U-Space: Uncovering When and Why Uncertainty Arises in Language Models

U-Space extracts an interpretable low-dimensional subspace from language model hidden states to map evolving token-level uncertainty without training or repeated generations, outperforming baselines.

Tobias Braun, Nils Loose, Alexander Herzog, Virginia Ceccatelli and 3 more

Published Oct 6, 2026 · ▲ 3 on Hugging Face · Code ★ 2

Lingtai: What Concept Geometry Reveals—and Does Not Reveal—About LLM Inference

Lingtai introduces a training-free concept telemetry layer that reveals inference-time uncertainty-linked activity and execution-specific trajectory structures in LLMs without tracking correctness.

Jiangang Chen

Published Oct 1, 2026 · 0 citations

Making LLMs Say What They Think: Measuring and Improving CoT-Interpretability Alignment

We introduce CIA to measure chain-of-thought alignment with internal reasoning, finding low alignment that post-training improves substantially while maintaining accuracy.

Yihuai Hong, Shauli Ravfogel, Chen Zhao, Eunsol Choi

Published Sep 30, 2026 · 0 citations · ▲ 21 on Hugging Face · Code ★ 2

CircuitSeer: Mining High-Quality Data by Probing Mathematical Reasoning Circuits in LLMs

Shaobo Wang, Yongliang Miao, YuanCheng Liu, Qianli Ma and 3 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Cross-Model Circuit Discovery

Harrish Thasarathan, Matthew Kowal, Thomas Fel, Kosta Derpanis

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Concept-Localized Generative Representations

Amanda Merkley

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Spectral-Spatial Interpretation

Haotian Ma, Ruqi Yang, Philip Townsend

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

LoRAcles: Self-Supervised Weight-Space Interpretability at Scale

Celeste De Schamphelaere, Jan Bauer, Neel Nanda, Euan Ong

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

How Do Language Models Understand Tables? A Mechanistic Analysis of Cell Location

Xuanliang Zhang, Dingzirui Wang, Keyan Xu, Qingfu Zhu and 1 more

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026

What Arranges Features in Activation Space? Non-Classical Predictive Geometry in Next-Token Predictors

Adam Shai, Thomas J Elliott, Paul Riechers

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Gluing Local Contexts into Global Meaning: A Sheaf-Theoretic Decomposition of Transformer Representations

Bryce Grant, Peng Wang

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Attention Sinks as Spectral Spikes: A Mechanism Analysis of Gated Attention

Seojin Kim, Yehjin Shin, Noseong Park

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Reading Positional Coupling in Transformers with Diffusion Scores

Savik Kinger, Johannes Bertram, Luciano Dyballa, Andy Keller and 1 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

The Sign Code: The Hidden Binary Nature of Deep Networks

Niclas A Göring, Shuofeng Zhang, Roi Holtzman, yoonsoo nam and 1 more

Paris Poster Session 1, Wed, Dec 9, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Decision Path Tracing for Causal Analysis in Transformers

Won Jo, Dahee Kwon, Jongeun Baek, Cheongwoong Kang and 1 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Knowing Without Saying: How Contextual Evidence Survives but Fails to Surface in Transformers

Ruochen Jin, Tianyu Pang, Chenxi Lin, Lei Hsiung and 4 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Learnable Spectral Activations

Tamir Shor, Or Litany, Alex M Bronstein

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Right Results, Wrong Reasons: Auditing Behavioral Reliance in Motion Forecasting

Geonyeong Park, Byounghun Park, Nayoung Kim, Kyungmin Kim and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

How Language Models Compress and Compare: Understanding Selection with Token Covariance Maps

Moshe Eliasof

Paris Poster Session 3, Thu, Dec 10, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Iterative Nonlinear Computation Underlying Abstract Reasoning

Zitian Gao, Yilong Chen, Yihao Xiao, Xinyu Yang and 4 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Concept-Based Mechanistic Interpretability Needs a Concrete Evaluation Paradigm

Yiming Tang, Qinglin Qi, LIN Zheng, Dianbo Liu

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

JEPAWG: Interpretable Hypernetworks for Weight-Space Physics

Tobias Göbel, Julian R Ebelt, Zier Mensch, Mathis Gerdes and 1 more

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026

IDEA: Unwrapping Visual Black-box Models by Interaction Decomposition

Chaojie Ji, Jian Xu, He Zhang, Hongyan Wu and 2 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

What Transformer FFNs Never See: Theory, Diagnosis, and Lightweight Remediation

Tinghe Zhang, Yucheng Xiao, Alex Lamb

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

From Groups to Rings: Causal Evidence for Algebraic Decomposition in Grokked Transformers

Shurui Zheng, Fanhong Li, Zhixing Huang, Zixi Li and 1 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

ScopeSAE: Model-Scope Feature Discovery with Interpretable Layer Selection

Qingwen Zeng, Zehao Fu, Shuyu Meng, Linghan Huang and 4 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Can Circuit Alignment Predict OOD Generalization?

Ayan Banerjee, ABHRA CHAUDHURI, Josep Llados, Umapada Pal and 1 more

Paris Poster Session 3, Thu, Dec 10, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Where Reusable Computation Becomes Detectable: Solution-Frame Path Triage for Modular-Arithmetic Grokking

Ruian Lei, Ruoxi Jiang, Marley M Vellasco, M. Tanveer and 1 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Mechanistic Interpretability with Sparse Autoencoder Neural Operators

Bahareh Tolooshams, Ailsa Shen, Animashree Anandkumar

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

SASA: Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability

Seyed Arshan Dalili, Mehrdad Mahdavi

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Matched-Control Tests of Partition-Source Claims in One Routed Distillation Family

Bob Li, Chloe Cao, Constance Liu

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

What does a Bayes-filtered transformer believe? A predictive Monte Carlo approach

Afiq Abdillah Effiezal Aswadi, Haotong Ma, Susan Wei

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Introspection Tools Help LLMs Understand and Control Themselves

Xiao Liu, Junsol Kim, Shiyang Lai, Jacy Anthis and 3 more

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Think about how AIs think about themselves

Raymond Douglas, Jan Kulveit, Ondřej Havlíček, Theia Pearson-Vogel and 1 more

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026

Claims of AI emergence should be grounded in information decomposition

Christoph Riedl, Fernando Rosas

Paris Poster Session 6, Fri, Dec 11, 2:30 PM–4:30 PM, Paris Poster Hall · Published 2026

The parameters in weight-sparse transformers are interpretable

Arnau Marin-Llobet, Stefan Heimersheim

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Loyalty Capture: Reporting Relationships and Structural Sycophancy in Frontier AI Models

Eric So, Alex Imas

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Reading Attribution from Attention: Evidence Heads as Latent Attribution Mechanisms in LLMs

Qi Wu, Jianfeng Qu, Peng-Fei Zhang, Yanzhe Ji and 4 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Algorithmic Impact Reveals the Hidden Structure of Alignment

Zachary Wojtowicz, Michelle Si, Finale Doshi-Velez, Ariel Procaccia

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Do Latent-CoT Models Think Step-by-Step? A Mechanistic Study on Sequential Reasoning Tasks

Jia Liang, Liangming Pan

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Specialist Mediators: Causal Localization of Dense Fine-Tuning

Amir Asiaee

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

GEM: Interpretable Language Models via Geometric Embedding Alignment

Dimitrios Tsaras, Yankun Hong, Lei Chen, Zhiyao Xie and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

The Unembedding Bottleneck: A Mechanistic Analysis of Single Digit Counting in LLMs

Satwik Sunnam, Raghav Magazine, Vatsalya Singh, Lavanya Kotha and 2 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Transfer Entropy as a Measure of Information Flow in VLMs and LLMs

Jessica E Liang, Jianbo Shi

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Beyond Incremental Beam Search for Compositional Explanations of Neurons

Biagio La Rosa, Leilani Gilpin

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Spectral Measures of Mamba Conductances Predict and Shape Effective Receptive Fields

Zy Li

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

A Mechanistic Investigation of Theory of Mind in a Large Language Model

Idil K Sahin, Steven M Frankland, Taylor Webb

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Conclusions from Circuit Extraction Depend on the Level of Description: A Controlled Comparison

yang sheng, Jie Fu

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Sink vs. diagonal patterns as mechanisms for attention switch and oversmoothing prevention

Sinks and diagonal patterns serve as attention switches and anti-oversmoothing mechanisms, with sinks favored in pretrained transformers due to lower representation costs.

Peter Súkeník, Cristina Lopez Amado, Christoph Lampert, Marco Mondelli

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunks

Protein folding models share a two-stage trunk mechanism initializing biochemical signals then spatial features, with causally steerable, interchangeable representations across architectures.

Kevin Lu, Jannik Brinkmann, Stefan T Huber, Aaron Mueller and 3 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Emergent representations of graphical structure in mechanistic neural models of causal judgment

Task-optimized recurrent neural networks learn to judge causal relations and implicitly represent complete graphical structures via low-level neural mechanisms.

Marcus Triplett, Kenneth Kay

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Training Language Models to Explain Their Own Computations

Fine-tuning language models on interpretability ground truth teaches them to describe their internal computations, with self-explanation outperforming larger external explainers.

Belinda Z Li, Zifan Carl Guo, Vincent Huang, Jacob Steinhardt and 1 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 1 on Hugging Face · Code ★ 38

Artificial Aphasias in Lesioned Language Models

Lesioning language models to mimic aphasia reveals distinct artificial symptom profiles across components and depths, indicating human aphasias depend heavily on learning and processing specifics.

Nathan Roll, Jill Kries, Laura Gwilliams, Cory Shain

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

OceanCBM: A Concept Bottleneck Model for Mechanistic Interpretability in Ocean Forecasting

OceanCBM is a concept bottleneck model that predicts ocean heat content through physically derived intermediate concepts, yielding consistent interpretable representations without sacrificing predictive skill.

Sanah Suri, Kieran Ringel, Maike Sonnewald

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracing

DifFRACT extends transcoder-based circuit tracing to multimodal diffusion transformers, yielding exact feature attribution, interpretable circuits for attribute binding and cross-stream propagation, and precise circuit-guided interventions outperforming standard SAE steering.

Artyom Mazur, Nina Konovalova, Aibek Alanov

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

The SuperActivator Mechanism: Transformers Concentrate Reliable Concept Signals in the Tail

Transformers use superactivator mechanisms to amplify concept activation gaps, concentrating reliable evidence into sparse high-activation token tails that improve concept detection F1 by up to 0.14.

Cassandra Goldberg, Chaehyeon Kim, Adam Stein, Eric Wong

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

FiTS: Interpretable Spiking Neurons via Frequency Selectivity and Temporal Shaping

FiTS introduces frequency-selective, temporally shaped spiking neurons that improve feedforward SNN accuracy on auditory tasks while yielding interpretable frequency and timing organization.

Jongmin Choi, Joon Son Chung

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

A Mechanistic Analysis of Looped Reasoning Language Models

Looped reasoning models converge to cyclic fixed points that stabilize attention and repeat feedforward inference stages iteratively, with recurrence size and normalization affecting stability.

Hugh Blayney, Alvaro Arroyo, Johan Obando Ceron, Pablo Samuel Castro and 3 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 1 on Hugging Face

Cross-Family Universality of Behavioral Axes via Anchor-Projected Representations

An anchor-projected framework maps hidden representations into shared coordinates to extract and transfer behavioral directions across model families without fine-tuning, achieving high cross-family steering and detection accuracy.

Su-Hyeon Kim, Yo-Sub Han

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Unveiling Memorization–Generalization Coexistence: A Case Study on Arithmetic Tasks with Label Noise

Over-parameterized networks memorize noisy arithmetic labels yet internally form generalizable structures recoverable via frequency methods, revealing distributed memorization-generalization coexistence.

Linyu Liu, Pinyan Lu

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Geometry-Adaptive Explainer for Faithful Dictionary-Based Interpretability under Distribution Shift

Out-of-distribution shifts rotate active subspaces, misaligning dictionary explainers; a geometry-adaptive realignment using unlabeled OOD activations closes the faithfulness gap and restores causal interpretability without training.

Sungjun Lim, Heedong Kim, Andrew Lee, Kyungwoo Song

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models

Visual pathways dominate VLA action generation via spatial motor programs, language matters only with ambiguous goals, and expert pathways encode motor control separately from VLM semantic pathways.

Bryce Grant, Xijia Zhao, Peng Wang

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Rare Events, Real Signals: Functional Ensembles as Units of Computation in Deep Spiking Networks

Deep spiking ResNets preserve cortical-like functional connectivity where rare, coordinated 1FC ensemble cofiring reliably predicts downstream responses via ReLU-like scaling, encodes class identity, and breaks under adversarial perturbation and weight permutation.

Aditi Aravind, Konstantinos Ladakis, Mario Alexios Savaglio, Stelios Smirnakis and 1 more

Paris Poster Session 1, Wed, Dec 9, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models

CRAFT localizes arbitration and brake failure in medical vision-language models to disjoint attention head sets, enabling targeted interventions that reduce misleading text influence and restore abstention without retraining.

Chunzheng Zhu, Jiaqi Zeng, Hongbo Zhao, Yihang Chen and 2 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Reading Between the Dots: Decoding Hidden Computation across Filler Tokens

Open-weight LLMs perform hidden multi-step reasoning over filler tokens that unsupervised hidden-state decoding recovers at 82-94% accuracy, showing monitorability requires internal traces.

Kaley Brauer, Claudio Mayrink Verdun, Samuel Marks

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Task Vector Geometry Underlies Dual Modes of Task Inference in Transformers

Task-vector geometry governs dual inference: in-distribution tasks use convex combinations of learned vectors, while out-of-distribution tasks use nearly orthogonal extrapolative subspaces.

Hao Yan, Haolin Yang, Yiqiao Zhong

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Neuron Populations Exhibit Divergent Selectivity with Scale

Rosetta neuron populations grow sublinearly and become more selective and specialized as language and vision models scale, while non-Rosetta neurons stay less selective.

Amil Dravid, Yasaman Bahri, Alexei Efros, Yossi Gandelsman

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Geometric Factual Recall in Transformers

Transformers memorize facts geometrically via linear superpositions and MLP selectors, needing only logarithmic dimensions and enabling zero-shot MLP transfer.

Shauli Ravfogel, Gilad Yehudai, Joan Bruna, Alberto Bietti

Paris Poster Session 3, Thu, Dec 10, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026 · ▲ 1 on Hugging Face

Transcoder Adapters for Reasoning-Model Diffing

Transcoder adapters approximate reasoning fine-tuning differences in MLPs, revealing sparse interpretable features where only ~8% relate to reasoning behaviors like hesitation tokens.

Nathan Hu, Jake Ward, Thomas Icard, Chris Potts

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · Code ★ 5

What Cohort INRs Encode, and Where to Freeze Them

Freezing cohort INR layers at highest stable-rank depth improves fitting, with sparse autoencoders revealing SIREN learns localized atoms and FFMLP learns image-spanning memorized contours.

Vasiliki Sideri-Lampretsa, Sophie Starck, Robbie Holland, Julian McGinnis and 1 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Sparse Reward Subsystem in Large Language Models

LLM hidden states contain sparse value and dopamine neurons forming a reward subsystem that predicts confidence and guides search.

Guowei Xu, Mert Yuksekgonul, James Zou

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 13 on Hugging Face

Dual-Pathway Circuits of Object Hallucination in Vision-Language Models

Vision-language models contain separate visual grounding and hallucination pathways whose components flip polarity to drive errors, and suppressing them cuts object hallucination by up to 76%.

Jiaxin Liu, Ding Zhong, Yue Wang, Zhidong Yang and 5 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching

Activation patching's natural indirect effect embeds hidden interaction effects between components, which cause conditional importance to be invisible or inflated, explain faithfulness instability, scale with activation distance, and diagnose when greedy component ranking misses combinatorial mechan

Sankaran Vaidyanathan, David Arbour, Aaron Mueller, Scott Niekum and 1 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Decoding the Critique Mechanism in Large Reasoning Models

Large reasoning models use hidden critique abilities to recover from uncorrected reasoning errors, and steering with critique vectors improves error detection and test-time scaling without training.

Hoang Phan, Nguyen Hung-Quang, Thanh Quoc Hung Le, Xiusi Chen and 2 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · Code ★ 1

On the Recall Scaling Laws in Mamba: A Theoretical and Mechanistic Study via Hashing

Mamba performs associative recall via implicit linear hashing, and Recall Scaling Laws predict required dimensions and success probabilities for perfect recall.

Yuval Koren, Assaf Ben-Kish, Raja Giryes, Lior Wolf and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants

Predictive Concept Decoders train end-to-end interpretability assistants that encode neural activations into sparse concepts to predict model behavior, scaling with data to detect jailbreaks, hidden hints, and latent attributes.

Vincent Huang, Dami Choi, Daniel D Johnson, Sarah Schwettmann and 1 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining

Mechanistic analysis reveals a Commit-Abstain Circuit where early commitment signals overpower later abstention corrections, causing hallucinations; training on its activations improves abstention accuracy by 12.2 points.

Gavin Vy Nguyen, Ziqi Xu, Jeffrey Chan, Estrid He and 4 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Understanding Generalization through Decision Pattern Shift

Decision Pattern Shift uses GradCAM channel contributions to define generalization as stable internal decision patterns, finding DPS correlates linearly with the generalization gap and unifies failure modes.

Huiqi Deng, Yibo Li, Quanshi Zhang, Peng Zhang and 2 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Finding Interpretable Prompt-Specific Circuits in Language Models

ACC++ improves circuit tracing to extract interpretable prompt-specific language model circuits from single passes, revealing clustered indirect-object mechanisms and language-specific reused components.

Gabriel Franco, Lucas M Tassis, Azalea Rohr, Mark Crovella

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Interpreting Latent Protein Language Model Features with Geometric Annotations

Sparse autoencoders in ESM-2 are interpreted via Cα geometric features, revealing localized structural patterns and substructure within biological labels that sequence annotations miss.

Siddharth Setlur, Djordje Mihajlovic, Darrick Lee

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Few Channels Draw The Whole Picture: Revealing Massive Activations in Diffusion Transformers

A small subset of hidden-state channels in diffusion transformers drives image semantics, spatial structure, and prompt transfer without training.

Evelyn Turri, Davide Bucciarelli, Sara Sarto, Lorenzo Baraldi and 1 more

Paris Poster Session 4, Thu, Dec 10, 5:30 PM–7:30 PM, Paris Poster Hall · Published 2026

Predictive Geometry of Hidden Trajectories in Transformers

Hidden-state loss-to-go geometry near successful trajectories is governed by a pullback Fisher operator that reveals output-sensitive directions, enabling rank allocation, pruning, and distillation improvements.

Timur Mudarisov, Mikhail Burtsev, Tatiana Petrova, Radu State

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026

Mechanistic Circuit Identification for Controllable Data Generation

A circuit-grounded framework links mechanistic interpretability to data valuation via learnable, challenge, and alignment circuits, using them to steer generation and improve QA performance.

Nakyung Lee, Sangwoo Hong, Jungwoo Lee

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026