6/1/2026
EchoRL: Reinforcement Learning via Rollout Echoing
Reinforcement learning method via rollout echoing addresses reward signal collapse in LLM post-training for improved reasoning capabilities.
6/1/2026
Spectral Reach: Understanding Neural Scaling as Progress into the Spectral Tail
Introduces spectral position metric to explain neural scaling laws via eigenvalue distribution of empirical neural tangent kernel.
6/1/2026
Toward Identifiable Sparse Autoencoders
Sparse autoencoders for interpreting neural network representations lack stability across training runs; characterizes instability sources and proposes solutions.
6/1/2026
Why Linear Recurrent Memory Works in Partially Observable Reinforcement Learning
Theoretical analysis of linear recurrent networks as memory units in partially observable RL, proving they can reproduce belief state logits in HMMs.
6/1/2026
Algorithmic Recourse of In-Context Learning for Tabular Data
In-context learning enables LLMs to perform tabular prediction without fine-tuning; studies post-hoc recourse methods for high-stakes predictions on tabular data.
6/1/2026
Survival Reinforcement Learning: Toward Scalable Self-Supervised RL
Self-supervised contrastive RL framework with 64+ layer networks for long-horizon planning using survival value learning instead of contrastive losses.
6/1/2026
The Terminal Representation in Reinforcement Learning
arXiv paper on representation learning in RL using successor and default representations for spatio-temporal abstraction and credit assignment.
6/1/2026
Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning
Study of localized failures in machine unlearning showing pointwise prediction discrepancies masked by aggregate evaluation metrics.
6/1/2026
Generalized Intention Modeling in Multi-Agent Reinforcement Learning
Opponent modeling approach for multi-agent reinforcement learning that generalizes intent representation beyond fixed a priori episode information.
6/1/2026
Inconsistency-Aware Minimization: Improving Generalization with Unlabeled Data
Generalization improvement method using unlabeled data and local inconsistency measures derived from information-geometric analysis.
6/1/2026
dashi: A Python library for Dataset Shift Characterization to Support Trustworthy AI Development and Deployment
Python library for characterizing dataset shifts and temporal/multi-source distribution changes to support robust AI development and deployment.
6/1/2026
Trading Complexity for Expressivity Through Structured Generalized Linear Token Mixing
Structured token mixing framework for language models balancing decoding speed and memory efficiency in long-range dependency learning.
6/1/2026
A Unifying View of Variational Generative Wasserstein Flows
Unified theoretical framework for generative models using Wasserstein gradient flows and Jordan-Kinderlehrer-Otto discretization scheme.
6/1/2026
Scaling Higher-Order Graph Learning with Maximal Clique Complexes
Scalable higher-order graph neural networks using maximal clique complexes and simplified Weisfeiler-Leman tests for improved expressivity.
6/1/2026
Constrained Multi-Objective Reinforcement Learning with Max-Min Criterion
Multi-objective reinforcement learning framework combining max-min fairness criterion with explicit constraint satisfaction for policy optimization.
6/1/2026
DG-CoLearn: An Efficient Collaborative Learning Framework for Dynamic Graphs
Collaborative learning framework for dynamic graph models with privacy-preserving partitioned data training on evolving graph structures.
6/1/2026
DRIFT: Decoupled Rollouts and Importance-Weighted Fine-Tuning for Efficient Multi-Turn Optimization
Efficient multi-turn LLM optimization method combining decoupled rollouts with importance-weighted fine-tuning for interactive reinforcement learning settings.
6/1/2026
PithTrain: A Compact and Agent-Native MoE Training System
Compact MoE training framework designed to be agent-native, enabling AI coding agents to automate training-framework development for mixture-of-experts models.
6/1/2026
GPU Forecasters: Language Models as Selective Surrogates for Kernel Runtime Optimization
Using LLMs as surrogate models to predict GPU kernel runtime, reducing optimization costs for kernel search with evolutionary and coding agents.
6/1/2026
Balanced LoRA: Removing Parameter Invariance to Accelerate Convergence
Improved LoRA fine-tuning method addressing parameter invariance through balanced conditioning to accelerate convergence of large language models.
6/1/2026
Assign and Add: A Mechanistic Study of Compositional Arithmetic
Mechanistic study of how transformers compose skills for arithmetic tasks, analyzing compositional generalization in LLMs through controlled experiments.
6/1/2026
Scalable Inference-Time Annealing with Surrogate Likelihood Estimators
Methods for efficiently sampling molecular Boltzmann distributions using iteratively finetuned diffusion models with importance sampling.
6/1/2026
On Efficient Scaling of GNNs via IO-Aware Layers Implementations
Research on optimizing Graph Neural Network scalability through I/O-aware kernel implementations to reduce memory bottlenecks in sparse, irregular computations.
6/1/2026
Skill Reuse as Compression in Agentic RL
ReuseRL framework for training LLM agents via RL using Minimum Description Length principle to learn reusable skill dictionaries.
6/1/2026
On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoders
Studies feature death in sparse autoencoders that decompose neural network activations into interpretable features, analyzing activation outliers.
6/1/2026
Positional versus Symbolic Attention Heads: Learning Dynamics, RoPE Geometry, and Length Generalization
Analysis of positional vs symbolic attention heads in Transformers, studying learning dynamics and length generalization with RoPE.
6/1/2026
Effective Biological Representation Learning by Masking Gene Expression
Masked representation learning approach for gene expression data using foundation models, addressing technical noise in RNA sequencing.
6/1/2026
Giving Sensors a Voice: Multimodal JEPA for Semantic Time-Series Embeddings
CHARM: Transformer encoder with channel-level textual descriptions for multivariate time-series representation learning using JEPA.
6/1/2026
A Tight Theory of Error Feedback Algorithms in Distributed Optimization
Theoretical analysis of error feedback algorithms in distributed optimization, comparing variants for communication-efficient gradient compression.
6/1/2026
TRINE: A Token-Aware, Runtime-Adaptive FPGA Inference Engine for Multimodal AI
TRINE: single-bitstream FPGA accelerator for end-to-end multimodal AI inference combining ViTs, CNNs, GNNs without reconfiguration.
6/1/2026
Kernel Foundry: A Diagnosis-driven Evolutionary Kernel Optimizer with Multi-Experts
Kernel Foundry: diagnosis-driven evolutionary framework combining LLMs and retrieval-augmentation for automatic GPU kernel optimization.
6/1/2026
Gradient-Free Training of Spiking Neural Networks via Low-Rank Evolution Strategies
Gradient-free training approach using low-rank evolution strategies for spiking neural networks on neuromorphic hardware.
6/1/2026
Evolutionary Algorithm for Reservoir Learning and Yielding
EARLY framework using evolutionary algorithms to optimize architecture and hyperparameters of Echo State Networks for temporal learning tasks.
6/1/2026
Smaller and Faster 3DGS via Post-Training Dictionary Learning
Dictionary learning compression technique for 3D Gaussian Splatting models to reduce memory footprint and improve deployment efficiency.
6/1/2026
Exploring Autonomous Agentic Data Engineering for Model Specialization
Framework for autonomous LLM-based agents to execute end-to-end data engineering pipelines for model specialization without human-designed workflows.
6/1/2026
Generalistic or Specific Embeddings, Which is Better? An Empirical Study on Search for Clinical Coding in Non-English Languages
Evaluates generalistic vs. domain-specific embeddings for clinical coding retrieval in non-English languages, proposes LLM data generation to improve recall.
6/1/2026
True Self-Avoiding Walk for Accelerating Markov-Chain Monte Carlo Integration
True self-avoiding walk mechanism for improving empirical integral estimation via adaptive MCMC with penalized transition probabilities.
6/1/2026
Speculative Decoding Across Languages
Compares strategies to improve speculative decoding efficiency for multilingual LLM inference, addressing poor draft model performance in non-English languages.
6/1/2026
ZAPS-DA: Zero-Phase Action Policy Smoothing with Decoupled Actor for Continuous Control in Reinforcement Learning
ZAPS-DA reduces action jitter in continuous control policies without phase lag by decoupling smoothness from RL gradient via framework approach.
6/1/2026
CacheProbe: Auditing Prompt Cache Isolation in Gateway APIs
CacheProbe audits prompt cache isolation in LLM gateway APIs, exposing timing attack and metadata disclosure vulnerabilities in caching implementations.
6/1/2026
Reward Learning from Best-of-$N$ Preference Data: Targets, Tradeoffs, and Design Principles
Analyzes Bradley-Terry reward learning from best-of-N preference data, clarifies target quantity and design principles for N and base distribution selection.
6/1/2026
The Architecture of Errors: From Universal Impossibility to Patch-Local LLM Reliability
Demonstrates universal LLM reliability is impossible across infinite tasks, proposes operationally-bounded patch approach for deployed system error management.
6/1/2026
Rationalize: Shared Semantic Reasoning for Human-AI Alignment
Rationalize framework enables shared semantic reasoning between humans and LLMs through complementary role pairs for collaborative data-driven sensemaking.
6/1/2026
Depth-Dependent Indirect Prompt Injection in Tool-Calling ReAct Agents: Injection Depth, Payload Framing, and Turn-Budget Sensitivity
Characterizes indirect prompt injection attacks on ReAct agents via tool observation loops, analyzing injection depth, payload framing, and turn-budget sensitivity.
6/1/2026
Mathematical Morphology in Machine Learning
Applies mathematical morphology theory to clustering algorithm preserving shape/density with intrinsic noise removal and diverse growth patterns.
6/1/2026
Equivariant Latent Alignment via Flow Matching under Group Symmetries
Equivariant latent alignment via flow matching under group symmetries for geometry-aware generative models and novel view synthesis.
6/1/2026
SAGE: A Novelty Gate for Efficient Memory Evolution in Agentic LLMs
SAGE: novelty-detection gate using von Mises-Fisher density estimation to manage memory evolution in agentic LLMs with principled write-side control.
6/1/2026
Is the Last Layer Sufficient for Uncertainty Quantification?
Analyzes whether final layer is sufficient for uncertainty quantification in DNNs via Bayesian GLM linearization for safety-critical AI deployment.
6/1/2026
Pairwise Reference Alignment as a Model-Level Ordinal Observable
Formalizes pairwise reference alignment as ordinal observable for model evaluation, clarifying what quantity is measured in preference-based LLM alignment.
6/1/2026
XLGoBench: Detecting cross-lingual skill gaps with algorithmic tasks
XLGoBench synthetic benchmark detects cross-lingual capability gaps in LLMs through algorithmic tasks with scalable complexity levels.