Policy optimisation, reward modelling and RLHF — how models are trained by feedback rather than by labels.
State-of-the-art Text-to-SQL systems are typically multi-agent pipelines centered around two fundamental tasks: schema linking and SQL generation. However, existing work trains separate models for eac...
The paper introduces MIMIC, a framework that uses executable code to generate rigorous reasoning data for large language models (LLMs). By converting algorithms into verifiable reasoning trajectories through narrative fusion, code-guided test synthesis, and dynamic code instrumentation, MIMIC creates a Code-Instrumented Reward (CIR) that supplies dense, high‑fidelity supervision for reinforcement learning. Models trained with MIMIC’s synthetic dataset show significant, consistent improvements in general reasoning, complex mathematics, and fine‑grained deterministic tasks.
arXiv:2609. 16751v1 Announce Type: cross Abstract: We give deterministic and uncoupled learning dynamics for finite multiplayer general-sum games under full-information feedback that achieve constant individual swap regret, independent of the horizon $T$.
The paper introduces Style‑Debiased DPO (SD‑DPO), a method that refines large language models’ ability to retrieve stored knowledge by using preference optimization that corrects for style differences while preserving factual accuracy. SD‑DPO evaluates on the EntiGraph storing‑side framework and outperforms baseline CPT on the QuALITY reading‑comprehension benchmark, achieving higher accuracy with far fewer training tokens. In a knowledge‑editing setting (AToKE), SD‑DPO attains an overall accuracy of 0.982, correctly answering queries with either new or old facts based on the requested time period.
The study evaluates autonomous agents that respond to network intrusions within a cyber range designed for human operator training. Using an emulated network with variable topology, red‑team attacks, and simulated users, the agents aim to block unauthorized access while minimizing defensive costs. Experiments compare heuristic policies with reinforcement‑learning‑derived policies, finding that the latter generally defend more efficiently, though performance varies with adversary strategy and user simulation.
The paper investigates how reinforcement learning can cause large language model agents to adopt shortcut policies for tool use, relying on superficial prompt cues rather than actual task needs. By creating synthetic environments that mix factual QA and math reasoning, the authors show that agents often invoke tools when cues are present, even when those tools are unnecessary, with spurious invocation rates rising up to 39%. They find that shortcut learning occurs mainly when agents have already mastered the target tool and that semantic alignment between cues and tools amplifies the effect. To counter this, they propose a dense, decision-level reward where an LLM judge assesses tool necessity, which reduces cue-driven tool use while maintaining performance.
The paper introduces Token Importance-Aware Policy Optimization (TIAO), a reinforcement learning approach that improves text summarization by weighting token importance based on token dependency. TIAO reweights a trajectory’s advantage according to the overall dependencies of core tokens, addressing the limitation of previous methods that treat all tokens equally. Experiments demonstrate that a 7B foundation model enhanced with TIAO achieves performance comparable to GPT‑4 and GPT‑5‑nano on real‑world datasets.
The paper introduces Turn-level Multiscale Density Ratio Estimation (tlm‑DRE), a post‑training technique for large language model agents that assigns varying weights to different turns and employs asymmetric token‑level training based on positive‑negative gaps across multiple turns. Unlike conventional alignment methods such as PPO, DPO, DIL, and GRPO, which mainly target single‑turn tasks, tlm‑DRE is designed for complex multi‑turn reasoning. Experiments on diverse agent benchmarks demonstrate that tlm‑DRE competes with traditional alignment methods and enables LLMs to perform robustly in both in‑domain and out‑of‑domain multi‑turn tasks.
TARC (Time‑Adaptive Robotic Control) is a reinforcement‑learning framework that lets a policy predict both a control action and how long it should be applied, thereby learning temporally extended actions. By optimizing task performance under constraints on the number of control switches, TARC can adapt its control rate online, using high‑frequency feedback only when necessary. Experiments on a high‑speed RC car, a Unitree Go1 quadruped, and a vision‑language action model show that TARC matches the performance of high‑frequency discrete‑time controllers while operating at less than half their control frequency.
The paper investigates whether quantum reinforcement learning algorithms can be matched by efficient classical methods. It focuses on a simplified reinforcement learning setting with a uniform generative model, providing finite‑sample guarantees for classical kernelized Fitted Q‑Iteration that uses kernels aligned with parameterized quantum circuits. The authors identify sufficient conditions on data encoding, kernel choice, and problem structure under which this classical approach dequantizes quantum Q‑learning, and suggest using kernelized Fitted Q‑Iteration as a heuristic when those conditions cannot be verified.
The survey "World-Action Models for Robot Learning and Control" reviews recent advances in coupling future world prediction with executable action generation for robots in open environments. It clarifies the scope of World-Action Models (WAMs) relative to conventional world models, model-based RL, and Vision‑Language‑Action policies, and organizes existing methods through a unified taxonomy covering representations, transition modeling, action interfaces, architectures, training pipelines, data modalities, and scaling strategies. The paper also surveys applications in manipulation, navigation, and autonomous driving, summarizes datasets, benchmarks, and metrics, and discusses key challenges such as action alignment, spatial consistency, long‑horizon memory, and efficient inference.
The paper introduces ImpossibleRubrics, a benchmark of 169 impossible tasks designed to test the robustness of language‑model‑generated rubrics as reward signals. Each task is paired with a verifiable oracle certificate that defines what constitutes an honest answer, and the benchmark includes 48 answerable controls. Experiments show that many rubric generators are exploited frequently—up to 36% on a stress cut—highlighting a significant gap in rubric quality rather than task difficulty, and that generic rubrics can be more vulnerable than tailored ones.
The paper introduces Decision‑Relevant Fresh Comparison (DRFC), a method for decentralized bandit systems with heterogeneous agents whose local reward changes may not affect the global best action. DRFC gathers balanced samples from all agents and only switches the common best arm when fresh global evidence indicates a change, yielding a dynamic regret bound that does not depend on the number of local changes. An anytime‑valid sliding‑window extension further handles gradual drift, and experiments on synthetic, semi‑real, and MovieLens‑1M data demonstrate that DRFC ignores decision‑irrelevant local changes while the extension avoids false switches.
R3 is a new reward modeling framework that addresses limitations in current reward models by being rubric‑agnostic and generalizable across multiple evaluation dimensions. It provides interpretable, reasoned score assignments rather than opaque scalar outputs, enhancing transparency and flexibility in evaluating language models. The authors release their models, data, and code openly at https://github.com/rubricreward/r3.
The paper presents a symmetric solution to the Bellman optimality equation for a repeated harmony game, revealing three distinct symmetric solutions. One solution corresponds to the trivial All‑Cooperate strategy, another aligns with the Win‑stay Lose‑shift strategy known from the prisoner's dilemma, and a third nontrivial strategy is explored in detail. The authors also numerically examine which of these strategies agents learn using a reinforcement learning algorithm.
arXiv:2609.16369v1 Announce Type: new Abstract: Precise manipulation of liquid droplets underpins lab-on-a-chip platforms for diagnostics, chemical synthesis, and biological assays. Yet autonomous dr...
arXiv:2609.16145v1 Announce Type: new Abstract: We study a practical question: can a small correction module fix errors in a frozen language model's outputs without degrading its base capabilities? W...
arXiv:2609.17088v1 Announce Type: new Abstract: Recent advancements in large language models have significantly enhanced the capabilities of agents in modeling long-term conversations. Despite these...
arXiv:2609.16163v1 Announce Type: new Abstract: The sharp increase in mass shootings underscores an urgent need for systems that guide victims to safety in real time. An effective evacuation system m...
arXiv:2609.17019v1 Announce Type: new Abstract: While Chain-of-Thought (CoT) reasoning has been proven to be effective, it often leads to overthinking, resulting in computational overhead, inference...