Introspective Coupling: Self-Explanation Training Tracks Behavioral Change Despite Fixed Supervision

Abstract: When does training LMs to generate explanations of their predictions yield faithful introspection, rather than superficial imitation? We study LMs trained to explain which features of their inputs influenced their behavior, using models' counterfactual behavior on modified inputs as supervision. Surprisingly, we find that LMs trained on fixed counterfactual explanations derived from earlier checkpoints of themselves, or even from behaviorally similar models in different families, frequently produce explanations more faithful to their own current behaviors than to those of their training targets. This "introspective" coupling between LM explanations and behaviors occurs when training explanations remain sufficiently correlated with current behaviors over the course of training, even as behaviors themselves shift. We also show that introspective coupling tracks behavior shifts: when explanation training is provided concurrently with other post-training objectives, explanations track those shifts without requiring updated supervision. This phenomenon appears in multiple tasks, including sycophancy and refusal, and is robust to label noise. Overall, our results show that even fixed datasets of counterfactual explanations can provide scalable and generalizable post-training signal for introspection.

• The paper establishes that fixed self-explanation supervision paired with behavioral regularization enables models to generate explanations that accurately track evolving behaviors.
• Empirical findings reveal a 'Self > Orig' effect where explanation predictions align more closely with current model outputs than with original labels.
• Mechanistic analysis, including activation patching with a strong correlation (r=0.89), confirms that the same internal circuits drive both behavior and explanation.

Introspective Coupling in LLM Self-Explanation Training

Problem Formulation and Motivation

"Introspective Coupling: Self-Explanation Training Tracks Behavioral Change Despite Fixed Supervision" (2606.32038) investigates the fidelity of self-explanation systems in LLMs when explanation supervision is fixed but underlying behaviors shift. The central question is whether training LMs to explain their predictions using counterfactual examples and explanations from earlier checkpoints, or even different model families, results in faithful introspection—i.e., explanations that accurately track the model's current behavioral policy, instead of merely imitating static label distributions.

Traditional fine-tuning paradigms in explanation training depend on static sets of explanation labels from an initial model snapshot. However, post-training objectives, distribution shifts, and behavior drift can result in misalignment between the supervised explanation and the model's current policy, potentially undermining interpretability and oversight. This paper presents an empirical, mechanistic, and theoretical analysis of this mismatch, highlighting the emergence of a phenomenon termed introspective coupling, where explanations consistently track the model’s current behavior more faithfully than the original behaviors used for supervision—even without updated labels.

Figure 1: Method overview illustrating how behaviors and explanations are sampled, explanations are predicted after drift, and introspective coupling is identified by comparing explanation faithfulness against original and current behaviors.

Methodology

The study follows a counterfactual simulability framework for explanation evaluation: given an input $x$ and cue $C$, the model is assessed for how its answer changes upon ablation ($x_{\setminus C}$) and for its ability to articulate this dependence. Training uses SFT with explanation labels constructed from initial behaviors, regularized via KL-divergence to align behaviors with the original checkpoint.

The main evaluation metrics are:

• Explanation Exact Match (EM): Agreement between generated explanations and the ground-truth explanation label, comparing both against the original label (Orig) and the current self label (Self).
• Behavioral EM: Measures behavioral drift by quantifying agreement between original and current behaviors on both cue-included and cue-ablated inputs.

Three core tasks probe sycophancy (Hint-MMLU, AITA) and refusal behaviors using naturalistic and synthetic prompts. Detailed experimental setups manipulate regularization intensity, explanation label provenance, and auxiliary behavioral shifts to dissect the conditions under which introspective coupling emerges.

Strong Empirical Findings: Self > Orig Emergence

A robust outcome across all tasks and model variants is that explanation predictions from regularized self-explanation training match the model's current behaviors (Self) more accurately than the original behaviors (Orig), despite training solely on fixed labels. This Self > Orig effect is strictly tied to behavioral regularization; unregularized models fail to couple explanations to current behavior, instead favoring label-driven imitation.

Figure 2: Self > Orig emerges only with regularization; explanation faithfulness to self labels exceeds faithfulness to original labels in Hint-MMLU, AITA, and Refusal tasks (left block), while unregularized training loses this effect (right block).

Fine-grained breakdowns rule out trivial solutions: behavioral drift does not drive the model to a degenerate or easier-to-explain mode, and explanation accuracy is high in both change and unchange subsets. External explainer baselines confirm that this coupling is not present unless an introspection-trained model, with privileged access to its own representations, acts as explainer.

Mechanistic Analysis

The mechanistic origin of introspective coupling is scrutinized using activation patching interventions. Layer/token-wise patching swaps activations between paired prompts differing in cue-induced behavior. The resulting changes in behavior logits are strongly correlated ($r=0.89$) with changes in explanation logits, confirming that the same internal circuits drive both behavioral output and meta-explanation when regularized introspection training is employed.

Figure 3: Mechanistic signature—patching activations affects both behavior and explanation logits in a correlated manner, indicating overlapping circuits behind introspective coupling.

Baselines show diminished or absent correlation, verifying that regularization-induced introspective training structurally couples explanations to behaviors at the representational level.

When and Why Introspective Coupling Emerges

A series of ablations and parameter sweeps interrogate the determinants of introspective coupling:

• Regularization Sensitivity: Sweeping regularization strength ($\lambda$) reveals that even minimal behavioral regularization suffices for Self > Orig to emerge, so long as explanation supervision remains compatible with online behaviors.

Figure 4: Increasing behavioral regularization $\lambda$ precipitates Self > Orig, with explanation fidelity tracking behavioral agreement.

• Label-Self Agreement Thresholds: Continuous online relabeling studies establish that coupling requires a minimum threshold ($\nu \approx 0.7$) of agreement between explanation supervision ($E_\text{sup}$) and the current behavior-generated explanations ($E_{\text{self}}$). Below this threshold, explanation faithfulness collapses.

Figure 5: Coupling only persists when online label-self agreement exceeds threshold; below it, explanations fail to generalize to current behaviors.

• Cross-Model Label Reusability: Training on explanations from a behaviorally similar but distinct model (e.g., mixing Qwen and Llama labels) maintains the Self > Orig effect as long as current behavior is regularized irrespective of label source, motivating scalable label reuse.

Figure 6: Explanation labels from another model retain introspective coupling, as long as behavior is regularized to match the current model.

• Learning Rate and LoRA Rank: High learning rates and higher-rank LoRA adapters widen the Self > Orig gap and encourage coupling, even with substantial behavioral drift, indicating that multiple interacting factors—not merely online label-self similarity—control emergence.

Generalization: Tracking New and Shifted Behaviors

Introspective coupling is tested in two auxiliary data scenarios:

• Acquisition of Novel Behaviors: Models trained concurrently on explanation supervision in the original domain and behaviors in a synthetic, nonsensical "Jabberwocky" domain (with no explanation supervision) exhibit high explanation faithfulness to their new behaviors, far surpassing chance and naive baselines.

Figure 7: Explanations generalize to newly acquired behaviors not seen in explanation training.

• Behavioral Drift Induced by Realistic Post-Training: Auxiliary training on datasets (WildChat, warm-assistant dialogue, FineWeb, direct refusal) shifts both behaviors and explanations in lockstep; explanation accuracy tracks the post-training induced drift and reflects emergent behavioral change.

Figure 8: Models' explanations track behavioral shifts induced by auxiliary post-training, confirming generalization and practical utility for monitoring emergent behaviors.

Theoretical and Practical Implications

The demonstrated introspective coupling fundamentally alters the requirements for scalable explanation training pipelines. Explanation labels need not be constantly updated to track model drift; fixed or cross-model supervision, coupled with moderate behavioral regularization, suffices to induce faithful introspection. Label sharing across model families becomes feasible, and explanation systems retain utility even under significant behavior shifts from fine-tuning or distributional changes.

The mechanistic coupling between explanation and behavior suggests that introspective capabilities leverage internal privileged access, potentially enabling more accurate interpretability than externally trained explainers. This has direct relevance for safety monitoring, debugging, and oversight of alignment training, post-training regimes, and behavioral shifts in deployed models.

Limitations and Future Directions

Despite the strong empirical and mechanistic evidence, several open questions remain:

• The coupling signal disappears in domains lacking sufficient behavioral variation, indicating a need for more diverse supervision and task selection.
• The interplay between label-self similarity, learning rate, adapter architecture, and other factors remains not fully characterized.
• The internal circuit structure behind introspective coupling is revealed but not completely explained; further probing is required to map out introspection pathways and possible asymmetric effects.

Conclusion

The paper establishes that LLMs trained via fixed counterfactual explanation supervision and behavioral regularization consistently produce explanations that more faithfully track their own current policy than the original policy used for supervision. The introspective coupling effect is empirically robust, mechanistically grounded, and generalizes to new and shifted behaviors with realistic post-training signals. These insights engender practical improvements in scalable interpretability and self-explanation-based oversight, motivating further study into mechanistic origins and cross-domain generalization of introspective phenomena.