Semantic Resonance Architecture routes tokens to experts in MoE models via cosine similarity to learnable semantic anchors, improving interpretability of routing decisions.
AISysRev is a containerized LLM-based tool for automated title-abstract screening in systematic literature reviews.
DeepPrune optimizes parallel LLM reasoning by eliminating inter-trace redundancy, reducing 80% computational waste in multi-trace inference.
LLM watermarking technique using syntactic predictability for verifiable output attribution while preserving text quality.
E2EDev benchmark evaluates LLMs on end-to-end software development tasks with fine-grained requirements and behavior-driven evaluation.
PatMD detects harmful memes using MLLMs and misjudgment risk patterns. ML research on content moderation.
Chart2Code benchmark evaluating chart understanding and code generation in multimodal LLMs across three difficulty levels with real-world scenarios.
Vector symbolic architectures using histogram recovery from random linear codes for neurosymbolic AI systems and hardware implementations.
Interlat framework enables LLM-based agents to communicate directly in latent space bypassing tokenization, improving information depth for collaborative problem-solving.
VLA-Pilot enables inference-time steering of vision-language-action models for robotic manipulation without fine-tuning via embodied evolutionary diffusion.
DocVAL distills validated chain-of-thought reasoning from large VLMs to compact models for document VQA with precise spatial grounding and low latency.
End-to-end circuit analysis system leveraging multimodal LLMs for diagram interpretation and mathematical reasoning in engineering problem solving.
Hierarchical RL-MPC fusion approach for sample-efficient planning problems that couples reinforcement learning with model predictive control.
Few-shot prompting approach using LLMs for automated neural network architecture design in computer vision with efficient validation methods.
RLPO applies listwise preference optimization to long-context e-commerce review ranking, improving calibration of top-k rankings over pointwise approaches.
Adaptive layer selection method for layer-wise token pruning in LLM inference that learns which layers benefit most from KV cache reduction.
ORBIT enables controllable multi-budget reasoning in Large Reasoning Models by dynamically balancing exploration-exploitation for appropriate inference-time computation allocation.
TopoDIM framework optimizes communication topology in LLM-based multi-agent systems using one-shot diverse interaction mode generation for reduced latency and computation.
LLMOrbit comprehensive circular taxonomy surveying 50+ LLMs from 2019-2025 across architectural innovations, scaling approaches, and agentic AI system evolution.
HERMES uses hierarchical KV cache memory for efficient streaming video understanding in multimodal LLMs without training, maintaining real-time performance and low GPU overhead.
Multi-Persona Thinking framework reduces social bias in LLMs through inference-time multi-perspective reasoning guidance with contrasting viewpoints.
LLM-driven heuristic design framework that uses LLMs to generate and refine combinatorial optimization solvers with dynamics-aware evaluation beyond endpoint-only metrics.
POP proposes stage-aware structured pruning for LLMs/VLMs that differentiates between prefill and decode stages to improve inference efficiency while maintaining accuracy.
Study of attack selection vulnerability in monitored AI systems, analyzing how AI can evade detection by selecting which attacks to execute.
Framework combining foundation models with imitation learning for open-vocabulary robot skill adaptation from limited data for industrial deployment.
Evaluation framework for persona-adaptive LLM agents in dual-control settings with multimodal capabilities.
Analyzes prompt injection attacks as role confusion where models infer text source from style not origin.
CAP-TTA test-time adaptation framework for debiasing LLMs on out-of-distribution high-bias prompts using LoRA.
V-Reflection enables MLLMs to actively interrogate visual inputs and reduce hallucinations via visual reasoning.
Multi-bit watermarking method for LLM-generated text enabling attribution and tracing with improved efficiency.
VideoStir uses spatio-temporally structured RAG to scale MLLMs to long videos while preserving structure.
Technical report on client-side optimization for LLM-based agents deployed in real-world applications.
Unified framework connecting post-training methods for LLMs including SFT, preference optimization, RL, and process supervision.
Studies inference-time safety improvements in LLMs via deliberative alignment and attribution of unsafe behavior to base models.
Defense framework (Critical-CoT) against reasoning-level backdoor attacks in LLMs using chain-of-thought robustness.
Discriminative reward model scoring multiple LLM responses in single forward pass for efficient comparative reasoning.
Method to induce dialogue strategies from expert transcripts using LLMs for non-collaborative dialogue agents.
Benchmark (METER) evaluating causal reasoning capabilities in LLMs across three levels of causal hierarchy.
Studies forgetting dynamics in fine-tuned image classifiers across architectures to improve curriculum design and data pruning.
Multi-agent framework (C.E.H.) autonomously built and tested a 295-test RAG codebase using an 80B local model. Open-source implementation.
Technical analysis demonstrating how easy it is to poison AI training data via fake web content. Includes practical attack example.
Research paper on reinforcement learning approach (Sim2Reason) for solving Physics Olympiad problems using physics simulators. Code and paper available.
Aurora: research on speculative decoding for LLM inference with draft model staleness solutions. 40+ models benchmarked. Paper and code available.
Opinion piece on implications of AI making software development 10x cheaper and faster. Discussion of SaaS business model impact.
Developer tool using AST and LLM to navigate large codebases by intent. Creates navigable summaries for faster code exploration.
Systems-level technical overview of LLM inference infrastructure, serving runtimes, and ecosystem from infrastructure developer perspective.
W3C WebMCP standard enables AI agents to discover business tools via signed manifests, allowing agents like ChatGPT/Claude to complete tasks like booking services.
Parcae: stable looped language model architecture achieving Transformer-scale quality with fewer parameters and predictable training.
Platform with 7 parallel AI agents executing specialized tasks simultaneously on complex problems using Groq and Cerebras.
Using sandboxes for LLM agents to safely execute autoresearch: agents propose code changes, run them, measure metrics iteratively.