLP). Despite their success, LMs often struggle with adaptation to new contexts in real-time applications. A promising approach to task adaptation is activation intervention, which steers the LMs' generation process by identifying and manipulating the activations. However, existing interventions are highly dependent on heuristic rules or require many prompt inputs to determine effective interventions. This paper proposes a layer-wise additive activation intervention framework that optimizes the intervention process, thus enhancing the sample efficiency. We benchmark our framework on various datasets, demonstrating improvements in the accuracy of pre-trained LMs and competing intervention baselines.
Intervention intuition is often used in model explanation where the intervention effect of a feature on the outcome is quantified by the difference of a model prediction when the feature value is changed from the current value to the baseline value. Such a model intervention effect of a feature is inherently association. In this paper, we will study the conditions when an intuitive model intervention effect has a causal interpretation, i.e., when it indicates whether a feature is a direct cause of the outcome. This work links the model intervention effect to the causal interpretation of a model. Such an interpretation capability is important since it indicates whether a machine learning model is trustworthy to domain experts. The conditions also reveal the limitations of using a model intervention effect for causal interpretation in an environment with unobserved features. Experiments on semi-synthetic datasets have been conducted to validate theorems and show the potential for using the model intervention effect for model interpretation.
Imitation learning techniques have been shown to be highly effective in real-world control scenarios, such as robotics. However, these approaches not only suffer from compounding error issues but also require human experts to provide complete trajectories. Although there exist interactive methods where an expert oversees the robot and intervenes if needed, these extensions usually only utilize the data collected during intervention periods and ignore the feedback signal hidden in non-intervention timesteps. In this work, we create a model to formulate how the interventions occur in such cases, and show that it is possible to learn a policy with just a handful of expert interventions. Our key insight is that it is possible to get crucial information about the quality of the current state and the optimality of the chosen action from expert feedback, regardless of the presence or the absence of intervention. We evaluate our method on various discrete and continuous simulation environments, a real-world robotic manipulation task, as well as a human subject study. Videos and the code can be found at https://liralab.usc.edu/mile .
We propose an algorithm for change point monitoring in linear causal models that accounts for interventions. Through a special centralization technique, we can concentrate the changes arising from causal propagation across nodes into a single dimension. Additionally, by selecting appropriate intervention nodes based on Kullback-Leibler divergence, we can amplify the change magnitude. We also present an algorithm for selecting the intervention values, which aids in the identification of the most effective intervention nodes. Two monitoring methods are proposed, each with an adaptive intervention policy to make a balance between exploration and exploitation. We theoretically demonstrate the first-order optimality of the proposed methods and validate their properties using simulation datasets and two real-world case studies.
As autonomous AI agents move from conversational systems to long-horizon software execution, runtime safety layers that decide when to interrupt an agent have become essential. We study this timing problem using a continuous 18-dimensional affective-dynamics engine (HEART) as a diagnostic probe, evaluating four intervention trigger families - absolute state thresholds, composite state-action patterns, regex reasoning-feature extraction, and zero-shot LLM-as-judge - against human-annotated intervention points on SWE-bench-Verified debugging traces. We report three findings. First, a State Saturation Trap: agents show no recovery signal under sustained difficulty, so modeled frustration quickly crosses the threshold and stays at its maximum, converting threshold-on-state triggers from moment detectors into near-constant indicators that fire on 39-83% of actions across five trajectories. Second, a capability-and-context floor for LLM judges: a small model (gpt-5.4-mini) never fires, while frontier and cross-vendor models escape the zero-firing floor only with full-trajectory context, and even then reach only F1 0.17-0.40 at up to 90x the cost. Third, and most importantly, the supervis
This article discusses the application of stochastic intervention to find the optimal treatment distribution yielding a high value of expected potential outcome under the setting where the number of treatments is allowed to vary with $n$. The primary motivation is to obtain a novel summarization of the effect of various treatments which would guide practitioners towards better decision regarding which intervention to choose.
Causal games are probabilistic graphical models that enable causal queries to be answered in multi-agent settings. They extend causal Bayesian networks by specifying decision and utility variables to represent the agents' degrees of freedom and objectives. In multi-agent settings, whether each agent decides on their policy before or after knowing the causal intervention is important as this affects whether they can respond to the intervention by adapting their policy. Consequently, previous work in causal games imposed chronological constraints on permissible interventions. We relax this by outlining a sound and complete set of primitive causal interventions so the effect of any arbitrarily complex interventional query can be studied in multi-agent settings. We also demonstrate applications to the design of safe AI systems by considering causal mechanism design and commitment.