Publications

2026

1. arXiv

   The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching

   Sankaran Vaidyanathan, David Arbour, Aaron Mueller, and 2 more authors

   arXiv, Jun 2026

   Abstract arXiv Bib PDF

   Activation patching is the primary tool in mechanistic interpretability. It attributes causal responsibility for a model behavior to each of its individual components by estimating its natural indirect effect (NIE). Re-deriving the activation patching estimand from causal mediation analysis, we find that the NIE does not solely capture the causal effect through the specific component. It also contains interaction effects (INT) that measure how much the component’s causal effect itself depends on the state of other components in the model. A natural response may be to try to eliminate INT by adjusting the estimator or unit of analysis, but each of these potential remedies has predictable failure modes. We demonstrate these failure modes in the GPT-2 IOI circuit; components whose causal importance is conditional on the state of other components are either invisible or artificially inflated, and INT variance explains the previously documented instability of faithfulness scores. We prove that INT scales with the distance between clean and patched component activations, is negligible when the model is locally affine, and decomposes combinatorially into pairwise and higher-order group interactions. Despite its inevitability, INT is not a nuisance to be eliminated, but rather a diagnostic for interpretability studies. Its individual and group-level magnitude and sign signal when causal conclusions are prompt-dependent, and when greedy NIE-based component ranking will miss mechanisms only discoverable through combinatorial search.

   @article{vaidyanathan2026curse,
     title = {The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching},
     author = {Vaidyanathan, Sankaran and Arbour, David and Mueller, Aaron and Niekum, Scott and Jensen, David},
     journal = {arXiv},
     year = {2026},
     month = jun,
   }

2. RLC RLBrew

   Hierarchical Experimentalist Agents

   Abhranil Chandra, Sankaran Vaidyanathan, Utsav Dhanuka, and 2 more authors

   RLC Workshop on Reinforcement Learning Beyond Rewards: Towards Scalable General-Purpose Agents, Jun 2026

   Abstract arXiv Bib PDF

   Large language models (LLMs) are increasingly used to take actions in the real world and support human decision-making, yet most agents rely on parametric knowledge, fixed post-training data, retrieval, or search. This paradigm breaks down in novel domains and for sophisticated queries that cannot be answered from prior knowledge alone. Knowing the laws of physics, for instance, does not by itself enable LLMs to answer queries or complete long-horizon tasks in a complex physical system. To address this, we introduce Hierarchical Experimentalist Agents (HExA), an in-context self-improvement framework to learn from active experimentation. HExA iteratively designs and refines query-relevant experiments, learns a reusable library of composable skills from experience, and integrates experimental evidence to answer queries or take actions. HExA is training-free, compatible with any black-box model, and does not require external supervision, oracles, or offline data. To evaluate active experimentation, we introduce Interphyre, a tool-calling benchmark built on the PHYRE 2D procedural physics environment, where agents propose interventions and test hypotheses through simulation APIs. Experiments show that current LLM agents struggle in these settings, especially on the hardest levels of Interphyre. Claude Sonnet 4.6 achieves only 2% success, while HExA improves the same model to up to 77% success. HExA also improves open-weight models and outperforms agentic baselines such as ReAct and Reflexion. Moreover, using only skills learned from easier levels and transferred without active experimentation, HExA achieves 44% success, demonstrating the reusability and generalization of its learned skills. Overall, HExA shows that learning through active experimentation can help agents discover useful knowledge, acquire reusable skills, and make efficient progress on novel long-horizon tasks.

   @article{chandra2026hierarchical,
     title = {Hierarchical Experimentalist Agents},
     author = {Chandra, Abhranil and Vaidyanathan, Sankaran and Dhanuka, Utsav and Gandhi, Varun and Niekum, Scott},
     journal = {RLC Workshop on Reinforcement Learning Beyond Rewards: Towards Scalable General-Purpose Agents},
     year = {2026},
     month = jun,
   }

3. arXiv

   Evaluation Awareness Is Not One Capability: Evidence from Open Language Models

   Nilesh Nayan^*, Aishwarya Sampath Kumar^*, Rishiraj Girmal^*, and 5 more authors

   arXiv, Jun 2026

   Abstract arXiv Bib PDF

   Safety benchmarks assume that test-condition behavior predicts deployment behavior, an assumption that fails if models detect evaluation cues and adapt. This opens a gap between benchmark performance and deployment behavior: compliance measured under test conditions becomes an optimistic upper bound that overstates how safely a model behaves once the evaluation harness is removed. We characterize this evaluation awareness through eight experiments across 37 open-weight models and seven families. (i)Detection is moderate and training-driven (24/37 models exceed chance, best AUROC 0.714 vs.0.819 human, with instruction tuning dominating over scale). (ii)Detection shifts safety behavior (hard refusal drops 5.8 percentage points under hypothetical framing, and 21/140 HarmBench framing effects are significant, with compliance rising up to +30 percentage points. (iii)Representations survive behavioral collapse (probes retain AUROC 0.98 under rewrites that drive behavior below chance, and multi-layer steering causally moves three downstream tasks while random controls do not). (iv)These axes are weakly coupled (only 1/15 correlations are significant, the sole robust link being behavioral detection versus framing resistance, ρ=−0.79, p<0.001). We call this gap the benchmark illusion: because detectability, behavioral manifestation, and controllability vary independently, it is multivariate rather than a single number, so no single awareness score is a reliable proxy for deployment safety.

   @article{nayan2026evaluation,
     title = {Evaluation Awareness Is Not One Capability: Evidence from Open Language Models},
     author = {Nayan, Nilesh and Kumar, Aishwarya Sampath and Girmal, Rishiraj and Anilkumar, Shivani and Vaidyanathan, Sankaran and Palacio, David A. Nader and Ghosh, Reshmi and Srinivasan, Soundararajan},
     journal = {arXiv},
     year = {2026},
     month = jun,
   }

2025

1. NeurIPS Mech Interp

   Detecting and Characterizing Planning in Language Models

   Jatin Nainani, Sankaran Vaidyanathan, Connor Watts, and 2 more authors

   NeurIPS Mechanistic Interpretability Workshop, Dec 2025

   Abstract arXiv Bib PDF

   Modern large language models (LLMs) have demonstrated impressive performance across a wide range of multi-step reasoning tasks. Recent work suggests that LLMs may perform planning - selecting a future target token in advance and generating intermediate tokens that lead towards it - rather than merely improvising one token at a time. However, existing studies assume fixed planning horizons and often focus on single prompts or narrow domains. To distinguish planning from improvisation across models and tasks, we present formal and causally grounded criteria for detecting planning and operationalize them as a semi-automated annotation pipeline. We apply this pipeline to both base and instruction-tuned Gemma-2-2B models on the MBPP code generation benchmark and a poem generation task where Claude 3.5 Haiku was previously shown to plan. Our findings show that planning is not universal: unlike Haiku, Gemma-2-2B solves the same poem generation task through improvisation, and on MBPP it switches between planning and improvisation across similar tasks and even successive token predictions. We further show that instruction tuning refines existing planning behaviors in the base model rather than creating them from scratch. Together, these studies provide a reproducible and scalable foundation for mechanistic studies of planning in LLMs.

   @article{nainani2025detectingcharacterizingplanninglanguage,
     title = {Detecting and Characterizing Planning in Language Models},
     author = {Nainani, Jatin and Vaidyanathan, Sankaran and Watts, Connor and Assis, Andre N. and Rigg, Alice},
     journal = {NeurIPS Mechanistic Interpretability Workshop},
     year = {2025},
     month = dec,
   }

2. ACL GEM

   Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judges

   Aman Singh Thakur^*, Kartik Choudhary^*, Venkat Srinik Ramayapally^*, and 2 more authors

   In Proceedings of the Fourth Workshop on Generation, Evaluation and Metrics (GEM), Jul 2025

   Abstract arXiv Bib HTML PDF

   The LLM-as-a-judge paradigm offers a potential solution to scalability issues in human evaluation of large language models (LLMs), but there are still many open questions about its strengths, weaknesses, and potential biases. This study investigates thirteen models, ranging in size and family, as ‘judge models’ evaluating answers from nine base and instruction-tuned ‘exam-taker models’. We find that only the best (and largest) models show reasonable alignment with humans, though they still differ with up to 5 points from human-assigned scores. Our research highlights the need for alignment metrics beyond percent agreement, as judges with high agreement can still assign vastly different scores. We also find that smaller models and the lexical metric contains can provide a reasonable signal in ranking the exam-taker models. Further error analysis reveals vulnerabilities in judge models, such as sensitivity to prompt complexity and a bias toward leniency. Our findings show that even the best judge models differ from humans in this fairly sterile setup, indicating that caution is warranted when applying judge models in more complex scenarios.

   @inproceedings{thakur2024judging,
     title = {Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judges},
     author = {Thakur, Aman Singh and Choudhary, Kartik and Ramayapally, Venkat Srinik and Vaidyanathan, Sankaran and Hupkes, Dieuwke},
     booktitle = {Proceedings of the Fourth Workshop on Generation, Evaluation and Metrics (GEM)},
     year = {2025},
     month = jul,
     address = {Vienna, Austria and virtual meeting},
     publisher = {Association for Computational Linguistics},
     isbn = {979-8-89176-261-9},
     pages = {404--430},
   }

3. arXiv

   Quantitative LLM Judges

   Aishwarya Sahoo, Jeevana Kruthi Karnuthala, Tushar Parmanand Budhwani, and 9 more authors

   arXiv preprint arXiv:2506.02945, Jul 2025

   Abstract arXiv Bib PDF

   LLM-as-a-judge is a framework in which a large language model (LLM) automatically evaluates the output of another LLM. We propose quantitative LLM judges, which align evaluation scores of existing LLM judges to human scores in a given domain using regression models. The models are trained to improve the score of the original judge by using the judge’s textual evaluation and score. We present four quantitative judges for different types of absolute and relative feedback, which showcases the generality and versatility of our framework. Our framework is more computationally efficient than supervised fine-tuning and can be more statistically efficient when human feedback is limited, which is expected in most applications of our work. We validate these claims empirically on four datasets using two base judges. Our experiments show that quantitative judges can effectively improve the predictive power of existing judges through post-hoc modeling.

   @article{sahoo2025quantitativellmjudges,
     title = {Quantitative LLM Judges},
     author = {Sahoo, Aishwarya and Karnuthala, Jeevana Kruthi and Budhwani, Tushar Parmanand and Agarwal, Pranchal and Vaidyanathan, Sankaran and Siu, Alexa and Dernoncourt, Franck and Healey, Jennifer and Lipka, Nedim and Rossi, Ryan and Bhattacharya, Uttaran and Kveton, Branislav},
     journal = {arXiv preprint arXiv:2506.02945},
     year = {2025},
   }

2024

1. Neural Networks

   Data-driven learning of chaotic dynamical systems using Discrete-Temporal Sobolev Networks

   Connor Kennedy, Trace Crowdis, Haoran Hu, and 2 more authors

   Neural Networks, Jul 2024

   Abstract Bib HTML PDF

   We introduce the Discrete-Temporal Sobolev Network (DTSN), a neural network loss function that assists dynamical system forecasting by minimizing variational differences between the network output and the training data via a temporal Sobolev norm. This approach is entirely data-driven, architecture agnostic, and does not require derivative information from the estimated system. The DTSN is particularly well suited to chaotic dynamical systems as it minimizes noise in the network output which is crucial for such sensitive systems. For our test cases we consider discrete approximations of the Lorenz-63 system and the Chua circuit. For the network architectures we use the Long Short-Term Memory (LSTM) and the Transformer. The performance of the DTSN is compared with the standard MSE loss for both architectures, as well as with the Physics Informed Neural Network (PINN) loss for the LSTM. The DTSN loss is shown to substantially improve accuracy for both architectures, while requiring less information than the PINN and without noticeably increasing computational time, thereby demonstrating its potential to improve neural network forecasting of dynamical systems.

   @article{kennedy2024data,
     title = {Data-driven learning of chaotic dynamical systems using Discrete-Temporal Sobolev Networks},
     author = {Kennedy, Connor and Crowdis, Trace and Hu, Haoran and Vaidyanathan, Sankaran and Zhang, Hong-Kun},
     journal = {Neural Networks},
     pages = {106152},
     year = {2024},
     publisher = {Pergamon},
   }

2. arXiv

   Automated Discovery of Functional Actual Causes in Complex Environments

   Caleb Chuck^*, Sankaran Vaidyanathan^*, Stephen Giguere, and 3 more authors

   arXiv preprint arXiv:2404.10883, Jul 2024

   Abstract arXiv Bib PDF

   Reinforcement learning (RL) algorithms often struggle to learn policies that generalize to novel situations due to issues such as causal confusion, overfitting to irrelevant factors, and failure to isolate control of state factors. These issues stem from a common source: a failure to accurately identify and exploit state-specific causal relationships in the environment. While some prior works in RL aim to identify these relationships explicitly, they rely on informal domain-specific heuristics such as spatial and temporal proximity. Actual causality offers a principled and general framework for determining the causes of particular events. However, existing definitions of actual cause often attribute causality to a large number of events, even if many of them rarely influence the outcome. Prior work on actual causality proposes normality as a solution to this problem, but its existing implementations are challenging to scale to complex and continuous-valued RL environments. This paper introduces functional actual cause (FAC), a framework that uses context-specific independencies in the environment to restrict the set of actual causes. We additionally introduce Joint Optimization for Actual Cause Inference (JACI), an algorithm that learns from observational data to infer functional actual causes. We demonstrate empirically that FAC agrees with known results on a suite of examples from the actual causality literature, and JACI identifies actual causes with significantly higher accuracy than existing heuristic methods in a set of complex, continuous-valued environments.

   @article{chuck2024automated,
     title = {Automated Discovery of Functional Actual Causes in Complex Environments},
     author = {Chuck, Caleb and Vaidyanathan, Sankaran and Giguere, Stephen and Zhang, Amy and Jensen, David and Niekum, Scott},
     journal = {arXiv preprint arXiv:2404.10883},
     year = {2024},
   }

3. arXiv

   Adaptive Circuit Behavior and Generalization in Mechanistic Interpretability

   Jatin Nainani^*, Sankaran Vaidyanathan^*, AJ Yeung, and 2 more authors

   arXiv preprint arXiv:2411.16105, Jul 2024

   Abstract arXiv Bib PDF

   Mechanistic interpretability aims to understand the inner workings of large neural networks by identifying circuits, or minimal subgraphs within the model that implement algorithms responsible for performing specific tasks. These circuits are typically discovered and analyzed using a narrowly defined prompt format. However, given the abilities of large language models (LLMs) to generalize across various prompt formats for the same task, it remains unclear how well these circuits generalize. For instance, it is unclear whether the models generalization results from reusing the same circuit components, the components behaving differently, or the use of entirely different components. In this paper, we investigate the generality of the indirect object identification (IOI) circuit in GPT-2 small, which is well-studied and believed to implement a simple, interpretable algorithm. We evaluate its performance on prompt variants that challenge the assumptions of this algorithm. Our findings reveal that the circuit generalizes surprisingly well, reusing all of its components and mechanisms while only adding additional input edges. Notably, the circuit generalizes even to prompt variants where the original algorithm should fail; we discover a mechanism that explains this which we term S2 Hacking. Our findings indicate that circuits within LLMs may be more flexible and general than previously recognized, underscoring the importance of studying circuit generalization to better understand the broader capabilities of these models.

   @article{nainani2024adaptive,
     title = {Adaptive Circuit Behavior and Generalization in Mechanistic Interpretability},
     author = {Nainani, Jatin and Vaidyanathan, Sankaran and Yeung, AJ and Gupta, Kartik and Jensen, David},
     journal = {arXiv preprint arXiv:2411.16105},
     year = {2024},
   }

4. JACS

   Assessing Intraoperative Cognitive Workload by Leveraging Deep Learning Networks

   Jake Awtry, Sankaran Vaidyanathan, Heather M Conboy, and 6 more authors

   Journal of the American College of Surgeons, Oct 2024

   Abstract DOI Bib

   Surgeons’ cognitive workload (CWL) fluctuates in response to intraoperative events and cognitive overload may negatively impact operative performance and outcomes. We sought to use a deep neural network model to predict surgeons’ CWL during coronary artery bypass grafting (CABG). The root mean square of successive differences (RMSSD), a heart rate variability metric that reflects CWL, was collected via 3-lead electrocardiogram monitors and Kubios software for surgeons during non-emergent CABG procedures (n = 26). RMSSD was predicted at 5-minute intervals throughout operation via a long short-term memory (LSTM) neural network integrating time, surgical phase, and the RMSSD of surgeons at previous timepoints. Predictions were compared with a random model, linear ridge regression, and a simple autoregressive model in which RMSSD for the surgeon at time interval t equals RMSSD at interval t-1. The LSTM, linear ridge regression, and autoregressive models all performed similarly in predicting dynamic changes in surgeon RMSSD while outperforming the random model. Correlation coefficients for measured and predicted RMSSD values for all 3 models across all cases were 0.47, 0.48, and 0.49, respectively, compared with 0.03 for the random model, and indistinguishable from one another. Shapley additive explanations (SHAP) analysis revealed that a surgeon’s RMSSD at t-1 was the dominant predictor of RMSSD at time t across the range of RMSDD values. The deep LSTM model converged toward, and did not outperform, an autoregressive model, suggesting sustained trends in intraoperative surgeon CWL that would otherwise be difficult to effectively model with machine learning.

   @article{awtry2024leveraging,
     title = {Assessing Intraoperative Cognitive Workload by Leveraging Deep Learning Networks},
     author = {Awtry, Jake and Vaidyanathan, Sankaran and Conboy, Heather M and Kennedy-Metz, Lauren and Clarke, Lori A and Avrunin, George and Dias, Roger and Jensen, David and Zenati, Marco},
     volume = {239},
     issn = {1879-1190},
     url = {http://dx.doi.org/10.1097/XCS.0000000000001159},
     doi = {10.1097/xcs.0000000000001159},
     number = {5},
     journal = {Journal of the American College of Surgeons},
     publisher = {Ovid Technologies (Wolters Kluwer Health)},
     year = {2024},
     month = oct,
     pages = {S71-S79},
   }

2020

1. Complex Networks

   A new measure of modularity in hypergraphs: Theoretical insights and implications for effective clustering

   Tarun Kumar^*, Sankaran Vaidyanathan^*, Harini Ananthapadmanabhan, and 2 more authors

   In Complex Networks and Their Applications VIII: Volume 1 Proceedings of the Eighth International Conference on Complex Networks and Their Applications COMPLEX NETWORKS 2019 8, Oct 2020

   Abstract Bib HTML

   Many real-world systems consist of entities that exhibit complex group interactions rather than simple pairwise relationships; such multi-way relations are more suitably modeled using hypergraphs. In this work, we generalize the framework of modularity maximization, commonly used for community detection on graphs, for the hypergraph clustering problem. We introduce a hypergraph null model that can be shown to correspond exactly to the configuration model for undirected graphs. We then derive an adjacency matrix reduction that preserves the hypergraph node degree sequence, for use with this null model. The resultant modularity function can be maximized using the Louvain method, a popular fast algorithm known to work well in practice for graphs. We additionally propose an iterative refinement over this clustering that exploits higher-order information within the hypergraph, seeking to encourage balanced hyperedge cuts. We demonstrate the efficacy of our methods on several real-world datasets.

   @inproceedings{kumar2020new,
     title = {A new measure of modularity in hypergraphs: Theoretical insights and implications for effective clustering},
     author = {Kumar, Tarun and Vaidyanathan, Sankaran and Ananthapadmanabhan, Harini and Parthasarathy, Srinivasan and Ravindran, Balaraman},
     booktitle = {Complex Networks and Their Applications VIII: Volume 1 Proceedings of the Eighth International Conference on Complex Networks and Their Applications COMPLEX NETWORKS 2019 8},
     pages = {286--297},
     year = {2020},
     organization = {Springer International Publishing},
   }

2. Appl. NetSci

   Hypergraph clustering by iteratively reweighted modularity maximization

   Tarun Kumar, Sankaran Vaidyanathan, Harini Ananthapadmanabhan, and 2 more authors

   Applied Network Science, Oct 2020

   Abstract Bib HTML

   Learning on graphs is a subject of great interest due to the abundance of relational data from real-world systems. Many of these systems involve higher-order interactions (super-dyadic) rather than mere pairwise (dyadic) relationships; examples of these are co-authorship, co-citation, and metabolic reaction networks. Such super-dyadic relations are more adequately modeled using hypergraphs rather than graphs. Learning on hypergraphs has thus been garnering increased attention with potential applications in network analysis, VLSI design, and computer vision, among others. Especially, hypergraph clustering is gaining attention because of its enormous applications such as component placement in VLSI, group discovery in bibliographic systems, image segmentation in CV, etc. For the problem of clustering on graphs, modularity maximization has been known to work well in the pairwise setting. Our primary contribution in this article is to provide a generalization of the modularity maximization framework for clustering on hypergraphs. In doing so, we introduce a null model for graphs generated by hypergraph reduction and prove its equivalence to the configuration model for undirected graphs. The proposed graph reduction technique preserves the node degree sequence from the original hypergraph. The modularity function can be defined on a thus reduced graph, which can be maximized using any standard modularity maximization method, such as the Louvain method. We additionally propose an iterative technique that provides refinement over the obtained clusters. We demonstrate both the efficacy and efficiency of our methods on several real-world datasets.

   @article{kumar2020hypergraph,
     title = {Hypergraph clustering by iteratively reweighted modularity maximization},
     author = {Kumar, Tarun and Vaidyanathan, Sankaran and Ananthapadmanabhan, Harini and Parthasarathy, Srinivasan and Ravindran, Balaraman},
     journal = {Applied Network Science},
     volume = {5},
     number = {1},
     pages = {52},
     year = {2020},
     publisher = {Springer International Publishing Cham},
   }