The Archive

Utility-Verified Abstraction teaches code agents to query cloud models while masking private code via role/type abstraction instead of identifier removal.

ZO-FO hybrid framework systematically explores adversarial jailbreak attacks on LLM agents by coordinating zeroth-order and first-order optimization strategies.

TinyLlama-1.1B's tokenizer severely penalizes Indic scripts (22.6× subword bloat for Kannada vs. 1.34× for Latin), undermining reported efficiency metrics.

System enables LoRA fine-tuning on phone NPUs by computing input gradients using existing inference-held weights, avoiding device upload.

Propensity standardization technique identifies and mitigates fine-grained fairness disparities in conformal prediction beyond Mondrian calibration.

Online error localization framework for streaming ASR enables targeted speaker clarification by identifying low-confidence spans before propagation.

Risk-sensitive LLM-judge framework for medical QA evaluates six harm categories (directives, contraindications, dosage) beyond factual correctness.

Closed-loop myoelectric control study compares calibration-free cross-user models against calibrated baselines on 306-user EMG dataset.

Table serialization format (row-major vs. column-major) significantly affects instruction-tuned LLM's table lookup accuracy depending on table aspect ratio.

Sampler-induced bias in diffusion language models: commit rules favoring high-probability tokens underrepresent minority groups in generation.

Reciprocal Q-Augmentation enables multi-agent cooperation in general-sum games without shared rewards via latent opponent modeling.

Empirical study of 1,500 LLM refactoring attempts across 15 Java repos shows multi-agent approaches improve dependency-aware code generation over single-prompt baselines.

IndustrialCAD-1400 benchmark evaluates text-to-CAD models on 1,400 detailed industrial assembly requests with 17+ component specifications and numeric constraints.

Theoretical analysis of entropy rates in random composition trees as models of hierarchical text structure, with limited practical implications for LLM development.

Proposes Martingale Posterior Cox Processes for Bayesian nonparametric inference on event data, primarily theoretical contribution with limited LLM relevance.

Demonstrates data-poisoning attacks that plant Trojans increasing LLM response length and token costs, and proposes defenses against semantic trigger activation.

Legal analysis arguing AI rights are compatible with mandatory safety off-switches through rights-balancing and due-process frameworks.

Meta-post anticipating API usage quota reset.

SiamSA-PPM framework uses self-supervised Siamese learning and semantic-aware augmentation for predictive process monitoring with limited labeled data.

Theoretical convergence analysis of deep networks via Langevin dynamics and thermodynamic inequalities; incremental contribution to optimization theory.

Evidence Field architecture produces interpretable medical image classifications via spatial programs and evidence fields, bridging prediction and explainability.

DialogMem benchmark evaluates embodied agents integrating dialogue and visual memory across 142 episodes for object navigation and QA in 3D environments.

ECHO replaces multi-token prediction auxiliary modules with efficient self-recursion via dynamic depth sampling, reducing inference cost.

BanglaBBQ introduces a bias benchmark for QA systems in Bangladeshi context, extending KoBBQ and PakBBQ with localized stereotypes.

Paired-outcome learning optimizes spacecraft attitude control under torque constraints via selective controller modification.

Fisher-Geometric Pruning post-hoc reduces overparameterized symbolic regression models by exploiting information geometry.

CorrFlow augments flow-matching generative models with a learned refiner to correct velocity predictions using intermediate representations.

ReTRACE integrates LLM agents with chemistry tools and route scoring for retrosynthesis planning, handling stereochemistry and cost tradeoffs.

Study reveals token count and aggregate accuracy mask true VLM pruning quality; proposes per-example and latency-aware evaluation metrics.

Collective Model Intelligence framework characterizes diverse model capabilities and orchestrates complementary models for collaborative inference.