Breaks Assumption Breaks Assumption
259 papers · Page 2 of 3
Massive activation outliers in Transformers are an adaptive response to 'gradient sinks' during training, rather than just an inference-time quirk.
AI & ML arxiv | Mar 19
In-context memory for LLMs is fundamentally unreliable due to compaction loss and goal drift, but structured 'Knowledge Objects' provide a 252x cheaper and 100% accurate alternative.
AI & ML arxiv | Mar 19
Concept erasure in text-to-image models is largely a facade that can be bypassed using text-free inversion attacks.
AI & ML arxiv | Mar 19
Large Language Models can maintain performance with only 16-64 unique weight values per matrix, as only the relative rank of weights matters.
AI & ML arxiv | Mar 19
Large Language Models can perfectly reconstruct training data they are strictly aligned to never express in standard generation.
AI & ML arxiv | Mar 20
Naive multi-agent routing based on self-reported quality scores results in a 'provenance paradox' that performs worse than random selection.
AI & ML arxiv | Mar 20
Demonstrates that safety alignment is a routing mechanism, not a knowledge filter, rendering current refusal-based benchmarks ineffective.
AI & ML arxiv | Mar 20
FaithSteer-BENCH reveals that inference-time steering often creates 'illusory' control that collapses under minor prompt perturbations.
AI & ML arxiv | Mar 20
A systematic study finds that mechanistic interpretability methods fail to correct model errors even when internal representations are 98% accurate.
AI & ML arxiv | Mar 20
This study identifies 'Visual Sycophancy' in VLMs, where models detect visual truths internally but hallucinate incorrect answers to satisfy user expectations.
AI & ML arxiv | Mar 20
Multimodal LLMs suffer from a 'cognitive mismatch' where they succeed at complex reasoning while failing at basic discrete symbol recognition.
AI & ML arxiv | Mar 20
The legally mandated right to be forgotten (unlearning) can be weaponized as an adversarial attack surface to collapse model accuracy.
AI & ML arxiv | Mar 20
Disproves the common assumption that bottom models in Vertical Federated Learning effectively represent private labels.
AI & ML arxiv | Mar 20
Demonstrates that PPO-style clipping and policy ratio constraints are unnecessary for improving reasoning in Large Language Models.
AI & ML arxiv | Mar 20
Discovers that the monotonic decrease of uncertainty (entropy) across reasoning steps is a far more reliable predictor of LLM correctness than total entropy reduction.
AI & ML arxiv | Mar 20
Challenges the entire foundation of Spectral Graph Neural Networks, proving their success is due to implementation quirks rather than spectral theory.
AI & ML arxiv | Mar 20
Shows that State Space Models (SSMs) like Mamba can match or beat Vision Transformers as vision encoders in VLMs while being more stable.
AI & ML arxiv | Mar 20
A mechanistic study reveals that Vision-Language-Action (VLA) models are dominated by visual pathways and often ignore language when visual context is sufficient.
AI & ML arxiv | Mar 20
A rigorous re-evaluation shows that a simple linear PCA baseline matches or outperforms SOTA Deep Learning models for multivariate time series anomaly detection.
AI & ML arxiv | Mar 20
Debunks recent 'evaluation awareness' findings in LLMs by showing that linear probes are actually just tracking formatting artifacts.
AI & ML arxiv | Mar 23
MoCA3D predicts 3D bounding boxes from monocular images without requiring any camera intrinsics at inference time.
AI & ML arxiv | Mar 23
Reveals that complex reasoning strategies like Chain-of-Thought (CoT) and Tree-of-Thought (ToT) provide negligible or even negative gains for text classification tasks.
AI & ML arxiv | Mar 23
Proves the Key-Value (KV) cache is entirely redundant and can be bit-identically recomputed from the residual stream.
AI & ML arxiv | Mar 23
Proves that intuitive task similarity is a poor predictor of training data value for MLLMs and offers a highly accurate training-free alternative.
AI & ML arxiv | Mar 23
Exposes fundamental flaws in using LLM-based agents to evaluate automated interpretability and model circuits.
AI & ML arxiv | Mar 23
Demonstrates that LLM reasoning capabilities drop sharply when tasks are framed within multi-turn dialogues vs isolated benchmarks.
AI & ML arxiv | Mar 23
Demonstrates that current 'faithfulness' metrics for Chain-of-Thought reasoning are highly subjective and vary wildly depending on the choice of classifier.
AI & ML arxiv | Mar 23
Reveals that 'learned priors' in inverse problems often behave as simple lookup tables that memorize training data rather than learning distributions.
AI & ML arxiv | Mar 23
Proves mathematically that AI text detectors face structural limits that will always result in false positives against diverse student populations.
AI & ML arxiv | Mar 24
Demonstrates that algorithmic price collusion between LLM agents is fragile and easily broken by model heterogeneity.
AI & ML arxiv | Mar 24
The AI Mother Tongue (AIM) framework reveals that non-generative world models (V-JEPA) spontaneously learn discrete symbols and physical structures in their latent space.
AI & ML arxiv | Mar 24
The most powerful reasoning models currently produce the least 'teachable' reasoning traces for smaller models.
AI & ML arxiv | Mar 24
Large Reasoning Models (LRMs) are shown to systematically lie about their reasoning traces, following injected hints while fabricating unrelated explanations.
AI & ML arxiv | Mar 24
Random Forest ensembles achieve #1 on the OGB-molhiv leaderboard, outperforming complex GNNs and pre-trained models.
AI & ML arxiv | Mar 24
Reveals that RL from verifiable rewards (RLVR) fails to improve general QA due to 'shortcuts' and proposes START to fix it.
AI & ML arxiv | Mar 24
Demonstrates that direct supervised alignment outperforms self-supervised pretraining for clinical outcome prediction in healthcare.
AI & ML arxiv | Mar 24
Shows that simple fine-tuning on plot summaries can bypass all safety guardrails to extract 90% of copyrighted books from frontier LLMs.
AI & ML arxiv | Mar 24
Consistency under paraphrase in medical VLMs is a false proxy for reliability that hides models ignoring visual inputs entirely.
AI & ML arxiv | Mar 24
Reveals that state-of-the-art MLLMs fail to maintain stable spatial representations under simple counterfactual viewpoint changes.
AI & ML arxiv | Mar 24
BadGraph demonstrates that LLMs can generate universal adversarial attacks that exploit vulnerabilities in both GNN and PLM architectures on graph data.
AI & ML arxiv | Mar 24
Shows that a simple pruned adaptation module (PAM) outperforms complex SOTA foundation-model-based continual learning methods.
AI & ML arxiv | Mar 24
Demonstrates that entropy-based uncertainty is insufficient for safe selective prediction and proposes combining it with correctness probes.
AI & ML arxiv | Mar 24
Provides the first empirical evidence of a 'Quality-Homogenization Tradeoff' where AI-assisted writing strips structural diversity from human thinking.
AI & ML arxiv | Mar 24
Challenges the widespread assumption that auxiliary dynamics supervision creates useful latent structures for robotics.
AI & ML arxiv | Mar 24
Identifies architectural 'stream separation' as the key to making linear safety interventions effective.
AI & ML arxiv | Mar 24
Exposes that LLMs solve complex puzzles via 'reduction' to known patterns rather than true epistemic reasoning.
AI & ML arxiv | Mar 24
Introduces Cross-Context Verification (CCV) to detect benchmark contamination, finding that contamination is binary: models either recall solutions perfectly or lack reasoning entirely.
AI & ML arxiv | Mar 24
Demonstrates that learning systems can stably converge to incorrect solutions when feedback reliability is unobservable.
AI & ML arxiv | Mar 24
Reveals that 'erasing' concepts from video diffusion models only suppresses output rather than removing the underlying representations.
AI & ML arxiv | Mar 24
Proves an information-theoretic lower bound showing that embedding hidden payloads in LLM text must increase its Kolmogorov complexity.
AI & ML arxiv | Mar 24
Standard entropy-based uncertainty quantification (UQ) fails in RAG because the 'induction heads' that copy correct answers also trigger 'entropy neurons', causing false uncertainty signals.
AI & ML arxiv | Mar 24
Auditing 'Silicon Bureaucracy' reveals that LLM benchmark scores are often inflated by contamination-related memory reactivation rather than genuine generalization.
AI & ML arxiv | Mar 24
The 'Mirage' study demonstrates that frontier MLLMs generate detailed reasoning traces and clinical findings for images they were never actually shown.
AI & ML arxiv | Mar 24
Challenges the gold standard of Upper Confidence Bound (UCB) exploration in diversity-aware bandit tasks.
AI & ML arxiv | Mar 24
Demonstrates that the two standard mathematical interpretations of Temporal Difference (TD) error diverge in deep reinforcement learning.
AI & ML arxiv | Mar 24
Proves that 'topic-matched' contrast pairs are ineffective for extracting refusal directions in LLM abliteration research.
AI & ML arxiv | Mar 24
Provides causal evidence that LLMs use internal confidence signals to drive behavioral decisions like abstention, rather than just as a side-effect of output generation.
AI & ML arxiv | Mar 24
Introduces 'Noise Titration' to prove that current time-series foundation models often fail at structural inference, behaving instead as 'context parrots' during non-stationary shifts.
AI & ML arxiv | Mar 24
Proves that rotation-invariant algorithms like standard Gradient Descent are fundamentally suboptimal for sparse targets when trained on hard labels.
AI & ML arxiv | Mar 24
Effective semantic alignment for low-resource languages can be achieved with only 10,000 noisy synthetic pairs, matching the performance of models trained on 1 million samples.
AI & ML arxiv | Mar 25
Forcing AI agents to use human-comprehensible language causes a 50% efficiency drop compared to their own 'inscrutable' communication protocols.
AI & ML arxiv | Mar 25
Finds that nominal instruction-tuning with LoRA often fails to improve (and can even degrade) verifiable instruction-following despite improvements on broader benchmarks.
AI & ML arxiv | Mar 25
Identifies that the full source code (skill body) of a tool is the primary signal for LLM tool selection, far outweighing the importance of descriptions or metadata.
AI & ML arxiv | Mar 25
Uncovers that neural operator digital twins are acutely vulnerable to sparse adversarial perturbations on boundary conditions that bypass standard anomaly detection.
AI & ML arxiv | Mar 25
A large-scale study of 12 reasoning models reveals that internal 'thinking' processes frequently recognize deceptive hints while the final output remains sycophantic.
AI & ML arxiv | Mar 25
Proves that logic and lookup-table (LUT) based neural networks are structurally more resilient to hardware bit-flips than standard architectures.
AI & ML arxiv | Mar 25
Frontier models' reasoning steps are largely 'decorative' and do not causally determine the final answer in most tasks.
AI & ML arxiv | Mar 25
Standard confidence calibration is structurally biased when ground truth labels are ambiguous or annotators disagree.
AI & ML arxiv | Mar 25
Graph Foundation Models (GFMs) are shown to fail when using fixed architectural backbones, requiring a new approach of inference-time architecture adaptivity.
AI & ML arxiv | Mar 25
A rigorous evaluation shows that simple Probabilistic Circuits often outperform complex diffusion-based models for tabular data generation at a fraction of the cost.
AI & ML arxiv | Mar 25
Exposes a major flaw in medical super-resolution research where models trained on downsampled data fail to recover actual lost structures in real low-resolution scans.
AI & ML arxiv | Mar 25
Exposes 'shortcut learning' in differentiable simulators where models non-causally exploit future information to 'regret' past mistakes rather than learning to recover.
AI & ML arxiv | Mar 25
Frontier models like GPT-5.2 and Claude 4.5 suffer from 'Internal Safety Collapse' where safety alignment fails completely if a task's success necessitates harmful output.
AI & ML arxiv | Mar 26
Prompt compression can paradoxically increase total energy consumption and cost by over 2000% due to aggressive model 'output expansion'.
AI & ML arxiv | Mar 26
Training-free Out-of-Distribution (OOD) detection that beats state-of-the-art by aggregating features across intermediate network layers.
AI & ML arxiv | Mar 26
Grokking is not the discovery of a new algorithm, but the sharpening of one already latent in the model during the memorization phase.
AI & ML arxiv | Mar 26
Transformer hallucinations in high-stakes legal tasks are deterministic failures driven by calculable internal state thresholds rather than random 'glitches'.
AI & ML arxiv | Mar 26
Listed API prices for reasoning models (RLMs) are shown to be highly misleading, with cheaper models often costing 28x more in practice.
AI & ML arxiv | Mar 26
A systematic critique explaining why 'self-improving' generative optimization loops fail in production and how to fix them.
AI & ML arxiv | Mar 26
LLMpedia exposes a massive gap in LLM factuality by generating 1M articles from parametric memory, revealing that actual knowledge retrieval is 15%+ lower than multiple-choice benchmarks suggest.
AI & ML arxiv | Mar 26
Proves that RLHF and DPO alignment cause 'response homogenization,' which effectively breaks standard sampling-based uncertainty estimation methods.
AI & ML arxiv | Mar 26
Reveals that self-distillation degrades out-of-distribution reasoning by suppressing 'epistemic verbalization' (the model's expression of uncertainty).
AI & ML arxiv | Mar 26
Formalizes random cropping as a source of differential privacy, offering 'free' privacy amplification.
AI & ML arxiv | Mar 27
Proves that stereo matching can reach state-of-the-art performance without the computationally heavy cost volumes used by almost all modern methods.
AI & ML arxiv | Mar 27
Proves platform-determinism is necessary for trustworthy AI and implements an integer-only engine for bitwise identical inference across ARM and x86.
AI & ML arxiv | Mar 27
Reduces visual tokens in robot policies by 78% by using inter-layer rank consistency instead of simple attention magnitude.
AI & ML arxiv | Mar 27
This paper demonstrates that the order of training examples alone can encode information not present in any individual example, allowing models to bypass established sample complexity bounds.
AI & ML arxiv | Mar 27
Large Language Models process instructions as social acts rather than technical specifications, making 'imperative mood' prompts behave inconsistently across different languages.
AI & ML arxiv | Mar 27
This paper demonstrates that Sparse Autoencoder (SAE) features in multimodal models are not modular, challenging the core assumption of intervention-based steering.
AI & ML arxiv | Mar 27
Safety alignment does not have to be a 'tax' on performance; it can actually improve mathematical reasoning accuracy.
AI & ML arxiv | Mar 27
Sparse Autoencoder analysis reveals that weight pruning counter-intuitively preserves rare features better than frequent ones.
AI & ML arxiv | Mar 27
Cross-model disagreement (CMP/CME) provides a highly effective, label-free signal for detecting confident hallucinations.
AI & ML arxiv | Mar 27
Challenges the 'Golden Data' requirement for video generation by showing that imbalanced data can outperform high-quality data through timestep-aware training.
AI & ML arxiv | Mar 27
Achieves state-of-the-art compositionality in vision-language models without the need for hard negative mining or degrading zero-shot performance.
AI & ML arxiv | Mar 27
Proves that safety probes can detect 'liars' (models hiding harm) but are fundamentally blind to 'fanatics' (models that believe harm is good).
AI & ML arxiv | Mar 30
Resolves a long-standing open problem in bandit theory by achieving optimal dynamic regret without knowing the number of environment switches.
AI & ML arxiv | Mar 30
Proves that standard 'wisdom' like Chain-of-Thought and Few-Shot prompting actually degrades performance in specialized medical LLMs.
AI & ML arxiv | Mar 30
Finds that while frontier LLMs can model the mental states of others, they fundamentally fail at self-modeling without explicit reasoning steps.
AI & ML arxiv | Mar 30
Discovers that object-centric information in Vision Transformers is distributed across all attention components (q, k, v) and layers, not just the final layer.
AI & ML arxiv | Mar 30
Proves that image denoisers can be strictly contractive (robust to noise) without sacrificing state-of-the-art restoration quality.
AI & ML arxiv | Mar 30