MemoraX AI Researchers and Collaborators Have 10 Papers Accepted to NeurIPS 2026

SAN FRANCISCO, Oct. 22, 2026 (GLOBE NEWSWIRE) -- The research explores advances in LLM training, adaptive reasoning, and reinforcement learning.

Ten papers involving researchers from MemoraX AI’s founding research Team and its academic and industry collaborators have been accepted to the 2026 Conference on Neural Information Processing Systems (NeurIPS), one of the leading international conferences in artificial intelligence and machine learning.

The research spans large language model (LLM) post-training, inference optimization, adaptive reasoning, and online reinforcement learning. Across these areas, the work explores how AI systems can learn more reliably, allocate computational resources more effectively, and make better decisions in complex environments.

NeurIPS 2026 will take place December 6–12 in Sydney, Australia, with satellite events in Atlanta, Georgia, and Paris, France, from December 9–13.

Advancing AI Learning, Reasoning, and Decision-Making

As AI agents take on increasingly complex and long-running tasks, researchers face fundamental challenges: making model training more reliable, aligning training objectives with actual inference performance, allocating reasoning resources more effectively, and developing robust decision-making strategies for complex environments.

MemoraX AI is building Next -generation AI memory infrastructure while pursuing research in memory modeling, reinforcement learning, and agent systems. Its research explores how memory can move beyond information storage and retrieval to become a learnable capability that helps agents retain context, manage information, and apply accumulated experience to future tasks.

The company’s work spans text memory, work and coding memory, and multimodal memory. For developers building AI applications, Memory API supports memory writing, search, updates, and deletion. MemoraX Code, the company’s persistent-memory tool for coding agents, helps agents retain repository structure, architectural decisions, coding preferences, and debugging experience across sessions and agents. The underlying technical challenge is how to make memory itself learnable, enabling systems to determine what to retain, when to update or retrieve information, and how to turn historical interactions into reusable experience.

Four Representative Papers

Four representative papers from the NeurIPS 2026 selections illustrate different approaches to improving AI learning, reasoning, and decision-making.

1. Beyond Correctness: Robustness-Driven Evolutionary Self-Training for Large Language Models

Research focus: LLM post-training and reasoning reliability

A correct answer does not necessarily indicate a reliable reasoning process. This paper introduces TREST, a robustness-driven evolutionary self-training framework that evaluates the stability of reasoning paths under sampling perturbations and prioritizes more robust trajectories for training.

Across five mathematical reasoning benchmarks, TREST achieved an average accuracy of 49.83%, outperforming the GRPO and Iterative RFT baselines in the reported experiments.

The work offers an approach to improving the reliability of LLM reasoning by considering not only whether a solution is correct, but also whether the reasoning strategy remains effective under variation.

2. The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning

Research focus: Reinforcement learning and inference optimization

Improvements during training do not always translate into better performance when a model is deployed. This paper examines inconsistencies between training and inference in LLM reinforcement learning, where improvements in a training policy may not translate into better performance from the policy used at inference time.

The researchers propose Monotonic Inference Policy Improvement (MIPI), a policy-improvement principle focused on inference-time performance, and introduce Monotonic Inference Policy Update (MIPU), a framework designed to implement that principle.

Experiments using Qwen3-4B and Qwen3-1.7B showed improvements in average reasoning performance and training stability under the tested conditions. The work provides a framework for better aligning reinforcement learning objectives with model behavior during deployment.

3. Does Your Large Language Model Have An Intuitive Sense of The Difficulty of A Question?

Research focus: Adaptive reasoning and computational efficiency

AI systems do not necessarily need to devote the same amount of reasoning effort to every question. This paper investigates whether LLMs encode signals related to question difficulty in their internal representations and introduces RLDP, a method for estimating difficulty from hidden representations.

Building on this approach, RLDP-AdaSwitch dynamically adjusts reasoning effort according to the estimated difficulty of a question.

The research points toward a more adaptive approach to inference, allowing models to allocate computational resources according to the demands of individual queries rather than applying the same reasoning strategy to every problem.

4. What Kind of Diffusion Models Do We Need in Online Reinforcement Learning?

Research focus: Online reinforcement learning and complex decision-making

In complex environments, AI agents must make decisions across high-dimensional Action spaces. Developing policies that can represent complex Action distributions while remaining efficient to optimize is a key challenge in online reinforcement learning.

This paper examines the design of diffusion models for online reinforcement learning and proposes Coupled Flow, an approach that explores the relationship between expressive Action modeling and efficient policy optimization.

The evaluation covers 36 tasks across MuJoCo, DeepMind Control (DMC), MyoSuite, and HumanoidBench. The study contributes to ongoing research into policy learning for complex control environments.

From Foundational Research to Long-Term AI Systems

These four studies address complementary challenges in AI development, from reliable model training and training-inference consistency to adaptive reasoning and policy optimization in complex environments.

For AI systems designed to operate over extended periods, strong reasoning at a single point in time is only part of the equation. Agents also need to retain relevant context, learn from previous interactions, and apply accumulated experience to subsequent tasks.

“From reinforcement learning to agent memory, the fundamental question has remained the same: how can an agent use experience from past interactions to make better decisions in the future?” said Jianye Hao, founder and CEO of MemoraX AI. “We believe memory should be a learnable capability and an independent infrastructure layer, rather than an extension of a foundation model. Our work focuses on training memory models to determine what to retain, when to update or retrieve information, and how to transform historical interactions into reusable experience that can guide future decisions.”

MemoraX AI will continue working with academic and industry research partners to advance foundational AI research and develop memory capabilities that help intelligent systems retain context, accumulate experience, and use that experience in future tasks.

About MemoraX AI

MemoraX AI builds Next -generation AI memory infrastructure that helps AI agents retain relevant context, learn from interactions, and apply accumulated experience to future tasks and decisions.

Its research spans AI memory, reinforcement learning, and agent systems. Through foundational research, evaluation, and product development, the company aims to make persistent memory a core capability of intelligent systems, supporting agents that can maintain context, reuse experience, and work more effectively across extended tasks.

Learn more: www.memoraxai.com

Media Contact: rosanne@memorax.net

Attachment

• MemoraX AI Researchers and Collaborators Have 10 Papers Accepted to NeurIPS 2026

CONTACT: rosanne@memorax.net