All entries

What Visual Generators Need from Teachers: Rethinking Representation Alignment

arXiv:2609.34732v1 Announce Type: new Abstract: Representation alignment speeds up diffusion transformer training by pulling an intermediate block of the model (student) toward features of a frozen pr

safetyarxiv-cs-cv

Content type

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

What Would Falsify It? A Variable Specific Evidence Standard for Mechanistic Claims About Self Explanation

DGX agent

arXiv:2609.32670v1 Announce Type: cross Abstract: When a language model explains an answer it has already given, does it reuse the computation that produced the answer or reconstruct a story from the

researcharxiv-cs-lg

29 Sept 2026

When an Evaluation Rule Writes Training Labels: Measuring Human-Reference Forgiveness in NAVSIM

arXiv:2609.33189v1 Announce Type: cross Abstract: When the human reference scores zero on a metric, the released GTRS-Dense label generator for NAVSIM marks every candidate trajectory in the scene as

When Better Gets Worse: Improvement Fidelity for Self-Improving Agents in Adaptive Worlds

arXiv:2609.32677v1 Announce Type: cross Abstract: Self-improving agents increasingly rely on proxy verifiers to choose policy updates, yet deployment can change the world in which those updates are ev

Tutorials

When Can Attention Heads Be Statically Defined?

arXiv:2609.34650v1 Announce Type: new Abstract: Some attention heads learn similar patterns across inputs. Reusing these patterns could reduce training cost by avoiding repeated query-key score comput

tutorialsarxiv-cs-lg

When Can Old Evaluations Certify a New Model? Label-Efficient Release Decisions under Evaluator Drift

arXiv:2609.32267v1 Announce Type: new Abstract: Releasing a model update requires certifying that its current-population risk stays below a threshold. Trusted labels are expensive, while a cheap evalu

When Can We Trust the Sparse Lens? A Certification Framework for SAE Faithfulness

arXiv:2606.18383v2 Announce Type: replace Abstract: Sparse autoencoders (SAEs) are increasingly used to study internal representations in language models (LMs), but it remains unclear when an SAE-base

When Do Model Internals Help? Exploring the Role of Representation Engineering in LLM Safety

arXiv:2609.34771v1 Announce Type: cross Abstract: Reliable AI safeguards require both control mechanisms that reduce unsafe behavior and monitoring mechanisms that detect safety risks during model int

When Do Models Admit They Are Wrong? Failure Disclosure Is Unstable Under Reinforcement Learning

arXiv:2609.33220v1 Announce Type: new Abstract: Outcome-based reinforcement learning can produce models with similar task performance but very different ways of communicating about their mistakes. We

When do Random Forests work?

arXiv:2504.12860v2 Announce Type: replace-cross Abstract: We study the effectiveness of randomizing split-directions in random forests. Prior literature has shown that, randomization can reduce varian

When Does an Image Determine the Answer? Benchmarking Visual Answerability across Charts and Scenes

arXiv:2609.34480v1 Announce Type: new Abstract: Reliable visual question answering requires correct answers when evidence is sufficient and abstention when it is not. We introduce a benchmark that con

When Does Backpropagating Through Policy Memory Matter? Physical Credit, Optimizer Updates, and Observability

arXiv:2609.33169v1 Announce Type: new Abstract: Policies with memory can learn along two backward paths: through the physical states their actions produce and through the representations they store. T

When Does Dense Retrieval Need Asymmetric Geometry? A Bias-Variance Theory of Shared and Dual Projections

arXiv:2609.32488v1 Announce Type: cross Abstract: Dense retrieval powers retrieval-augmented generation, semantic search, and question answering, yet the theoretical basis for choosing between shared

Tutorials

When Does Depth Matter For In-Context Learning? Adaptive Inference in Deep Transformers

arXiv:2606.16694v2 Announce Type: replace Abstract: Transformers perform computations through many successive attention and feedforward blocks, allowing them to learn complex correlations between a la

tutorialsarxiv-cs-lg

When Does Domain Adaptation Help on Physical Vibration Sensors? A Held-Out-Bearing Study of Neural-Operator and Convolutional Models

arXiv:2609.31639v1 Announce Type: new Abstract: Diagnosing rolling-element bearing faults from vibration is a canonical physical-sensing task and a widely used benchmark for domain adaptation under op

When Does Geometric View Synthesis Help Wine Label Retrieval? A Public One-Shot Benchmark Across Self-Supervised and Vision-Language Backbones

arXiv:2609.33359v1 Announce Type: new Abstract: Geometric view synthesis can expand a single wine-label photograph into a training set, but its value with pretrained image encoders is unclear. We stud

When Does Structured Knowledge Help Neural Theorem Proving?

arXiv:2609.34460v1 Announce Type: cross Abstract: Does structured mathematical knowledge help LLMs prove theorems in Lean 4? If so, for which models, and does the answer vary by problem? Formal librar

When Does Synergy Help Active Feature Acquisition? A PID-Based Study

arXiv:2609.32301v1 Announce Type: new Abstract: Active feature acquisition (AFA) sequentially selects informative features under budget constraints. However, existing policies rarely distinguish wheth

When Helpful Text Hurts: Option-Redirecting Bias in Vision-Language Models

arXiv:2609.32489v1 Announce Type: cross Abstract: In tri-modal visual question answering (VQA), auxiliary text is commonly used to complement visual and textual inputs, yet its reliability is often un

safetyarxiv-cs-cv

When Is an SAE Feature Interpretable? A Validation Ladder for EEG Foundation Models

arXiv:2609.34091v1 Announce Type: new Abstract: Sparse autoencoders (SAEs) decompose dense model activations into discrete latents, making individual features easy to interpret--and easy to misinterpr

When Keywords Drop but Classifiers Hold: Soft Refusals under KV Cache Compression

arXiv:2609.31678v1 Announce Type: new Abstract: KV cache compression is widely used for long context LLM inference under memory constraints, while deployed systems typically score refusals after gener

When Known Physics Helps Neural PDE Models: Residual Constraints Out-Regularize Generic Priors for Nonlinear Dynamics

arXiv:2609.34012v1 Announce Type: new Abstract: Neural PDE surrogates increasingly incorporate structural priors, yet it is often unclear whether their gains arise from physics-specific information or

Applications

When Less Compute Is More: Adaptive Early Exit Improves Pretrained Outlier Detection

arXiv:2609.32898v1 Announce Type: new Abstract: Pretrained tabular foundation models process every dataset at a fixed depth, with inference costs growing with dataset size. To address this, we present

applicationsarxiv-cs-lg

When Less Data Favors Smaller Teachers: Rethinking Teacher Capacity and Data Selection for Knowledge Distillation

arXiv:2609.34489v1 Announce Type: new Abstract: Data pruning reduces the training cost of knowledge distillation (KD). However, the preferred teacher capacity changes with the data budget: smaller tea

When local gains fail to transfer: Frozen Earth-observation embeddings across wildfires

arXiv:2609.34602v1 Announce Type: new Abstract: Frozen Earth-observation embeddings are judged almost entirely by spatially blocked cross-validation inside one study region. We show that this number d

When Noise Meets Long-Tail: Feature-Threshold Dual Calibration for Robust Pseudo-Labeling

arXiv:2609.33668v1 Announce Type: new Abstract: Pseudo-labeling has become a cornerstone of learning from unlabeled data in semantic segmentation. Yet its effectiveness drops sharply in real-world sce

safetyarxiv-cs-cv

Local Ai

When Privacy Moves ML-Mediated Decisions On Device: Information and Incentive Misalignment in Auctions

arXiv:2609.33312v1 Announce Type: cross Abstract: Moving ML-mediated decision making onto privacy-preserving clients decentralises the economic decision along with the inference. Shared budget constra

local-aiarxiv-cs-lg

When Retrieval Hurts: Measuring and Explaining Retrieval-Induced Hallucination in Chest X-ray Report Generation

arXiv:2609.31733v1 Announce Type: new Abstract: Retrieval-augmented generation is an attractive way to improve chest X-ray reporting, because reports from similar prior studies supply clinical context

researcharxiv-cs-cv

When Should a Human Take Back Control? Optimal Delegation under Turbulent AI Risk

arXiv:2609.32083v1 Announce Type: new Abstract: Deploying AI systems requires deciding when to delegate tasks and when humans should intervene to monitor and mitigate risk induced by AI operations. Th

When Should a Satellite Estimate Be Changed? Stress-Testing Neural Corrections for Evapotranspiration

arXiv:2609.35314v1 Announce Type: new Abstract: Neural residuals can improve satellite evapotranspiration (ET) estimates, but selectors must predict when a correction helps and reject unsupported inpu

When Should Humans Step In? Optimal Human Dispatching in AI-Assisted Decisions

arXiv:2603.13688v2 Announce Type: replace-cross Abstract: AI systems increasingly assist decision making by producing cheap factor-level assessments of complex inputs, but these assessments can be bia

Applications

When Should the Count Change? Learning State Maintenance for Causal Video Counting

arXiv:2609.35416v1 Announce Type: new Abstract: Continuous video counting requires distinguishing new observations from new objects or completed events. We introduce StaMina (State Maintenance), which

applicationsarxiv-cs-cv

When Sparse Reward Meets Dense Distillation: Training Dynamics of On-Policy Distillation

arXiv:2609.34849v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards provides a sparse post-training signal: a single binary outcome evaluates the entire rollout, and every t

When Text Matters: Design Principles for Visual Token Pruning in Vision-Language Model

arXiv:2609.34861v1 Announce Type: new Abstract: Visual token pruning has been widely studied as a practical approach to reducing the computational cost of large vision-language models. However, it str

researcharxiv-cs-cv

When the Merge Coefficient Stops Mattering: Proximity Regularized Merging for Continual LoRA Adaptation

arXiv:2609.32332v1 Announce Type: new Abstract: Rehearsal-free continual learning with parameter-efficient adapters can be cast as a sequence of task-vector write-in operations: for each new task, a l

When the Score Becomes the Target: Rethinking Metric Validity in Autonomous Driving

arXiv:2609.34440v1 Announce Type: new Abstract: Driving benchmark scores are increasingly used not only for evaluation but also as optimization targets. This raises a fundamental question: do score ga

When to Commit and When to Defer: Maturing Markov Decision Processes under Refining Information and Expiring Opportunities

arXiv:2606.18820v2 Announce Type: replace Abstract: Sequential decisions often become better informed as action opportunities expire. We introduce Maturing Markov Decision Processes (MMDPs), a structu

When to Evict, Not What to Keep: Draft-Guided Eviction for Training-Free KV-Cache Compression

arXiv:2609.33334v1 Announce Type: new Abstract: Training-free KV-cache compression methods such as SnapKV, H2O, and PyramidKV evict tokens at the end of prefill, aiming to preserve the attention mass

When VLMs Trust Context: Evaluating Scene Text Recognition under Misleading Context

arXiv:2609.34781v1 Announce Type: new Abstract: Vision-language models (VLMs) can read text in natural scenes, but their predictions may be influenced by the surrounding context. When the printed text

When World Models Lie: Adaptive Safety Analysis Under Wrong Imaginations

arXiv:2609.34300v1 Announce Type: new Abstract: World models offer a powerful substrate for safety reasoning in high-dimensional robotic systems, but they are also fallible: their predictions can be b

safetyarxiv-cs-ro

Where Activation Sparsity and KV-Cache Sparsity Cross in LLM Decoding

arXiv:2609.33889v1 Announce Type: new Abstract: At each step, decoding one sequence with a large language model rereads the projection weights, whose traffic is fixed, and the key-value (KV) cache, wh

Local Ai

'Where Can I Trust You?': Boundary-Aware Evaluation of Surrogate Fidelity

arXiv:2609.32230v1 Announce Type: new Abstract: Surrogate models are commonly evaluated by how often they agree with their teacher model over an evaluation set. Local variation in this agreement is we

local-aiarxiv-cs-lg

Where Does the Watermark Hide? Push-Pull Disentanglement for Invisible Watermark Removal

arXiv:2609.31722v1 Announce Type: new Abstract: Fixed image distortions do not cover an attacker that learns from paired clean and watermarked images. We study this paired-training threat with single-

researcharxiv-cs-cv

Where Memory Belongs: Ledger, an Object Ledger for Memory-Augmented VLAs

arXiv:2609.34554v1 Announce Type: new Abstract: Memory is essential for long-horizon, partially observed robotic manipulation: a robot must remember which object was placed in a drawer, whose cup it m

safetyarxiv-cs-ro

Which Self-Improvements Should We Trust? Reliable Self-Improvement When Agents Reuse Their Benchmarks

arXiv:2609.33180v1 Announce Type: cross Abstract: As recursive self-improvement (RSI) rapidly advances, reliable evaluation becomes critical for guiding adaptive search. RSI typically relies on finite

Which the Eye Fears: Writing with Read-Blindness Explains Massive Activations in Transformers

arXiv:2609.35630v1 Announce Type: cross Abstract: Massive activation features (MAs) in Transformers are extreme-value residual-stream features that persist across layers despite the model's ability to

WhiteCon: Semi-Supervised Domain Adaptation Regression Through Whitening Transform and Dual Consistency

arXiv:2609.34078v1 Announce Type: new Abstract: Domain adaptation is crucial for addressing distributional shifts that degrade model performance across domains. While most existing research has center

Who Gets a Token, and What Does It Carry? Unequal Name Support and Concept Access in Large Language Models

arXiv:2609.34065v1 Announce Type: cross Abstract: Names are personal identifiers, but they also carry social meaning and are widely used to evaluate how language models treat different people. Such ev