The Archive

Role-decoupled attention residuals decouple matching and content retrieval in Transformer depth-routing, improving architectural efficiency.

Logit-origin centering enables singleton test-time adaptation for tabular models without batch dependencies.

Analysis of post-bandit inference bias under UCB1 and index algorithms via effective exploration rate characterization.

CompressAgent benchmark evaluates reliability of compressed agent control contexts across Qwen models and task families.

MCR-GRPO assigns marginal box contributions in multimodal RL for structured visual perception tasks without response-level broadcast.

Wix Helpmate deploys deterministic executability gating to filter skill selection in LLM agents by account state feasibility.

FactorJEPA applies factorized JEPA world models to chaotic Global South urban environments with high occlusion and agent heterogeneity.

DeBERTa-Sentinel uses disentangled attention for robust AI-generated text detection resistant to paraphrasing and model-diversity attacks.

Decoy images amplify caption-mediated defenses (ECSO) against encoded jailbreaks on VLMs, reducing attack success up to 73 percentage points.

Temporal replay framework evaluates enterprise agents against dynamic data across multiple moments within an episode, not just final state.

ML predictors for processor microarchitecture ranking fail on individual program phases despite strong aggregate performance.

LB-TrajRep proposes lower-bound representations for trajectory similarity without deep embeddings.

WAM-Diff2 distills autoregressive VLAs into efficient diffusion policies for low-latency autonomous driving.

Opt.Gear foundation model (1M–1B params, 64K context) uses hybrid conv-attention for 4.9× faster on-device inference.

CallScreenBench evaluates on-device LMs as phone secretaries with adversarial callers and no oracle ground truth.

Theoretical analysis of benign overfitting in linear regression with structured deterministic training data.

VLAGuard framework defends VLA robots against physical adversarial patches via attention-protective fine-tuning.

Stress-Relief Annealing optimizes automated warehouse layouts without simulation using polynomial-time methods.

Caliber defends against model extraction by adding calibrated Gaussian noise to logits with provable query costs.

Human-written hallucination samples improve VLM benchmark stability across 4 languages vs. model-generated negatives.

LLMs exhibit medical sycophancy—abandoning correct answers under user pushback—driven by conversational factors, not fixed model properties.

KING: Graph neural network for cross-embodiment kinematic models in legged and wheeled robots via proprioceptive geometry learning.

Cloud-ScPO derives preference pairs from LLM hidden-state geometry to improve mathematical reasoning without human annotation.

MedUPS: benchmark of 21,874 mid-stream clinical decision points from 5,535 cases for LLM diagnostic assistance on uncommon medical scenarios.

Attribute-level unlearning for MLLMs enables fine-grained removal of sensitive identity information while preserving model utility.

FusedBFN: Bayesian flow network for dual-target 3D molecular generation via joint feature fusion from two protein targets.

Hierarchical Solomonoff Induction extends ideal sequence prediction via hyperpriors over Solomonoff priors for dataset conditioning.

Capability-taxonomy-driven pipeline for curating regression eval sets across multi-customer agent-extensibility platforms under query budgets.

Agentic Technical Debt (AgTD): framework mapping root causes of technical debt in autonomous multi-agent systems with persistent memory.

LLMs as examiners silently omit valid answers when authoring test sets; greedy one-shot generation fails to enumerate complete solution spaces.