acceptodds
Under review as a conference paper at ICLR 2027

Causal Intervention on Latent Confounding Bias Neurons in Large Language Models

Abstract

Pre-training corpora of large language models (LLMs) often contain latent confounding information, which can be encoded into model parameters and induce confounder-related neuronal activations during inference. Existing causal debiasing methods for LLMs predominantly rely on the front-door adjustment criterion, which assumes that latent confounders do not directly affect the reasoning process. In practice, however, latent confounders may also affect the reasoning process, violating the assumptions required for valid front-door adjustment. To address this limitation, we propose CILC (Causal Intervention on Latent Confounding Bias Neurons), a neuron-level causal intervention framework for mitigating latent confounding bias in LLMs. Specifically, CILC first employs a Variational Autoencoder (VAE) to infer latent representations from the data and maps these representations into the LLM’s input embedding space via a latent-to-soft-prompt transformation. It then constructs contrastive inputs with and without the resulting latent soft prompts and applies Backward Integrated Gradients (BIG) to localize neurons that are particularly sensitive to latent confounding information. Finally, CILC performs targeted causal interventions on the identified neurons to attenuate the influence of latent confounding on model predictions. Extensive experiments on three LLM backbones and eight real-world datasets demonstrate that CILC effectively mitigates latent confounding bias while consistently improving model performance. Our source code is available at https://anonymous.4open.science/r/CILC-3E63.

Primary area:
interpretability and explainable AI
Keywords:
Large language models, Latent Confounding Bias, LLM Debiasing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_81AWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes CILC, a method to mitigate latent confounding bias in LLMs by intervening on responsible neurons. CILC infers a latent representation of the input with a VAE aiming to encode the latent confounding, then maps it into soft prompts, identifies the neurons most sensitive to these soft prompts with integrated gradients attribution, and scales down their activations at inference time. The method is motivated by the argument that front-door-based debiasing methods are invalid when latent confounders also affect the reasoning process. Experiments on recommendation tasks (on top of TALLRec and CoLLM) and reasoning benchmarks (on top of ACPS) show consistent accuracy improvements.

Critical strengths
  • The paper is very well written and easy to follow. Fig. 2 is well crafted.
  • Terminology is precise, and notation is clear.
  • The motivation is convincing: latent confounders encoded in the model parameters can also affect the reasoning process, which violates the assumptions of front-door adjustment.
Critical weaknesses

Major

  • W1. The main text only reports accuracy. The RQ3 and an important causal analysis is all deferred to the appendix (Sec. 4.3), as is the related work. Studies beyond accuracy are important to support the main claim, so in my opinion they belong in the main text, and the appendix should only complement them. I understand these analyses could be moved to the main text if the extra camera-ready page is allowed, but the content was not well organized within the 9-page limit at submission time.

  • W2. It seems to me the paper misses convincing evidence that the improvement comes from mitigating latent confounding. The ablations (Tab. 4-5) only change which neurons are intervened on, while the rest of the pipeline is kept fixed. They show that the BIG-selected neurons matter, but not that the latent confounder does. Without such controls, the gains could equally come from other elements in the pipeline. Some results also raise some concerns:

    • In Fig. 3, the AUC at γ = 1 (i.e., the intervention leaves the activations unchanged) is about 0.686 on movie and 0.580 on book, well above TALLRec (0.672 and 0.564, Tab. 1). Unless I am missing something, most of the gain would then not come from the intervention. Please correct me if I misunderstood this point.
    • The ATE is measured on the injected soft prompt, which is exactly what the neurons are selected for, so a positive ΔATE (Tab. 6) is somewhat expected by construction. Moreover, even for inputs without the latent proxy, the intervention increases the average predicted probability by about 0.05-0.06 (Tab. 7), i.e., more than twice the effect of the proxy itself (about 0.02). This suggests that the intervention changes the model's predictions well beyond the confounding effect it targets, right?
  • W3. A synthetic controlled experiment is missing. The core claim is that CILC addresses the failure of front-door adjustment, yet CILC is never compared with front-door methods in a very controlled synthetic experiment where one can clearly see front-door failing and the proposed method succeeding. Given the length of the pipeline (VAE, soft-prompt mapping, BIG, intervention), a synthetic experiment with a known latent confounder affecting the mediator would help convince the reader.

  • W4. Uncertainties are missing. All experiments are repeated over five seeds (Sec. 4.1), but only means are reported. Standard deviations or statistical tests are needed to assess whether the gains are significant.

Minor

  • W5. The interpretation of γ in App. E.3 seems reversed: since the intervention sets HZ=γHZH_Z = \gamma H_Z, a small γ corresponds to a strong suppression, not an insufficient one.
Questions for authors
  • Q1. Could the authors expand on the choice of a VAE to extract the input representation?
  • Q2 (related to W2) Could the authors demonstrate that the improvements originate from the effect of the latent confounders?
  • Q3 (related to W2). Is it possible, at least in principle, to design a synthetic experiment where the confounder is known? How would the proposed method compare to prior in this setting? What could be measured?

Summarizing, while I find the paper well presented and well motivated, the lack of evidence that the improvements stem from mitigating latent confounding prevents me from recommending acceptance.

Reviewer_CQB9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper aims to address the problem of latent confounding bias in LLMs, where unobserved confounders in pretraining corpora are encoded into model parameters and induce biased predictions during inference. The proposed framework, CILC, first uses a VAE to infer a proxy representation of the latent confounder from observed user-history inputs, maps this representation into soft prompts via an MLP, and constructs contrastive inputs. Backward Integrated Gradients (BIG) is then applied to localize the neurons most sensitive to the latent confounding information. Finally, during inference, CILC attenuates the influence of these bias-related neurons by scaling their activations by a factor γ\gamma. Experiments across three LLM backbones and eight real-world datasets demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The problem of latent confounding bias in LLMs is interesting and practically relevant.

  2. The critique of existing front-door-adjustment-based methods is well-reasoned: their validity relies on the assumption that latent confounders do not affect the reasoning mediator, which may not hold in LLMs. The proposed direction of explicitly learning and intervening on the latent confounders is well-motivated, though the learnability of the latent confounder itself remains debatable.

  3. The experimental evaluation is comprehensive, spanning eight real-world datasets and three LLM backbones.

Critical weaknesses
  1. The identifiability of latent confounders is the most serious concern. The paper simply uses a VAE to learn a latent representation Z from the observed input X, without any theoretical guarantee that Z captures the true latent confounders. In causal inference and causal representation learning, identifying and learning latent confounders from observational data is highly non-trivial and relies on strong assumptions (e.g., proxy variables [1, 2], linear models [3, 4]) that are not satisfiable and verifiable in the LLM setting. That is, Assumption 1 is not enough for the identification of latent confounders and too strict to be satisfied in the LLM setting. Therefore, in practical applications, it is unclear whether Z actually encodes the confounding information, or merely some arbitrary latent factors correlated with X.

  2. The intervention strategy assumes that latent confounding is concentrated in a small subset of neurons (Assumption 2), which may not hold in practice. In practice, latent confounding may be distributed across many neurons, and intervening only on the top-K most sensitive neurons may leave substantial residual bias. Conversely, scaling all neurons by γ\gamma would risk degrading task-relevant information. The paper does not discuss how to balance debiasing effectiveness against information preservation, nor does it provide evidence that Assumption 2 holds.

  3. The training procedure for the latent-to-soft-prompt mapping module gψg_\psi is not specified. The paper does not describe how gψg_\psi is optimized, including its loss function, whether it is trained jointly with the VAE or separately, and what supervision signal drives the mapping.

  4. Assumption 3 and Assumption 4 are also strict. The authors should clarify why they could be satisfied in the LLM setting.


Reference

[1] Rissanen, Severi, and Pekka Marttinen. "A critical look at the consistency of causal estimation with deep latent variable models." Advances in Neural Information Processing Systems 34 (2021): 4207-4217.

[2] Miao, Wang, et al. "A confounding bridge approach for double negative control inference on causal effects." Statistical Theory and Related Fields 8.4 (2024): 262-273.

[3] Song, Xiangchen, et al. "LLM interpretability with identifiable temporal-instantaneous representation." Advances in Neural Information Processing Systems 38 (2026): 103963-104000.

[4] Li, Zijian, et al. "Towards identifiability of hierarchical temporal causal representation learning." Advances in Neural Information Processing Systems 38 (2026): 19680-19719.

Questions for authors

Please see Weaknesses.

Reviewer_XSn8OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that front-door debiasing methods for LLMs fail when hidden confounders affect the reasoning step. It proposes CILC: a VAE summarises user history into a latent vector, which is mapped to soft-prompt tokens; neurons that respond most to these tokens (found with integrated gradients) are labelled bias neurons and scaled down at inference. It reports small accuracy gains on recommendation and reasoning benchmarks.

Critical strengths
  1. Reasonable motivation. The point that front-door methods assume the reasoning step is unaffected by hidden confounders is fair, and a method that doesn't need an explicit reasoning step could apply more widely.

  2. Broad, consistent evaluation. Eight datasets, three backbones, recommendation and reasoning tasks, with consistent (if small) gains over strong baselines.

Critical weaknesses
  1. The method doesn't isolate confounding. Bias neurons are found by contrasting soft prompts from the user's latent with soft prompts from a zero latent (the average user). That isn't a valid counterfactual: it changes user-specific information, including the genuine preference a recommender should use, rather than varying a confounder with the evidence held fixed. The soft prompts are also inputs the LLM never saw, and the main text doesn't say how the mapping network is trained, so the neurons found may just react to unusual input. The confounder is never named or measured, so the "causal" framing isn't earned; the method is closer to attribution-guided neuron ablation. A contrast that varies a known confounder with the user's evidence fixed would make the claim testable.

  2. The results don't show that confounding is reduced. The main text reports only accuracy gains, which are small and notably uniform (about +0.01 in every setting, including all four very different reasoning tasks), with no variance or significance tests across the five seeds. Accuracy on standard splits isn't evidence of deconfounding: if the test data share the training bias, removing it could lower accuracy, and gains could instead come from regularisation or tuning. Metadata-based debiasing methods are discussed but not compared against.

Questions for authors
  1. How is the latent-to-soft-prompt MLP trained? Do the soft prompts lie within the normal embedding distribution?
  2. What does the VAE summarise on the reasoning datasets, which have no user history?
  3. How are the scaling strength and the number of neurons chosen?
  4. Why use soft prompts rather than contrasting real prompts with and without user history, or adding the latent to the residual stream? Have you ablated each component?
  5. Does mean ablation give the same results as scaling towards zero?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.