Policy optimisation, reward modelling and RLHF — how models are trained by feedback rather than by labels.
arXiv:2608.21864v1 Announce Type: cross Abstract: The current progress of Clinical Vision Large Language Models (C-VLLMs) has substantially improved digital diagnostics, still these frameworks often...
The paper introduces the Shared Recurrent Memory Transformer (SRMT), a decentralized multi‑agent reinforcement learning framework that uses a global memory workspace for agents to broadcast and query each other’s learned states. SRMT is evaluated on the Partially Observable Multi‑Agent Pathfinding (PO‑MAPF) problem, showing that shared memory enables emergent coordination even with minimal reward guidance and outperforms existing baselines on the Bottleneck task and scales competitively on larger POGEMA maps. The authors provide open‑source code for training and evaluation on GitHub.
arXiv:2608.21863v1 Announce Type: cross Abstract: Tool-Integrated Reasoning (TIR) is a fundamental capability for LLM agents to solve complex tasks by interacting with external tools iteratively. Rei...
arXiv:2608.21727v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) is deployed to make models better at reasoning tasks, but its side effect on what models will d...
arXiv:2608.22549v1 Announce Type: new Abstract: Batched simulators for autonomous driving have recently enabled training reinforcement learning (RL) agents at scale, encompassing thousands of traffic...
arXiv:2608.21386v1 Announce Type: cross Abstract: Given a task described by a few examples, how should a model be specialized to it? Four mechanisms are available -- zero-shot, in-context attention,...
arXiv:2608.22429v1 Announce Type: new Abstract: Multimodal Large Language Models (MLLMs) capable of thinking with images often rely on external tools for fine-grained perception. However, this relian...
arXiv:2608.22266v1 Announce Type: new Abstract: In the context of information seeking, conversational agents are undergoing an evolution from reactive tools to proactive, personalized assistants. A c...
arXiv:2608.22595v1 Announce Type: new Abstract: We develop a new framework for flexible, nonlinear, and interpretable off-policy evaluation for infinite-horizon reinforcement learning. To handle larg...
The paper introduces Gradient Fingerprint (GRIFT), a technique that uses a model’s internal gradient computations to detect reward hacking in reinforcement learning with verifiable rewards. GRIFT compresses gradients of a chain-of-thought (CoT) conditioned on a prompt into a compact representation, which is then used to assess whether the CoT reflects reward hacking. Experiments on math, code, and logical reasoning benchmarks show GRIFT outperforms baselines by over 25% and, when integrated into a rejection fine‑tuning pipeline, reduces reward hacking while improving task performance.
The paper introduces MEMENTO, a memory‑enhanced neural solver that improves routing problem solutions by using online data from repeated attempts to adjust action distributions during inference. It targets NP‑hard routing tasks such as the Traveling Salesman and Capacitated Vehicle Routing problems, outperforming existing tree‑search and policy‑gradient fine‑tuning methods. MEMENTO demonstrates strong scalability and data efficiency, achieving state‑of‑the‑art results on 11 of 12 evaluated tasks and enabling zero‑shot integration with diversity‑based solvers.
AdaR is a framework designed to enhance large language models (LLMs) with adaptive reasoning for mathematical tasks. It identifies and mitigates spurious reasoning—where models rely on superficial correlations—by generating logically equivalent queries and training with Reinforcement Learning with Verifiable Rewards (RLVR) to penalize incorrect logic and promote adaptive logic. The approach includes extracting problem‑solving logic, executing code to verify answers, and applying sanity checks, resulting in significant gains in mathematical reasoning performance and improved data efficiency.
iScheduler is a reinforcement‑learning‑driven framework that tackles large‑scale Resource Investment Problems (RIP) by modeling them as a Markov decision process over decomposed subproblems and building schedules through sequential process selection. The approach speeds up optimization and allows efficient reconfiguration by reusing unchanged process schedules and only rescheduling affected processes. Using the new L‑RIPLIB benchmark, iScheduler achieves competitive resource costs while cutting time to feasibility by up to 43× compared to leading solver‑backed baselines.
The paper introduces Continual Uncertainty Learning (CUL), a curriculum-based continual learning framework that decomposes robust control of nonlinear systems with multiple heterogeneous uncertainties into a sequence of tasks. By progressively expanding and diversifying plant uncertainties and applying memory-efficient anti-forgetting regularization, CUL enables a policy to acquire strategies for each uncertainty sequentially while a model-based controller provides a shared baseline performance. Applied to an active vibration controller for automotive powertrains, the approach demonstrates robustness to structural nonlinearities and dynamic variations, improving control performance and sample efficiency.
The paper introduces RARRL, a hierarchical framework that learns when and how an embodied robotic agent should invoke large language model reasoning. By adaptively deciding whether to reason, selecting the reasoning role, and allocating computational budget based on observations, execution history, and remaining resources, RARRL improves task success rates and reduces execution latency. Experiments on the ALFRED benchmark demonstrate that this resource‑aware orchestration outperforms fixed or heuristic reasoning strategies, highlighting the importance of adaptive reasoning control for reliable robotic agents.
arXiv:2608.21808v1 Announce Type: new Abstract: Multimodal Retrieval-Augmented Generation (RAG) with visual citation is crucial for ensuring the traceability and verifiability of MLLMs. However, curr...
GRACE is a step‑level benchmark for evaluating the faithfulness of chain‑of‑thought reasoning over context. It provides human annotations for each step in CoT traces from 10 models across 4 datasets, labeling faithfulness, error category, and natural‑language explanations. The benchmark introduces a data‑driven taxonomy that splits errors into GRACE‑Inference (deductive) and GRACE‑Grounding (factual) tracks, each with four categories, and demonstrates that incorporating step‑level faithfulness signals can improve downstream accuracy and reasoning reliability.
MPCFormer is a physics‑informed, data‑driven framework that explicitly models multi‑vehicle social interaction dynamics for autonomous driving. It uses a Transformer‑based encoder‑decoder to learn discrete state‑space dynamics from naturalistic data, enabling explainable, human‑like behavior planning within a Model Predictive Control (MPC) framework. In open‑loop NGSIM tests, it achieves the lowest trajectory prediction errors (ADE 0.86 m over 5 s), and in closed‑loop intense interaction scenarios it attains a 94.67 % planning success rate, 15.75 % efficiency gain, and reduces collisions from 21.25 % to 0.5 %.
arXiv:2608.21925v1 Announce Type: new Abstract: Emotional Support Conversation (ESC) systems aim to provide holistic support by balancing professional therapeutic competence with natural empathy. How...
arXiv:2512.13095v3 Announce Type: replace Abstract: To address the limited capability expansion and low sample efficiency of Reinforcement Learning (RL), recent methods have integrated ''hints'' into...