Explore our academic papers, original research, and technical advances in AI memory, learning, and intelligent systems. From new ideas to validated methods, we turn research into technological progress
Research from the MemoraX community in memory, learning, and intelligent systems
Reinforcement learningPost-training
Conference paper /
Squeeze the Soaked Sponge: Efficient Off-policy Reinforcement Finetuning for Large Language Model
ReMix combines mixed-policy updates, a KL convex constraint, and policy reincarnation to reuse off-policy data for reinforcement finetuning. Experiments show competitive reasoning accuracy with fewer rollout samples and training steps
Jing Liang, Hongyao Tang, Yi Ma et al.
Accepted at ICLR 2026
Agent memoryCoding agents
Preprint · Benchmark /
SWE Context Bench: A Benchmark for Context Learning in Coding
SWE-ContextBench evaluates whether coding agents can reuse experience across related software tasks. It links issues and pull requests that share context, enabling evaluation of context retrieval, task resolution, runtime, and token cost
Jiayuan Zhu, Junde Wu, Minhao Hu et al.
Preprint · arXiv:2602.08316
Code agent memory
Preprint /
MemDev-Bench: Does Memory Help Code Agents? Benchmarking Memory Systems over Real Repository Histories
MemDev-Bench comprises 555 development tasks and 4,869 historical tasks drawn from real software development histories to evaluate the impact of memory on coding agents. Memory can improve code localization and editing, but reusing past experience without checking its applicability can also cause failures
Guorui Chen, Tao Liu, Zhenzhen Huang, Yidan Zhang et al.
Preprint · arXiv:2610.27524
Multimodal agent memory
Preprint /
GraphMemix: Query-Aware Evidence Forests for Long-Term Multimodal Agent Memory
GraphMemix organizes multimodal agent memory into query-aware evidence forests, retrieving complementary evidence while reducing redundant lifecycle costs. Experiments on four long-term multimodal memory benchmarks establish a new accuracy–cost frontier
EngramBench: A Capability-Grounded Benchmark for Skill-Evolution Harnesses
EngramBench distinguishes genuine capability abstraction from solution leakage in interactive learning and transfer tasks. Evaluations show that procedural memory can guide agent execution, reduce redundant context, and cut coding time by more than 55%
From Memory to Guide: Spatio-Temporal Composer for Procedural Coding Memory
The spatio-temporal composer dynamically compiles retrieved coding experience into staged, environment-adapted runtime guides. Without updating model parameters, it improves the pass rate on complex EngramBench tasks by 7.2% while reducing token consumption by 32.2%
Beyond the Final Embedding: Retrieving Memories from Token-Level Evidence
This work improves memory retrieval through sparse token-level matching, combining local evidence with overall memory similarity for ranking. The method improves retrieval across eight benchmarks and further enhances long-term memory question answering
Han Luo, Zhan Zhuang, Xu He, Jujia Zhao, Yu Li, Suqin Yuan, Dong Li, Lei Feng
Preprint · arXiv:2610.13277
Memory retrieval
Preprint /
NEAR: Neighborhood-Aware Residuals for Memory Retrieval
NEAR uses information shared among neighboring memories to construct residual representations, helping distinguish records with similar topics but different facts. This training-free method consistently improves retrieval across eight long-context benchmarks and nine embedding models
Han Luo, Zhan Zhuang, Xu He, Jujia Zhao, Yu Li, Suqin Yuan, Dong Li, Lei Feng
Preprint · arXiv:2610.13314
Large language modelsLLM reasoning
Conference paper /
Beyond Correctness: Robustness-Driven Evolutionary Self-Training for Large Language Models
TREST combines evolutionary algorithms with supervised fine-tuning to improve mathematical reasoning in large language models. By prioritizing reasoning trajectories that remain correct under decoding variations, it favors robust, insight-driven strategies over brittle brute-force calculations.
Wei Guo, Hongyao Tang, Yi Ma, Jinyi Liu, Pengyi Li, Jing Liang, Yifu Yuan, Yan Zheng, Jianye Hao
NeurIPS 2026
Large language modelsReinforcement learning
Conference paper /
The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning
Under training–inference mismatch, improving a training policy does not necessarily improve the deployed inference policy. MIPI makes inference-policy improvement the optimization objective, while MIPU combines sampler-referenced updates with inference-gap-aware acceptance to improve reasoning performance and training stability.
Jing Liang, Hongyao Tang, Yi Ma, Yancheng He, Weixun Wang, Xiaoyang Li, Ju Huang, Wenbo Su, Jinyi Liu, Yan Zheng, Jianye Hao, Bo Zheng
NeurIPS 2026
Decision makingReinforcement learningControl
Conference paper /
What Kind of Diffusion Models Do We Need in Online Reinforcement Learning?
This work examines the challenges of probability evaluation, sampling efficiency, and distribution coverage when using diffusion policies in online reinforcement learning. It introduces CoupledFlow to retain expressive policy distributions while enabling tractable log-likelihood evaluation and more efficient policy optimization.
Zihao Wu, Hongyao Tang, Yi Ma, Yan Zheng, Jianrong Wang, Jianye Hao
NeurIPS 2026
Large language modelsDifficulty perceptionFast–slow thinkingInterpretability
Conference paper /
Does Your Large Language Model Have An Intuitive Sense of The Difficulty of A Question?
LLM hidden representations contain clues about question difficulty before explicit reasoning begins. RLDP estimates difficulty from a single forward pass without training or additional reasoning rollouts. RLDP-AdaSwitch uses these signals to switch between fast and slow reasoning, improving inference efficiency.
Ruitao Wang, Jinyi Liu, Hongyao Tang, Rong Cheng, Yi Ma, Shaojin Ma, Yiwen Zhu, Hebin Liang, Gangyi Zhao, Pengyi Li, Yan Zheng, Jianye Hao