ABCDEFGHIJKLMNOPQRSTUVWXYZ
1
Submission NumberTitlePoster SessionTime
2
2When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction110:50 - 11:55
3
8Decoy Direction Optimization: A Post-Hoc Defense Against LLM Abliteration110:50 - 11:55
4
22How Faithful Is Trajectory-Based Data Attribution? Error Sources, Remedies, and Practical Guideline110:50 - 11:55
5
29When the Loop Closes Quietly: Mechanistic Interpretability as Contestability Infrastructure for Deployed AI110:50 - 11:55
6
43Self-CTRL: Self-Consistency Training with Reinforcement Learning110:50 - 11:55
7
44Fuzzing Large Language Models to Elicit Hidden Behaviours110:50 - 11:55
8
52On the Non-Identifiability of Steering Vectors in Large Language Models110:50 - 11:55
9
64EMO: Pretraining Mixture of Experts for Emergent Modularity110:50 - 11:55
10
69Mechanistic Specialization Does Not Guarantee Performance: Why GPT-2 Outperforms Dual Attention Transformers on Identity Rule Tasks110:50 - 11:55
11
73Decodability is Not Causality: Dissociating Probe Readouts from Behavioral Drivers via SAE Decomposition110:50 - 11:55
12
76Generative Meta-Models are Adversarial Classifiers110:50 - 11:55
13
81Asymmetric Collapse in Model Merging: When Refusal Overwrites Recognition110:50 - 11:55
14
84Truco-Bench: Covert Signaling in LLM Agents and What Black-Box Monitors Miss110:50 - 11:55
15
88LangFIR: Discovering Sparse Language-Specific Features from Monolingual Data for Language Steering110:50 - 11:55
16
93From Refusal to Fabrication: An Affective Cause of Hallucination That Bypasses the Honesty Gate110:50 - 11:55
17
97Targeting the Attention Heads Behind Object Hallucination in LLaVA110:50 - 11:55
18
102LOCUS: Head Selection and Subspace Projection for Targeted Activation Steering.110:50 - 11:55
19
109Guardrail Vectors: Bidirectional Steering and Framing-Induced Safety Failures110:50 - 11:55
20
116LayerRoPE: Dynamic Depth-wise Magnitude & Angular Superposition110:50 - 11:55
21
118Activation Steering via Generative Causal Mediation110:50 - 11:55
22
120Reasoning Models Know What’s Important, and Encode It in Their Activations110:50 - 11:55
23
122Verbalizing LLMs' assumptions to explain and control sycophancy110:50 - 11:55
24
125Why Fine-Tuning Encourages Hallucinations and How to Fix It110:50 - 11:55
25
129Are Latent Reasoning Models Easily Interpretable?110:50 - 11:55
26
132Pando: Do Interpretability Methods Work When Models Won't Explain Themselves?110:50 - 11:55
27
6Localizing and Repairing Context-Failures in Audio-Language Models214:25 - 15:30
28
12Characterizing Model-Native Skills214:25 - 15:30
29
27Why Does Robustness Reduce Superposition?214:25 - 15:30
30
40Simple Activation Clipping Defeats Circuit Breakers214:25 - 15:30
31
50Sycophantic Agreement Transfers with Neutral Data via Contrastive Preference Optimization214:25 - 15:30
32
62Before You Act: Benchmarking Interpretability in Vision-Language Models214:25 - 15:30
33
66Shifting Mechanisms: How Positional Encoding Choice Shapes Long-Context Retrieval214:25 - 15:30
34
72Look Before You Steer: Geometry Predicts SAE Feature Steerability214:25 - 15:30
35
75Trait-space Monitoring for Emergent Misalignment During Supervised Finetuning214:25 - 15:30
36
78When Is a Probe Monitor Safe to Act On? A Transferability Stress Test for Eval-Awareness Probes under RLVR214:25 - 15:30
37
83Improving Confidence Estimation in Language Models with Linear Representations of Ambiguity214:25 - 15:30
38
86Limits of Activation-Direction Interventions for Emergent Misalignment Beyond Rank-1 LoRA214:25 - 15:30
39
89Facial and Vocal Emotion Elicit Partially Shared Emotion-Sensitive Neurons in Multimodal Foundation Models214:25 - 15:30
40
95Decided Upstream, Written Late: Locating and Pricing the Cross-lingual Circuit of a Multilingual MoE214:25 - 15:30
41
101MoRFI: Monotonic Sparse Autoencoder Feature Identification214:25 - 15:30
42
107Generalization of Steering Vectors for Chain-of-Thought Faithfulness214:25 - 15:30
43
113Who Is Left of Whom? Tracing Spatial Evidence and Role Binding in Relative-Position Reasoning214:25 - 15:30
44
119Minimal, Local, Causal Explanations of Jailbreak Success in Large Language Models214:25 - 15:30
45
121Do Language Models Consistently Encode the Current Year?214:25 - 15:30
46
123Source-Modality Monitoring in Vision-Language Models214:25 - 15:30
47
126Subliminal Steering: Encoding and Detecting Hidden Signals214:25 - 15:30
48
128Reasoning Fine-Tuning Induces Persistent Latent Policy States214:25 - 15:30
49
130Capability Provenance in Language Models: A Case Study in Social Reasoning214:25 - 15:30
50
137Disentangling MLP Neuron Weights in Vocabulary Space214:25 - 15:30
51
133Convergent Evolution: How Different Language Models Learn Similar Number Representations214:25 - 15:30
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100