Policy optimisation, reward modelling and RLHF — how models are trained by feedback rather than by labels.
arXiv:2609.27572v1 Announce Type: cross Abstract: Reinforcement learning (RL) has emerged as a key paradigm for improving the reasoning capabilities of large language models (LLMs). However, existing...
arXiv:2609.27421v1 Announce Type: new Abstract: Multi-constraint instruction following requires a model to respond to a query under many simultaneously active constraints. Even strong instruction-tun...
arXiv:2609.27532v1 Announce Type: new Abstract: Long-horizon agentic tasks require an agent to modify an environment through a sequence of tool calls, with success determined by the final state. The...
The paper investigates what a steerable valence direction in large language models (LLMs) actually represents, focusing on a good‑bad outcome direction in a maze task. By using controlled interventions that separate the realized outcome from the informational history that led to it, the authors find that directions trained on one explicit outcome encoding transfer well to another, suggesting the readout is not tied to surface form. However, when the same outcome is achieved through announced versus unannounced histories, transfer performance drops sharply, indicating that the post‑event readout remains strongly conditioned on the earlier announcement. In a matched maze‑reinforcement‑learning run, the post‑RL direction becomes more predictive of reference‑MDP return and the policy depends more on it, yet the history dependence persists. These findings support a functional, value‑related interpretation of the direction but argue against identifying it with a history‑invariant scalar valence state.
Group-based reinforcement learning (RL) methods, such as GRPO and its variants, have become a leading paradigm for training reasoning and agentic large language models (LLMs). While their group-normal...
In value-based reinforcement learning, improving the accuracy of policy evaluation has been shown to improve downstream policy optimization performance. The widely adopted family of approximations rel...
Reinforcement learning (RL) is widely used to sharpen reasoning in multimodal large language models (MLLMs), yet its effect on hallucination is uneven. We trace this to two weak points in the \emph{co...
Cooperation emergence is a central problem in multi-agent systems because decentralized agents must coordinate while adapting to the changing behavior of others. Evolutionary game theory identifies st...
Multi-constraint instruction following requires a model to respond to a query under many simultaneously active constraints. Even strong instruction-tuned models still routinely violate some of them. E...
arXiv:2609.26780v1 Announce Type: cross Abstract: Long-term conversational memory in multi-party settings requires more than retrieving relevant content from long-term conversations: it must distingu...
ForeDrive introduces a planning-relevant latent world model that is asymmetrically coupled to a Diffusion Transformer planner. The model learns multi‑horizon latent futures with a JEPA‑style world model, while planning gradients update the shared encoder and stop‑gradient routing trains the predictor with forecasting losses only. Gated visual fusion, future‑status injection, and Trajectory‑Adaptive Bias are used to guide trajectory generation without overriding current observations, achieving high performance on NAVSIM benchmarks using only front‑view images and pure imitation learning.
ToolCUA is an end‑to‑end agent that learns to optimally orchestrate GUI actions and tool calls for Computer Use Agents. It introduces a staged training pipeline that first scales interleaved GUI‑tool trajectories from static GUI data, then bootstraps decision making with single‑turn reinforcement learning, and finally refines performance via online agentic RL guided by a tool‑efficient path reward. On the OSWorld‑MCP benchmark, ToolCUA achieves 46.85% accuracy, a 66% relative improvement over the baseline and a 3.9% gain over GUI‑only models, setting a new state of the art for comparable‑scale models.
The paper introduces a novel end‑to‑end, size‑agnostic graph reinforcement learning framework for the one‑dimensional bin packing problem (1D‑BPP). It models packing as a Markov decision process on an item‑compatibility graph, where a graph neural network actor‑critic policy learns to merge compatible partial bins. Empirical results on the BPPLIB benchmark show that the learned policy reduces the mean optimality gap of a constructive heuristic from 2.66 % to 2.31 %, performs competitively against other learned methods, and outperforms a state‑of‑the‑art learned solver on the hardest benchmark family.
The paper introduces a black-box defense strategy for smart meter data that uses a proxy-guided hierarchical reinforcement learning framework to generate battery-based load-shaping policies. These policies inject realistic yet misleading appliance-level signatures into aggregate power signals, disrupting non-intrusive load monitoring attacks. Experiments on UK-DALE and REDD datasets show significant increases in appliance-level reconstruction error and reductions in attacker F1 scores across multiple unseen NILM models.
MAGIC introduces a dense‑reward reinforcement learning framework for generating mixed‑granularity agent graphs in large‑language‑model based multi‑agent systems. The method sequentially selects functional roles, instantiates them as either single agents or reusable groups, and connects them to existing units, optimizing the construction policy with intermediate feedback from probe‑based utility and structural signals. Experiments show MAGIC outperforms state‑of‑the‑art baselines on eight benchmarks and achieves strong inference efficiency.
The paper investigates how entropy over chain‑of‑thought tokens influences policy decisions such as gradient application, pruning, and collapse detection. By separating scaffold tokens from substantive content, the authors analyze entropy, Kullback–Leibler divergence, and entropy velocity for each channel, proving differences between raw and content conventions and bounding answer diversity. Empirical results across 23 configurations show that scaffold tokens can account for up to 41% of high‑entropy positions, with entropy share growing through distillation, while content conventions outperform raw surprisal on compression tasks and reveal significant answer leakage in re‑fed chains.
Matryoshka Attribution (MAttr) is a mask‑learning method that identifies nested subsets of a language model’s internal components by minimizing downstream loss. It uses a differentiable sigmoid top‑k operator and randomizes sparsity during training to produce an attribution ordering of components. MAttr tops the Mechanistic Interpretability Benchmark leaderboard and can be applied via reinforcement learning to pinpoint weight changes that control behaviors such as refusal in Llama 3.1 8B Instruct, where restoring just 1% of weights removes refusals while preserving capabilities.
Ladders-of-Thought (LoT) is a framework that enhances reasoning in small- to mid-scale large language models by automatically generating easier variants of reasoning problems and organizing them into difficulty buckets. It uses a self‑evolving bandit scheduler to adaptively allocate training, improving performance across math and multi‑hop reasoning tasks on 1–8 B models. LoT achieves significant gains (e.g., +32 pp on AddSub, +16 pp on QASC) and converges faster than staged curricula.
The paper introduces Network Feasibility Geometry Reinforcement Learning (NFG‑RL), a method that enforces multi‑layer network constraints—such as interference, power‑rate coupling, flow conservation, service chains, capacity, latency, and reliability—by transporting a proto‑policy through a differentiable feasibility map. By compiling heterogeneous constraints into typed residual blocks and using a variational transport operator, NFG‑RL ensures almost‑sure feasible execution and shapes exploration and gradients to respect active constraints. Experiments on two wireless‑edge surrogate environments show that NFG‑RL boosts feasible utility by 37.5–41.5 %, cuts raw‑action violations by 48.5–60.8 %, and reduces P99 delay by 57.0–75.5 % compared to leading baselines.
The paper introduces MGRL-RSCC, a multi‑granularity reward reinforcement learning framework for Remote Sensing Change Captioning (RSCC). It uses a CNN with hierarchical self‑attention to extract visual features, a Transformer decoder for visual‑to‑linguistic translation, and a dual‑decoding strategy combined with token‑level supervised learning and self‑critical reinforcement learning. Three reward functions—linguistic fluency, change state consistency, and structural‑semantic relevance—are employed to improve caption quality and reduce exposure bias and conservative generation.