Causal Interventions in Language Models

Follow this topic Ask about this topic

This field asks which internal states, features, heads, neurons, or context elements actually influence a language model’s decisions, rather than merely correlate with them. Researchers use counterfactual prompt and evidence edits, activation patching and ablation, causal tracing, sparse-autoencoder features, and inference-time steering or weight edits to test mechanisms and control behavior. Applications include checking whether chain-of-thought and retrieved evidence drive answers, and modifying refusal, hallucination, bias, tool use, or reasoning while preserving unrelated capabilities. Results show that readable representations are often not causally used or easily steered; effects vary by layer, task, model, and context, while redundancy and collateral effects make reliable, selective interventions a continuing challenge.

Latest papers

488 papers83 papers in the last four weeks, up 22% on the four weeks before. 0.5% of all new papers.

Closed Gates Leak: Decomposing Adjacent-Layer Cancellation in Gated Language Models
Oct 28, 2026Harvey Dam, Ganesh GopalakrishnanTransformer FFNs Multilayer Perceptrons
Kahlert School of Computing, University of Utah, Salt Lake City, USA.
Interventional Training Reduces Causal Bypass in In-Context Introspection
Oct 28, 2026Zilu Tang, Jongin Kim, Kerem Sahin +3Language Model Introspection Causal Interventions in Language Models
Boston University · Northeastern University · Monash University Indonesia
Can LLMs Faithfully Solve Ontology Engineering Tasks? A Mechanistic Interpretability Study
Oct 28, 2026Bohui Zhang, Albert Meroño-Peñuela, Elena SimperlMechanistic Interpretability Causal Interventions in Language Models
King’s College London, United Kingdom · Technical University of Munich, Germany
From Error Prediction to Causal Influence: Auditing SAE Features in Multiple-Choice QA
Oct 27, 2026Aria Nourbakhsh, Adelaide Danilov, Oleksandr Marchenko Breneur +1Sparse Autoencoders Causal Interpretability
Department of Computer Science (DCS), University of Luxembourg
Causal Localization of Evidence Effects in Transformers
Oct 27, 2026Jipeng Li, Pramod Khargonekar, Yanning ShenTransformer Interpretability Attention Head Analysis
EXCISE: Locating and Excising Distracting Sentences in Medical Question Answering
Oct 26, 2026Yerim Hong, Sohyun Son, Heeseung Chung +1Language Model Robustness Clinical QA
Kyung Hee University
Why Are LLMs Confidently Wrong? Correcting Overconfident Errors via Causal Head Intervention
Oct 26, 2026Jing Ren, Bowen Li, Ziqi Xu +2Language Model Calibration Confidence Estimation in Language Models
RMIT University
Intervention-Certified Amplification: A Validity Gate for $σ$ -Weighted Post-Training
Oct 24, 2026Leiyue Chen, Meng Wei, Xuefeng Hu +1RL Post-Training RL for Language Model Reasoning
ByteDance
Many Perturbations, Shared Failure Dynamics: Tracing Direction-Sensitive Perturbation Influence in Language Models
Oct 23, 2026Zhibo Hu, Chen Wang, Yanfeng Shu +3Language Model Robustness Faithfulness of Language Model Reasoning
The University of New South Wales Australia · CSIRO technology Australia
REACT: Representational Early Analysis and Control of Chain-of-Thought Trajectories
Oct 23, 2026Hanwen Zhang, Enyi Jiang, Ruike Zhu +2LLM Safety Causal Interventions in Language Models
Independent Researcher · Stanford University · University of Illinois Urbana-Champaign
Context or Knowledge: How LLMs Attribute Their Own Answers
Oct 23, 2026Kirill Afendulev, Alexey Dontsov, Alla Chepurova +2Claim Attribution in Language Models Data Provenance in Language Models
Concentrated Causal Influence, Distributed Diagnostic Information: Cognitive Impairment Decisions in Large Audio-Language Models
Oct 23, 2026Xiutian Zhao, Luqi Sun, Björn Schuller +1Audio-Language Models Circuits in Language Models
Center for Language and Speech Processing (CLSP), Johns Hopkins University, USA · Group on Language, Audio & Music (GLAM), Imperial College London, UK
Causal Chain of Thought: Mitigating Attention Leakage via Context Isolation
Oct 22, 2026Gabriele Dominici, Teague McMillan, Fazl Barez +2Structured Reasoning in Language Models Faithfulness of Language Model Reasoning
Università della Svizzera Italiana · ETH Zurich · University of Oxford
Learning to Be Monitored: Causal Bounds and Intervention-Aware Post-Training Recipes for Chain-of-Thought Monitorability
Oct 22, 2026Qiang Fu, Xinyue Li, Andre Wibisono +1CoT Monitoring LLM Safety Monitoring
Yale University
Spectral Smoothing and Spiking Perturbations Reconfigure Capability Profiles and Selective Task Performance in Large Language Models
Oct 22, 2026Qince Luan, Jianhui Xu, Yang XuCausal Interventions in Language Models LLM Interpretability
Southern University of Science and Technology (SUSTech) · Independent Scholar
Interpreting Model Reasoning through Behavioral Counterfactuals
Oct 22, 2026Shawn Im, James Oldfield, Adel Bibi +1Causal Interventions in Language Models Faithfulness of Language Model Reasoning
University of Wisconsin–Madison · University of Oxford
WRDO: Weighted Refusal Direction Optimization in Aligned Language Models
Oct 21, 2026Grigoriy Evseev, Vladimir Solodkin, Dmitry Bylinkin +3Refusal Behavior in Language Models Refusal Mechanisms in Language Models
Moscow Independent Research Institute of Artificial Intelligence · Basic Research of Artificial Intelligence Laboratory (BRAIn Lab) · T-Technologies +1
Language Impairment as a Lens onto Large Language Models
Oct 20, 2026Muxuan He, Wang Bill Zhu, Yue Chen +5LLM Interpretability Circuits in Language Models
University of Southern California · University of California, Berkeley · Independent Researcher +1
Hidden Messages Between Diffusion Language Model Agents: Causal Evidence for Latent Communication
Oct 20, 2026Alan-Barsag Gazzaev, Alexey Gavrilov, Sergey Muravyov +1Multi-Agent LLMs Diffusion Language Models
ITMO University · HSE University
Can Language Models Get "Aphasia" From Neuronal Ablation?
Oct 20, 2026Pratik Worah, Randall BatemanCausal Interventions in Language Models Disease Progression Modeling
Strength Controls Change How Interventions Compare Across Sparse Autoencoder Codes
Oct 19, 2026Jinchen Zhang, Zheyuan Liu, Yihan Zhu +1Transformer Interpretability Sparse Autoencoders
University of Notre Dame
Learning to Reason Efficiently with Interventional Step Credit
Oct 19, 2026Guanda Li, Xuefeng Bai, Yulong Chen +3Efficient Language Model Reasoning Step-Level Credit Assignment in RL
Harbin Institute of Technology, Shenzhen · University of Aberdeen · University of Cambridge
On Reasoning Correctness Directions
Oct 18, 2026Munachiso Samuel Nwadike, Zayd M. K. Zuhri, Rifo Genadi +5Causal Interpretability Activation Steering
MBZUAI, UAE · RIKEN AIP, Japan · Mila, Canada +1
One Head Away from Language Confusion: Single Attention Heads Across Model Families
Oct 18, 2026Jiwon Jeong, Anjelo Jann Laroza, Chaewoon Yoon +4Multilingual Language Models Attention Head Analysis
Lotte Innovate · Mapúa University · University of Wisconsin–Madison +4
WAQS: Weight Absorption for Quadratic Probing and Affine Steering
Oct 17, 2026Ding Zhu, Mohammad Mahdi KhaliliLanguage Model Steering Activation Steering
Department of Computer Science and Engineering, the Ohio State University, Columbus, OH 43210, USA
Many Biases, One Circuit: Post-Training Tunes a Pretrained Cue-Routing Mechanism
Oct 16, 2026Prakhar Gupta, Terry Jingchen Zhang, Zhijing JinCognitive Bias in Language Models Circuits in Language Models
University of Michigan · Jinesis Lab, University of Toronto & Vector Institute · University of Oxford +2
SAFETYEDIT: Sparse Internal Targets with Prompt-Wide Control for LLM Jailbreak Backdoors
Oct 15, 2026Yutong Ke, Yansu Wang, Zeyu Zhang +3LLM Backdoor Attacks Language Model Editing
Duke Kunshan University · Xi’an Jiaotong-Liverpool University · Wuhan University
Measuring Mechanistic Uncertainty in Large Language Models via Causal Robustness
Oct 15, 2026Aneesh Komanduri, Majid Afshar, Timothy A. Miller +1LLM Uncertainty Estimation Causal Interventions in Language Models
University of Colorado Anschutz Medical Campus · University of Wisconsin-Madison · Harvard Medical School +1
Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures
Oct 15, 2026Cesare Zavattari, Alessandro Tommasi, Giuseppe PrencipeLLM Agents Failure Attribution for LLM Agents
Dipartimento di Informatica, Università di Pisa, Pisa, Italy
Emotion Is Not External Knowledge: Internal Attention Signals for Affective State Reasoning
Oct 15, 2026Wenqiu Tang, Zhen Wan, Yihui Wang +2Emotion Representation in Language Models Emotion Recognition
Nagoya University, Nagoya, Aichi, Japan · National Institute of Informatics, Tokyo, Japan

Oct 28, 2026J/K move · Enter open · S save

Closed Gates Leak: Decomposing Adjacent-Layer Cancellation in Gated Language Models

Harvey Dam, Ganesh Gopalakrishnan

Kahlert School of Computing, University of Utah, Salt Lake City, USA.

Transformer FFNs Multilayer Perceptrons Transformer Training Dynamics Causal Interventions in Language Models

Open paper AskPDF arXiv

In transformer language models with gated multilayer perceptrons (MLPs), adjacent layers write partly against each other into the residual stream. Previous work interpreted this cancellation as layers proposing features and the next layer selectively rejecting the inappropriate ones. We decompose the cancellation by removing the previous layer's write from the input of the next MLP: what the next layer writes along that write splits into what it writes anyway and how much less its MLP writes because the write is present, which we call the damping, and we trace the damping to the neurons that carry it. In gated MLPs the damping is carried by neurons whose gates are closed. SiLU and GELU are slightly negative below zero, so a closed gate still passes a small negative multiple of what the neuron reads, and a neuron whose output weight points along its input weight then subtracts a fraction of the direction it read. Summed over such neurons, closed gates contract the average direction a gated MLP reads, and the previous layer's write is not singled out: the same layer's write at another token is damped as much or more, and the eigenvalue signature from which that work concluded that the correction is selective appears for another token's write as well. Over training the damping grows while what the next layer writes anyway stays flat, and cancellation begins when the damping overtakes it; annealing the negative side of the gate activation away in continued pretraining removes the damping and the cancellation. Adjacent-layer cancellation is therefore weak evidence about which features a model rejects: it is produced by a contraction that gated MLPs apply, on average over directions, to whatever they read.