TL;DR
The Self-Evolving Curriculum (SEC) casts curriculum as a non-stationary multi-armed bandit over difficulty bins, with per-bin reward equal to mean absolute advantage, and was demonstrated on Qwen-3B/7B with 120–240 RLOO steps.
Desc descriptive evidence is provided that long-horizon multi-tool post-training can change ways of working that transfer beyond its training domain, and both software-engineering benchmarks improve despite the training collection containing no software-engineering tasks.
Large language models achieve strong reasoning performance, but often at prohibitive training cost - a challenge that is especially acute for compact models ($\leq 4 \, \mathrm{B}$ parameters) trained under limited budgets. We introduce MADA-RL, a post-training framework that specializes compact models into generator and critic roles and trains them with a debate-aware learning signal, fine-tuning only a small subset of parameters via LoRA adapters. Our central contribution is a counterfactual critic advantage: a dynamic, role-conditioned baseline that redefines the critic's advantage as its reward minus the generator ensemble's per-instance accuracy. This explicitly optimizes critics to improve over generator consensus rather than to merely reproduce a correct answer, yielding more targeted credit assignment than static mean-reward normalization. At deployment, the specialized agents are composed in a lightweight multi-round protocol. Across five mathematical reasoning benchmarks, MADA-RL raises the accuracy of the DeepSeek-R1-Distill-Qwen-1.5B model from $39.9 \, \%$ to $41.9 \, \%$ ($+2.0$ points, $p<0.001$) using $16$ times fewer trainable parameters than fully fine-tuned baselines, placing it on the accuracy-trainable-parameter Pareto front. It approaches, but does not surpass, the strongest baselines (DeepScaleR, STILL-3), which are trained on substantially larger datasets; we analyse this gap and the associated inference-time cost directly. A controlled study isolates the source of MADA-RL's gains: the counterfactual advantage produces the highest critic improvement rate of any model evaluated, indicating that trained critics learn to correct generator errors rather than to imitate them.
The alignment of Small Language Models (SLMs) in the 70--500M parameter range using reinforcement learning is often considered unstable, though the underlying failure mechanisms have not been systematically investigated. In the State-of-the-Art (SOTA) research, fifteen (model, corpus) configurations were trained using Proximal Policy Optimization (PPO). The experiments included Pythia-70M, 160M, 410M and SmolLM2-135M, 360M on the TinyStories, CNN/DailyMail, and Wikitext-103 corpora. Three reproducible failure modes were identified in small-scale language models: silent LoRA parameter freezing in standard PEFT/TRL pipelines, numerical overflow in importance ratios when using bfloat16, and catastrophic policy collapse due to reward-model error. These issues were addressed using a merge-and-reinitialize adapter technique, float32 precision during PPO updates, and a three-layer safety mechanism comprising reward whitening, importance-ratio guarding, and weight rollback. In this paper, a capacity-headroom hypothesis is proposed, which states that PPO performance at the SLM scale depends on both a fluent supervised model ($\text{PPL}<20$) and a discriminative reward signal, rather than on the number of model parameters. The proposed system converged stably in all experiments and improved preference win rate over the SFT baseline in configurations with a fluent prior and an informative reward signal. Furthermore, it outperformed instruction-tuned baselines while requiring significantly less training data. All checkpoints, preference datasets, and training scripts are publicly released$^{\S}$.
Reinforcement learning with verifiable rewards (RLVR) is a powerful recipe for improving language-model reasoning, but it is expensive to repeat on every new strong model because the target model must generate many rollouts during training. As models scale, post-training itself becomes a bottleneck. We study a weak-to-strong alternative: run RL on a smaller model where rollouts are cheaper, then reuse what that RL run learned to improve a stronger target model. Directly distilling the post-RL weak teacher is not enough, because the teacher's final policy mixes useful RL gains with the limitations of the smaller model. We propose Direct On-Policy Distillation (Direct-OPD), which transfers the teacher's RL-induced policy shift instead. Direct-OPD compares the post-RL teacher with its own pre-RL reference and treats their log-ratio as a dense implicit reward for the student. In plain terms, the checkpoint pair tells us which actions RL made the weak model more or less likely to take, and Direct-OPD applies that signal on the stronger student's own on-policy states. This directly reuses the weak model's RL supervision signal without running sparse-reward RL on the target model. Empirically, Direct-OPD consistently leverages weaker teachers to improve stronger target models; notably, it boosts Qwen3-1.7B from 48.3% to 58.3% on AIME 2024 in just 4 hours on 8 A100 GPUs. It outperforms step-matched direct RL and enables the sequential composition of multiple policy shifts. Our results show that RL outcomes can be reused across model scales as implicit reward signals, not merely as final models to imitate.
The integration of Large Language Models (LLMs) with Reinforcement Learning (RL) for financial decision-making has grown rapidly in recent years, yet the literature remains fragmented and lacks systematic comparison across methods. In this survey we analyze 34 core studies (2023–2026), selected through a multi-stage process involving 84 initial candidates and 46 full-text reviews, and propose a three-paradigm taxonomy (feature-based, auxiliary-based, and policy-based) based on the functional role of LLMs within the RL pipeline. Analysis of these integration paradigms reveals an emergent architectural trade-off: while tighter policy-based coupling theoretically offers deeper contextual reasoning, it frequently introduces significant computational overhead and training instability. Conversely, simpler feature-based integration provides superior scalability and stability, though often at the expense of representational depth. Given the current benchmark fragmentation, the reported performance gains across these studies remain difficult to validate universally across different asset classes. Critical gaps identified include the insufficient handling of data leakage and look-ahead bias, standardized benchmarks, and limited alignment with regulatory frameworks such as MiFID II and the EU AI Act.
In the quantitative finance area, particularly in order execution, reinforcement learning (RL) has shown great promise due to its ability to interact with market environments based on real data. However, traditional RL methods suffer from slow research speed and rely on static market assumptions, which do not consider the impact of the agent's execution action on the environment. To address these, we propose a Self-Evolutional single-agent/multi-agent Reinforcement Learning (SE-RL) framework. The framework utilizes a Large Language Model (LLM) to design various RL algorithm modules, such as agent model design, reward function, profiling, communication, and state imagination, by leveraging the LLM generating module output or code. SE-RL could continuously improve the accuracy of LLM-generated RL algorithms through a dual-enhancement kit at both high-level (prompt refinement) and low-level (parameter fine-tuning). Additionally, we use a multi-agent system to simulate dynamic financial markets, accounting for the impact of order executions on market dynamics. To further enhance training in such a dynamic market, we develop a hybrid environment training method that could rebalance each environment's loss weight. Comprehensive experiments on 200 realistic stock datasets demonstrate that our proposed framework outperforms current state-of-the-art baselines. Project page: https://kdd2026-se-rl.github.io/.