Evgeny Kharlamov
Publications
Memory for Large Language Models
Memory has evolved into a foundational architectural dimension in large language models (LLMs), shifting from an implicit byproduct of computation to a spectrum of explicit, controllable mechanisms. While recent advances introduce diverse strategies---spanning transient attention, recurrent state dynamics, parameter-efficient adaptations, and scalable lookup storage---this rapid evolution has led to a highly fragmented research landscape. In this survey, we present a systematic, architecture-centric taxonomy of memory in LLMs. Our framework characterizes memory along three orthogonal axes: representation (implicit versus explicit), update dynamics (offline versus online), and persistence (short-term versus long-term). We further formalize the granular mechanisms dictating memory writing, routing, state transitions, and consolidation. This unified perspective elucidates the conceptual boundaries between computation-coupled and independently addressable memory, effectively bridging disparate architectural paradigms. Additionally, we critically analyze hybrid memory architectures, system-level efficiency trade-offs, and multi-dimensional evaluation methodologies. By consolidating these scattered advancements into a cohesive framework, this survey charts the trajectory of memory-centric LLM design and provides a principled foundation for future innovations in scalable and adaptive language modeling.
MADA-RL: Multi-Agent Debate-Aware Reinforcement Learning for Parameter-Efficient Reasoning in Compact Models
Large language models achieve strong reasoning performance, but often at prohibitive training cost - a challenge that is especially acute for compact models ($\leq 4 \, \mathrm{B}$ parameters) trained under limited budgets. We introduce MADA-RL, a post-training framework that specializes compact models into generator and critic roles and trains them with a debate-aware learning signal, fine-tuning only a small subset of parameters via LoRA adapters. Our central contribution is a counterfactual critic advantage: a dynamic, role-conditioned baseline that redefines the critic's advantage as its reward minus the generator ensemble's per-instance accuracy. This explicitly optimizes critics to improve over generator consensus rather than to merely reproduce a correct answer, yielding more targeted credit assignment than static mean-reward normalization. At deployment, the specialized agents are composed in a lightweight multi-round protocol. Across five mathematical reasoning benchmarks, MADA-RL raises the accuracy of the DeepSeek-R1-Distill-Qwen-1.5B model from $39.9 \, \%$ to $41.9 \, \%$ ($+2.0$ points, $p < 0.001$) using $16$ times fewer trainable parameters than fully fine-tuned baselines, placing it on the accuracy-trainable-parameter Pareto front. It approaches, but does not surpass, the strongest baselines (DeepScaleR, STILL-3), which are trained on substantially larger datasets; we analyse this gap and the associated inference-time cost directly. A controlled study isolates the source of MADA-RL's gains: the counterfactual advantage produces the highest critic improvement rate of any model evaluated, indicating that trained critics learn to correct generator errors rather than to imitate them.
Integrating Meta-Features with Knowledge Graph Embeddings for Meta-Learning
The vast collection of machine learning records available on the web presents a significant opportunity for meta-learning, where past experiments are leveraged to improve performance. Two crucial meta-learning tasks are pipeline performance estimation (PPE), which predicts pipeline performance on target datasets, and dataset performance-based similarity estimation (DPSE), which identifies datasets with similar performance patterns. Existing approaches primarily rely on dataset meta-features (e.g., number of instances, class entropy, etc.) to represent datasets numerically and approximate these meta-learning tasks. However, these approaches often overlook the wealth of past experimental results and pipeline metadata available. This limits their ability to capture dataset - pipeline interactions that reveal performance similarity patterns. In this work, we propose KGmetaSP, a knowledge-graph-embeddings approach that leverages existing experiment data to capture these interactions and improve both PPE and DPSE. We represent datasets and pipelines within a unified knowledge graph (KG) and derive embeddings that support pipeline-agnostic meta-models for PPE and distance-based retrieval for DPSE. To validate our approach, we construct a large-scale benchmark comprising 144,177 OpenML experiments, enabling a rich cross-dataset evaluation. KGmetaSP enables accurate PPE using a single pipeline-agnostic meta-model and improves DPSE over baselines. The proposed KGmetaSP, KG, and benchmark are released, establishing a new reference point for meta-learning and demonstrating how consolidating open experiment data into a unified KG advances the field.