NeurIPS 2026 reviews are up: reviews, scores, rebuttals and meta-reviews for 7,877 papers. See the highest rated
acceptodds: the papers of ICLR 2027
689 results for “keyword:"mechanistic interpretability"” in ICLR 2027. 142 withdrawn or desk rejected listed apart. clearsort:relevanceratingnewestmost read
[PDF]Affine Anticipation: LLM Answer Representations Are Largely Linearly Predictable from Context▼4.3k2.00
[PDF]Pinpoint Quantization: A Mechanistic Account of Tail-Capability Failures in Quantized LLMs▲5103.00
[PDF]Representation Without Disposition: Secret Loyalties Install as Preference but Not as Action, and What That Costs White-Box Audits▲3.0k2.33
[PDF]Don't Ignore Them: Scaffold Features Matter for Stable SAE Steering in Diffusion Models▲3952.67
[PDF]Localization Does Not Imply Steerability: A Mechanistic Study of Sign Errors in LLM Mathematics▼1361.33
[PDF]CRANE: Functional Selectivity of Language-Conditioned Neurons in Multilingual Large Language Models▼1.2k1.67
[PDF]Not Which Path, but What Happens Before You Would Stop: What Reasoning Post-Training Changes▼1361.33
[PDF]Who Grades the Edit? Frozen and Re-fit Readouts for Position Interventions in RoPE Models▼1361.33
[PDF]Computation Persists, Routing Degrades: A Mechanistic Analysis of Transformer Length Extrapolation▼4.3k2.00
[PDF]SAE Interventions Can Be Unreliable: Post-Intervention Recovery of Suppressed Behavior▼4.3k2.00
[PDF]Beyond Task-Conditioned Importance: Head-Resolved Output Jacobians Separate Broad Response from Behavioral Relevance▼1.2k1.67
[PDF]Disentangling Safety and Capability in Large Language Models via Causal Interventional Decomposition▼1361.33
[PDF]From Reweighting to Rewriting: Unlocking the Intervention Effects of Influential Samples in Training Data Attribution▲5103.00
[PDF]SAFETYEDIT: Sparse Internal Targets with Prompt-Wide Control for LLM Jailbreak Backdoors▼4.5k2.67
[PDF]Minority Welfare Under LLM-Governed Allocation: A Systematic Empirical and Mechanistic Study▼1361.33
[PDF]How Much Do Circuits Tell Us? Measuring the Consistency and Specificity of Language Model Circuits▲5103.00
[PDF]When Trees Are Not Enough: Learning Mixed-Topology Feature Graphs with Adaptive Graph Sparse Autoencoders▼1.2k3.00
[PDF]Do We Need Handcrafted Corruptions? Context-Preserving Activation Perturbation in Representation Space▼4.5k2.67
[PDF]Balanced Perturbation and Subspace Erasure Do Not Identify What an Activation Monitor Reads▼1.2k1.67
Withdrawn or desk rejected (142)
[PDF]Why Benign Post-Training Induces Safety Gate Collapse: Changes in Harmful Activation Geometry Relative to the Inherited Refusal GateWithdrawn1.67
[PDF]Concept Subspaces Compute Beyond the Logit Lens: A Weights-Only Test for Locating Representations Upstream of ReadoutWithdrawn