Title resolution pending

representative citing papers

citing papers explorer

Showing 38 of 38 citing papers.

• Geometric Collapse: When Vision Models Fail to Verify Physical Causality cs.CV · 2026-07-08 · accept · none · ref 28

  Modern depth predictors adopt physically unsupported edge cues as real geometry, causing global structural hallucinations that local repair cannot fix.

• The inattentional gap in task conditioned AI models that omit otherwise reportable safety critical signals cs.CL · 2026-06-25 · conditional · none · ref 34

  Task-conditioned language and vision models omit co-present safety-critical signals they report when unconstrained, decoupling benchmark safety from deployment safety.

• Discovering Latent Groups for Robust Classification cs.LG · 2026-06-22 · unverdicted · none · ref 1

  NCT uses iterative correctness-based routing in a tree architecture to disentangle latent subgroups for robust classification without subgroup labels, yielding competitive performance on spurious correlation benchmarks.

• Implicit Neural Representations of Individual Behavior cs.LG · 2026-06-10 · unverdicted · none · ref 37

  Behavioral INR adapts INRs to behavior by mapping states to actions with FiLM-modulated episode latents for self-supervised policy inference in unlabeled data, with new policy OOD definitions.

• The Identity Trap in EEG Foundation Models: A Diagnostic Audit cs.LG · 2026-06-04 · unverdicted · none · ref 9

  Subject identity variance dominates frozen representations in three EEG foundation models by 13-89x over null, and erasing the linear subject axis improves label decoding where within-subject label variation exists.

• Toward Calibrated, Fair, and accurate Deepfake Detection cs.LG · 2026-06-03 · unverdicted · none · ref 210

  Face-Feature Tuning is a label-free logit remapping method that reduces FPR/TPR gaps across groups in deepfake detection while preserving overall accuracy.

• Whose Alignment? Comparing LLM Process Alignment Across Diverse Organizational Decision Contexts cs.AI · 2026-05-24 · unverdicted · none · ref 8

  Empirical study finds strong heterogeneity in LLM process alignment across models and organizations; process alignment predicts output accuracy in legal decisions but is low and resistant in credit decisions where higher alignment may not be desirable.

• Likelihood scoring for continuations of mathematical text: a self-supervised benchmark with tests for shortcut vulnerabilities cs.LG · 2026-05-11 · unverdicted · none · ref 29

  Presents a likelihood-based benchmark for equation-suffix prediction in technical papers with controls to detect shortcut vulnerabilities in model forecasts.

• How Language Models Process Negation cs.CL · 2026-05-04 · unverdicted · none · ref 16

  LLMs process negation using both attention-based suppression and constructive representation mechanisms (construction dominant), with late-layer attention shortcuts explaining poor accuracy on negation tasks.

• The UNDO Flip-Flop: A Controlled Probe for Reversible Semantic State Management in State Space Model cs.LG · 2026-04-07 · unverdicted · none · ref 4

  Mamba-2 models fail to learn reversible state retrieval in the UNDO Flip-Flop task, defaulting to a toggle heuristic and achieving only 41% accuracy under adversarial conditions.

• Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models cs.AI · 2024-06-14 · conditional · none · ref 43

  LLMs trained on simple specification gaming generalize to zero-shot reward tampering including rewriting their own reward function.

• FLORA: A deep learning approach to predict forest attributes from heterogeneous LiDAR data cs.CV · 2026-06-30 · unverdicted · none · ref 40

  FLORA is an octree-based deep learning framework with auxiliary data fusion that predicts forest attributes from heterogeneous LiDAR, achieving rRMSE of 12.3% for dominant height and 39% for total volume on 32k French NFI plots.

• Correct codes for the wrong reasons? validating LLMs as measurement instruments for theoretical constructs cs.CL · 2026-06-26 · unverdicted · none · ref 16

  Grain calibration decomposes theoretical constructs into clause-level components, tests each with extractive evidence, and combines results through explicit theory-derived rules to validate LLM coding beyond agreement with human annotators.

• ToxiREX: A Dataset on Toxic REasoning in ConteXt cs.CL · 2026-06-26 · unverdicted · none · ref 217

  ToxiREX is a new dataset of 128k Reddit comments in six languages with hierarchical annotations for implicit toxicity in conversational context based on an existing reasoning schema.

• CFPO: Counterfactual Policy Optimization for Multimodal Reasoning cs.CV · 2026-06-22 · unverdicted · none · ref 2

  CFPO is a counterfactual policy optimization method that regularizes RL policies in LVLMs by maximizing prediction discrepancy under suppressed visual cues, reporting 3-6% gains over baselines.

• Graph-of-Differences: Anatomy-Structured Difference Alignment for Medical Image Re-Identification cs.CV · 2026-06-19 · unverdicted · none · ref 6

  GoD uses anatomy graphs and difference alignment to improve medical image re-identification accuracy and auditability, with +7.1 pp Rank-1 gains on fundus and +3.1 pp on CXR.

• Forecasting Is Not Attribution: Localizing Decoder Bypass in Graph-Based Neural Marketing Mix Models cs.LG · 2026-06-10 · unverdicted · none · ref 10

  DICE-MMM is a two-stage diagnostic framework that separates forecasting accuracy from graph-aligned attribution in neural MMMs and localizes decoder bypass via controlled graph tests.

• What's in a Name? Morphological Shortcuts by LLMs in Pharmacology cs.CL · 2026-06-04 · unverdicted · none · ref 8

  LLMs induce pharmacological meaning primarily from affix cues in drug names, as revealed by a framework applied to 653 drugs and localized via activation patching to early-mid layers.

• Consistency Training Can Entrench Misalignment cs.CL · 2026-06-02 · unverdicted · none · ref 18

  Consistency training suppresses reward hacking and emergent misalignment but amplifies sycophancy in controlled model organisms, driven by labeling-induced distribution shifts rather than selection operators.

• What Are We Actually Decoding? Source Attribution for Non-Invasive Brain-to-Language Retrieval cs.LG · 2026-05-23 · unverdicted · none · ref 6

  An auditing framework for brain-to-audio retrieval isolates structural, stimulus-locked, and contextual performance sources via controls and a new Group Context Bias intervention, showing reduced performance under strict settings and measurable contextual gains.

• Understanding Goal Generalisation in Sequential Reinforcement Learning cs.LG · 2026-05-22 · unverdicted · none · ref 22

  Empirical analysis of over 100 sequential RL training pipelines across 250+ OOD environments finds salient features drive generalization and early goals persist, with latent policy gradients simulating latent variable evolution to predict OOD behavior from training history.

• DeconDTN-Toolkit: A Library for Evaluation and Enhancement of Robustness to Provenance Shift cs.LG · 2026-05-11 · unverdicted · none · ref 51

  DeconDTN-Toolkit simulates provenance shifts to expose ERM vulnerabilities and provides tools plus a robust OOD indicator for mitigating confounding by data provenance.

• CauSim: Scaling Causal Reasoning with Increasingly Complex Causal Simulators cs.AI · 2026-05-09 · unverdicted · none · ref 48

  CauSim turns scarce causal reasoning labels into scalable supervised data by having LLMs incrementally construct complex executable structural causal models.

• Catching The Correct Answer Trap: Characterising AI Tutor Blind Spots When Analysing Student Reasoning cs.CY · 2026-04-20 · conditional · none · ref 7

  AI tutoring models systematically fail to detect student misconceptions when flawed reasoning coincidentally produces the correct answer, with 71% of failures concentrated in two predictable question types.

• The Geometric Canary: Predicting Steerability and Detecting Drift via Representational Stability cs.LG · 2026-04-20 · unverdicted · none · ref 90

  Task-aligned supervised geometric stability predicts linear steerability with high accuracy while unsupervised stability detects representational drift earlier and with lower false alarms than CKA or Procrustes.

• Safety, Security, and Cognitive Risks in World Models cs.CR · 2026-04-01 · unverdicted · none · ref 47

  World models enable efficient AI planning but create risks from adversarial corruption, goal misgeneralization, and human bias, demonstrated via attacks that amplify errors and reduce rewards on models like RSSM and DreamerV3.

• PaLM: Scaling Language Modeling with Pathways cs.CL · 2022-04-05 · accept · none · ref 54

  PaLM 540B demonstrates continued scaling benefits by setting new few-shot SOTA results on hundreds of benchmarks and outperforming humans on BIG-bench.

• Unraveling Machine Behavior by Multi-Level Bias Analysis and Detection: Methodology and Application to Computer Vision cs.CV · 2026-07-08 · conditional · none · ref 28

  Bias in CNNs is detectable from internal representations—latent space geometry, layer activations, and filter weights—using statistical tests or a trained classifier, validated on 127,000+ models.

• A quantum-like benchmark for context-sensitive associative memory with adaptive plasticity q-bio.NC · 2026-05-30 · unverdicted · none · ref 19

  A controlled benchmark for context-sensitive memory shows adaptive plasticity (especially homeostatic) enables recall under weak support, with quantum-like models preserving order sensitivity better than Markov controls but without universal advantage.

• PaCo-VLA: Passivity-Shielded Compliance Prior for Contact-Rich Vision-Language-Action Manipulation cs.RO · 2026-05-30 · unverdicted · none · ref 19

  PaCo-VLA adds an independent passivity shield to VLA outputs so that semantic proposals for compliance and admittance can be used in contact-rich tasks without violating passivity or causing damage.

• MinhwaNet: Faithful but Insufficient Object Grounding in Korean Folk Painting cs.MM · 2026-05-26 · unverdicted · none · ref 8

  Object presence alone fails to predict minhwa genres while image-text fusion succeeds, revealing faithful but insufficient object grounding where genre hinges on symbol arrangement rather than inventory.

• Response-free item difficulty modelling for multiple-choice items with fine-tuned transformers: Component-wise representation and multi-task learning cs.CL · 2026-05-16 · conditional · none · ref 35

  Fine-tuned transformers with multi-task learning recover substantial wording-derived signal for item difficulty at small sample sizes typical in applied testing.

• How Code Representation Shapes False-Positive Dynamics in Cross-Language LLM Vulnerability Detection cs.CR · 2026-04-30 · unverdicted · none · ref 9

  Text fine-tuning of 8B LLMs on C/C++ vulnerability data inflates cross-language false-positive rates through surface-cue memorization, which an AST inference probe can partially reverse while direct AST fine-tuning cannot.

• pAI/MSc: ML Theory Research with Humans on the Loop cs.AI · 2026-04-22 · unverdicted · none · ref 37

  pAI/MSc is a customizable multi-agent system that reduces human steering by orders of magnitude when turning a hypothesis into a literature-grounded, mathematically established, experimentally supported manuscript draft in ML theory.

• Text-to-Image Models and Their Representation of People from Different Nationalities Engaging in Activities cs.CV · 2025-04-08 · conditional · none · ref 14

  T2I models generate traditional attire in 28.4% of images across 206 nationalities and five activities, with higher rates in Middle East/North Africa and Sub-Saharan Africa, stronger image-prompt alignment when country names are present, and frequent insertion of the word 'traditional' by one model.

• DASH: Dual-View Self-Distillation with Multi-Layer Hidden Representations for Robust Speech Recognition eess.AS · 2026-06-17 · unverdicted · none · ref 28

  DASH applies dual-view self-distillation on multi-layer representations and prototype distributions to boost ASR noise robustness while keeping clean accuracy.

• Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization cs.AI · 2026-06-08 · unverdicted · none · ref 97

  Proxy RL produces a staged proxy-internalization capability that emerges before and predicts reward hacking in coding environments.

• AI Safety Landscape for Large Language Models: Taxonomy, State-of-the-art, and Future Directions cs.AI · 2024-08-23 · unverdicted · none · ref 240

  The paper introduces a taxonomy of AI safety for LLMs organized into Trustworthy AI, Responsible AI, and Safe AI perspectives, accompanied by a review of state-of-the-art methods, challenges, and future directions.