Search: “on-policy-learning”

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

MInTRL: Off-policy Intervention can boost On-policy RL

MInTRL injects sparse judge corrections into on-policy RL rollouts, expanding exploration beyond on-policy sampling while preserving learnability on math and code benchmarks.

Minimal Intervention Reinforcement Learning periodically has a judge-intervention policy replace erroneous suffixes of the current policy's output with short corrections, then returns control, keeping trajectories largely on-policy. Training uses a sequence-level advantage-regression objective that removes the need for importance sampling. Across math and code benchmarks it consistently beats standard on-policy and off-policy baselines, remains effective with self-intervention, and performs best at moderate intervention intensity.

Hugging Face daily papers·Sep 10, 20:00 UTC · 6d agoAI research

40

Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation

OPRD distillation enables weak-to-strong generalization by amplifying verifier-supported policy updates, outperforming existing RL and distillation methods with fewer student updates.

On-Policy Reverse Distillation (OPRD) evaluates a weak teacher's policy shift relative to its reference policy on student rollouts and amplifies the verifier-supported component of the student's policy gradient. This rescaling preserves the stationary points of policy optimization while letting the student learn beyond the teacher's capacity ceiling. In successive model transfer and multi-teacher distillation, OPRD achieves higher performance with fewer student updates than existing RL and distillation approaches, and response-style analysis shows students remain closer to verifier-RL-trained models than to their weak teachers.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research

33

One Symptom, Three Levers: A Critical Review of On-Policy Self-Distillation

A review paper frames on-policy self-distillation collapse as governed by three levers: token weighting, privileged information, and guidance decay.

The paper critically reviews On-Policy Self-Distillation (OPSD), where a language model trains on its own generations scored token-by-token by a teacher conditioned on privileged information such as reference solutions or environment feedback. It identifies collapse, the progressive narrowing of producible reasoning paths, as the dominant failure mode and analyzes it through three levers: signal weighting, the nature of privileged information, and teacher dynamics. The review is restricted to mathematical reasoning, reports no new experiments, and offers a shared vocabulary separating settled findings from disputed ones.

Hugging Face daily papers·Aug 25, 20:00 UTC · 22d agoAI research

45

Learning Multimodal One-step Flow Policy via Value-weighted Optimal Transport

OptiFlow learns one-step multimodal flow policies for offline RL via state-wise entropic optimal transport, avoiding critic overestimation and mode collapse.

The paper introduces OptiFlow, a framework that frames one-step flow policy learning as a structured sample-allocation problem in offline reinforcement learning. It jointly trains a value-aware reference flow policy and a one-step policy, coupling action samples through state-wise entropic optimal transport where critic values set distillation priority and action-distance cost ensures geometrically compatible pairings. By avoiding direct critic maximization, it anchors the policy to high-value dataset-supported modes without out-of-distribution divergence. Code is released on GitHub and the method performs strongly across diverse offline RL benchmarks.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:10 UTC · 2d agoAI research

14

Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

Research shows on-policy expert correction, not imitation fine-tuning, lets weaker agent models catch up under evolved harnesses.

Researchers study how to combine automated agent-harness evolution with lightweight fine-tuning across seven enterprise agent tasks. Naively training weaker models (Qwen3-Coder, Gemma 4) on expert trajectories under an evolved harness regressed performance by 4 to 30 points on all tasks. They propose an on-policy correction pipeline, automated by a meta-level MLE agent, where an expert rewrites only the failing turn of the weaker model's rollout, preserving model-harness fit.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:53 UTC · 8d agoAI research

35

Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement

Generalized Agent Iteration formally unifies iterative policy improvement and recursive self-improvement, defining axes that distinguish anchored, goal-drifting, and self-referential agents.

The paper proposes Generalized Agent Iteration (GAI), a formal framework that models learning as a cycle of agent evaluation and agent improvement, defining the agent as a configuration of modifiable components. Two dials—whether the improving mechanism is part of the agent and whether the evaluation standard is grounded outside it—separate generalized policy iteration (GPI) from recursive self-improvement (RSI) and classify systems as anchored, goal drift, or fully self-referential. The framework places existing systems on shared axes and makes defects of recursive self-improvement statable one condition at a time.

Hugging Face daily papers·Sep 10, 20:00 UTC · 6d agoAI research

30

Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

Research shows imitation of expert trajectories breaks weaker models' harness fit, while on-policy expert correction preserves gains across seven enterprise agent tasks.

The paper studies combining automated agent-harness evolution with lightweight fine-tuning across seven enterprise agent tasks using Qwen3-Coder and Gemma 4. Training weaker models on complete expert trajectories under an evolved harness regressed performance by 4-30 points on all tasks, disrupting model-harness fit. The authors propose an on-policy expert-correction pipeline, automated by a meta-level MLE agent, that rewrites only failing turns and preserves the model's planning style.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research

48

Entropy-Regularized Rank-Masked Policy Optimization for Test-Time Reinforcement Learning in Code Generation

Researchers propose ERPO, enabling test-time reinforcement learning for code generation via probe-executed consensus rewards, rank masking, and entropy regularization.

The paper introduces probe-driven test-time reinforcement learning (TTRL) for code generation, where output-free probe inputs are constructed from problem statements and candidate programs are executed on them to compute a Probe Consensus Reward (PCR). Because PCR can be gamed through spurious consensus, the authors propose Entropy-Regularized Rank-Masked Policy Optimization (ERPO), which turns low-PCR outcomes into conservative negative updates via rank masking and constrains policy drift with an entropy ceiling. On coding benchmarks, ERPO substantially improves pass@1 and pass@k in both in-domain adaptation and zero-shot transfer.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:54 UTC · 8d agoAI research1

38

Safe Meta-Reinforcement Learning via Information Space Reachability

Safe meta-RL framework reasons about safety in information space, learning a safety value function used for safety filtering and constrained policy optimization.

The paper proposes safe meta-RL that reasons about safety in information space, capturing both physical state and the agent's belief over the underlying task. A safety value function measures the probability of avoiding unsafe regions indefinitely and satisfies a self-consistency condition and Bellman equation, making it learnable via meta-RL. The resulting algorithm uses the learned function for safety filtering and constrained policy optimization, with effectiveness demonstrated on meta-RL benchmarks.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:27 UTC · 2d agoAI research

25

Multi-Agent Reinforcement Learning for Autonomous UAV Exploration in Wildfire Response

Researchers trained multi-agent deep reinforcement learning UAV agents for autonomous wildfire monitoring, with converging policies tracking fire boundaries in simulation.

The study develops a deep reinforcement learning framework for training UAV agents to navigate and monitor simulated wildfire environments. Agents showed increasingly stable and effective behavior over time, evidenced by converging loss trends, improved rewards, and consistent navigation patterns such as fire-boundary tracking. The findings highlight DRL-based UAV potential for autonomous wildfire monitoring and show that environmental structure and reward design influence policy effectiveness.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:45 UTC · 7d agoAI research1

18

Revisiting Complete Reasoning Traces for Post-Training

Researchers show full reasoning traces provide limited benefit in LLM post-training, with heavily truncated or endpoint-only trajectories performing comparably.

A pilot study plus attention-based analyses and controlled token-removal studies show intermediate tokens in reasoning trajectories contribute minimally to final reasoning quality. Partial trajectories remain effective even under heavy truncation, and training on endpoints alone leads to consistent changes in reasoning behavior. The finding also benefits reinforcement-learning and on-policy distillation post-training; code is released at github.com/naver-ai/revisiting-trace.

Hugging Face daily papers·Sep 6, 20:00 UTC · 10d agoAI research

32

Negative Self-Distillation: Learning to Reason by Avoiding Flaws

Researchers propose Negative Self-Distillation (NSD), a label-free LLM self-improvement method that diverges from self-generated flawed reasoning rather than imitating privileged solutions.

The authors show On-Policy Self-Distillation can degrade complex reasoning by forcing imitation of artificially confident traces built on privileged information, suppressing uncertainty and self-correction. NSD instead generates a question-specific negative condition — such as acting as a 'careless reasoner' — and pushes the model's distribution away from it without ground-truth labels. A dynamic gating mechanism isolates reasoning-critical tokens so gradient updates fix behavioral flaws without damaging foundational linguistic capabilities. NSD consistently outperforms OPSD and other label-free, self-bootstrapping reinforcement learning baselines.

Hugging Face daily papers·Sep 9, 20:00 UTC · 7d agoAI research1

30

Not All Prompts Are Equal: Exploration-Guided Prompt Scaffolding for Multimodal Reinforcement Post-Training

Exploration-guided prompt scaffolding rewrites training prompts by Exploration Potential Score, boosting multimodal RL post-training accuracy up to 11.5%.

The paper proposes dynamically adapting the training prompt distribution during online RL post-training of multimodal LLMs using the Exploration Potential Score (EPS), a lightweight rollout-based proxy for prompt utility computed from on-policy statistics with no additional overhead. Rather than discarding low-utility prompts, a teacher model generates scaffolded rewrites that preserve task intent while making training more informative. Integrated with GRPO on Geo3K and MMK12, the method achieves up to 9.7% relative in-domain improvement plus 11.5% on MathVision and 11.1% on MMMU-Pro.

Hugging Face daily papers·Sep 13, 20:00 UTC · 3d agoAI research

35

Verify Before You Distill: Prompt-Level Teacher Gating for On-Policy Distillation

TGOPD verifies teacher reliability per prompt before on-policy distillation, outperforming vanilla OPD across math, code, and instruction benchmarks.

Teacher-Gated On-Policy Distillation (TGOPD) estimates teacher reliability from verifier-scored teacher probes and routes each prompt either to dense on-policy distillation or to verifier-grounded GRPO, avoiding misleading updates from confidently wrong teachers under mode-seeking reverse KL. Across 4B and 35B students in mathematics, code, and instruction following, TGOPD outperforms vanilla OPD in all six single-domain settings and achieves higher seven-benchmark averages under multi-domain training. It also raises teacher-node GPU utilization from 9.8% to 78.9% in the measured 4B single-domain run by reusing idle teacher capacity.

Hugging Face daily papers·Sep 1, 20:00 UTC · 15d agoAI research

26

Environments as Scaffold: Enriching Feedback to Bootstrap Self-Evolving Agents in Long-Horizon Tasks

Researchers propose Feedback-Enriched Environments (FEEs) that reduce reward sparsity and improve RL training of Qwen3-based agents on SciWorld and BFCL.

The paper proposes shifting from agent-side warmup (SFT) to environment-side adaptation via Feedback-Enriched Environments to address severe reward sparsity in RL training of long-horizon LLM agents. A pilot study defines a feedback strategy that transitions from action guidance to observation enrichment in later training stages. Large-scale experiments on SciWorld and BFCL across Qwen3 model scales and GRPO, GSPO, and DAPO show consistent gains, plus stabilized training dynamics and proactive exploration.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research

25

Learning to Coach for Experiential Learning

Learning to Coach trains a dedicated LLM coach to extract transferable experiential knowledge from a frozen actor's trajectories, beating self-refinement.

Learning to Coach (L2C) trains an LLM-as-a-Coach to extract actionable experiential knowledge from a frozen actor model's previous solution trajectories, optimizing rewards based on the actor's guided response correctness. It studies same-instance and cross-instance rewards, where cross-instance elicits knowledge that transfers to other problems. Across mathematical reasoning and interactive text-games, L2C outperforms self-refinement and untrained coaches, scales better with extra inference iterations than larger decoding budgets, and transfers to out-of-distribution tasks.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 16:47 UTC · 2d agoAI research1

20

Bellman Policy Optimization

Bellman Policy Optimization, a critic-free RLVR method derived from Policy Mirror Descent, improves LLM mathematical reasoning without intermediate state-value estimation.

The paper introduces Bellman Policy Optimization (BPO), a critic-free reinforcement learning method for LLMs with verifiable rewards, derived from Policy Mirror Descent. BPO uses the Bellman equations to reformulate PMD as a trajectory-level objective for autoregressive generation with terminal rewards, avoiding state-value estimation at intermediate states. The authors prove BPO shares the same unique optimal solution as PMD and validate it on mathematical reasoning benchmarks.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:59 UTC · 2d agoAI research

38

ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents

ScienceBuddy couples harness evolution with model reinforcement learning so scientific agents continually self-improve from researcher feedback in an interactive workspace.

The authors release ScienceBuddy, an interactive scientific research workspace that transforms researcher requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. Its recursive-in-recursive self-improvement paradigm couples harness evolution (inner recursion, model fixed) with model reinforcement learning under the improved harness (outer recursion). Case studies span four scientific task families, and the system is released to the scientific community as a research product.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:55 UTC · 1d agoAI research 2 sources

35

FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience

FlowBalance is a verifier-grounded self-improvement method that beats FlowRL on Qwen3-4B and Qwen3-8B math reasoning while improving training stability.

FlowBalance calibrates dense self-guidance scores with verifier-derived group advantages: guidance is retained on positive-advantage trajectories, reversed on negative-advantage trajectories, and disabled when rollout groups show no outcome preference. The method exponentially reweights a reference policy via trajectory balance, with guarantees including within-group contrast preservation and a minimum-change reverse-KL characterization. On mathematical reasoning it outperforms FlowRL on Qwen3-4B and Qwen3-8B, trains faster and more stably, avoids direct OPSD's response-length collapse, and shows higher correct-strategy diversity on AIME24.

Hugging Face daily papers·Sep 2, 20:00 UTC · 14d agoAI research

32

5 useful things you'll learn in my new post-training textbook (shipping now!)

Nathan Lambert's new RLHF and post-training LLM textbook covers PPO, GRPO, GSPO, CISPO and related techniques, freely available online.

Nathan Lambert's book 'Reinforcement Learning from Human Feedback: Aligning and Post-training LLMs' is now shipping from Manning. It covers policy-gradient algorithms including PPO, GRPO, GSPO, CISPO, and RLOO, plus loss aggregation, truncated importance sampling, asynchronous RL systems, and post-training topics like rejection sampling, outcome reward models, and on-policy distillation. The book is freely available online with a 12-hour course, codebase, and exercises.

Interconnects·Aug 10, 13:02 UTC · Aug 10, 2026AI research

28

OmniHarness: Harnessing Generalizable Visual Generation via Symbolic Policy Learning

OmniHarness learns symbolic policies for visual generation agents, reaching a 95.0% resolve rate on ComfyBench Creative tasks, 27.5 points above the strongest baseline.

OmniHarness abstracts verified executions into symbolic policies for visual generation task families, which are instantiated, adapted, and composed for new tasks while model parameters remain fixed. Intermediate verification guides refinement and failure recovery during execution, and self-directed inquiry generates practice tasks near capability limits before downstream objectives are specified. Experiments across six benchmarks, three MLLM backbones, and three visual agent frameworks show strong performance; on ComfyBench Creative tasks it achieves a 95.0% resolve rate, exceeding the strongest baseline by 27.5 percentage points. Frozen policy snapshots improve existing visual agent systems through plug-and-play reuse.

Hugging Face daily papers·Sep 12, 20:00 UTC · 4d agoAI research

33

Difficulty-Adaptive Tree-Structured Policy Optimization for Expanding Reasoning Coverage in RLVR

Researchers propose DATPO, a difficulty-adaptive tree-structured RLVR training method that expands reasoning coverage (pass@k) and improves test-time scaling on math benchmarks.

The paper identifies three rollout design principles for RLVR: difficulty-adaptive rollout expands pass@k, tree-based rollout beats parallel sampling, and sentence-entropy-guided forking overcomes token-level branching localization. DATPO combines difficulty-adaptive tree search with a sibling-diversity advantage term to promote semantic diversity during training. On mathematical reasoning benchmarks, DATPO outperforms baselines in pass@k, directly translating to superior test-time scaling performance.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research

30

PLC-DPO: Posterior Label Correction in Noisy and Ambiguous Preference Optimization

PLC-DPO corrects noisy preference labels during DPO by routing each pair as clean, flip, or tie using calibrated policy-reference margins.

PLC-DPO addresses reversed, weak, or ambiguous preference labels in Direct Preference Optimization by using the calibrated policy-reference margin as online evidence to route each pair's training signal as clean, flip, or tie. This actively corrects supervision direction and strength rather than only filtering suspicious examples. Across 57 dataset-model-benchmark cells, PLC-DPO achieves the best mean win rate of 60.5 versus 55.5 for the next-best method. Injected-noise and tie stress tests plus human disagreement analysis show the routing stays stable and distinguishes flipped from weakly directional pairs.

Hugging Face daily papers·Aug 30, 20:00 UTC · 17d agoAI research

28

Expert-Space Exploration in MoE Reinforcement Learning

ESRL explores MoE expert-routing space during RL post-training, improving Qwen3-30B-A3B Pass@1 by 3.2 points over GRPO without extra compute.

The paper shows perturbing expert routing increases rollout diversity similarly to higher decoding temperature, but naive perturbation degrades quality. ESRL anchors high-confidence experts, restricts stochastic routing to a plausible candidate pool, adapts perturbation strength via router entropy, and replays recorded expert paths during policy optimization. It achieves the best results across top-K, top-1, and shared-expert MoE backbones on math, science, and code tasks; on Qwen3-30B-A3B it improves average Pass@1 and Pass@8 over GRPO by 3.2 and 4.5 percentage points.

Hugging Face daily papers·updated Sep 11, 16:58 UTC · 5d ago·first Sep 10, 20:00 UTC · 6d agoAI research 2 sources

45

Near-Optimal Reinforcement Learning with Multi-Step Transition Lookahead

Theorists prove multi-step lookahead RL planning is NP-hard for every fixed rational discount factor yet give a randomized polynomial-time approximation scheme.

The paper resolves open questions about reinforcement learning with multi-step transition lookahead. It shows exact planning remains NP-hard for every fixed rational discount factor in (0,1), not just discounts arbitrarily close to one, and introduces a randomized polynomial-time approximation scheme for every fixed lookahead depth. Extending to unknown transitions and stochastic rewards via optimism and variance-adaptive confidence bounds, the algorithm achieves cumulative regret matching classical tabular discounted RL up to logarithmic factors.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 16:58 UTC · 6d agoAI research1

22

Bridging Control, Inference, Transport, and Thermodynamics: From Theory to Applications in Learning

Review connects control theory, optimal transport, probabilistic inference, thermodynamics, and machine learning via free-energy optimization under constraints.

The review unifies five fields: control theory, optimal transport, probabilistic inference, non-equilibrium thermodynamics, and machine learning. The common conceptual thread is optimization of free-energy-like functionals under dynamical or statistical constraints. Selected applications are presented in reinforcement learning, variational inference, and generative modeling. The tutorial-style text assumes no prior familiarity and begins from physics principles.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:16 UTC · 2d agoAI research

15

Coupled Calibration and Learning: Mitigating Teacher Bias in LLM Distillation without Target-Domain Reward Feedback

CCL couples teacher calibration with student updates via token-level branching, provably removing teacher bias in LLM distillation.

The paper proposes Coupled Calibration and Learning (CCL), an LLM distillation algorithm that alternates teacher calibration using source-question reward feedback with student training on target questions under covariate shift. Each iteration calibrates the teacher on source feedback, trains the student on target questions, and lets the updated student inform subsequent calibration. The authors prove the student's expected KL divergence to the oracle student converges to zero at a polynomial rate, and show regularized direct matching error can remain bounded away from zero.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:15 UTC · 1d agoAI research

20

Drift-Constrained Optimization: Only Direction Matters in Fine-Tuning Instruct Models

Drift-Constrained Optimization reformulates fine-tuning as update-direction selection, letting Qwen3 models improve target tasks within a behavioral drift budget.

The paper specifies a behavioral drift budget before optimization and shows that update direction is the remaining degree of freedom, reformulating fine-tuning as a direction-selection problem. In a stringent QA-only setting where instruct models must still generate multi-step reasoning at inference, a coarse layer-selective probe reverses the failure of QA-only fine-tuning. Across Qwen3-8B and Qwen3-14B, these directions substantially improve scientific reasoning and multilingual translation, matching or outperforming dedicated translation systems over 100+ languages and giving stronger initialization for reinforcement learning.

Hugging Face daily papers·Sep 11, 20:00 UTC · 5d agoAI research

48

Distill Globally, Adapt Locally: Reasoning Distillation and Product-Type Test-Time Training for Scalable Trade-Up Recommendation

A distillation framework compresses LLM reasoning into a 15.5M-parameter trade-up recommendation model reaching AUC 0.941 with product-type test-time training.

The paper targets trade-up recommendation, which identifies higher-quality alternatives that preserve customer purchase intent. A retrieval-augmented few-shot LLM teacher generates labels and rationales that supervise a compact embedding-pair classifier; at inference the 15.5M-parameter student uses only two precomputed 768-dimensional embeddings with no LLM calls. On 8,352 annotated pairs, label-only training scored AUC 0.912, reasoning distillation reached 0.924, and product-type test-time training lifted it to 0.941 with average precision 0.940. The distilled student is roughly 5,000x faster and 10,000x cheaper than direct LLM inference on a 100K-pair proxy catalog.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:08 UTC · 12d agoAI research

25

Inoculation Midtraining with Learned Neologisms

Inoculation Midtraining confines unsafe LLM behavior to a neologism-marked context, reducing misalignment after unsafe post-training but leaking under nearby contextual cues.

The paper introduces Inoculation Midtraining, which teaches a base model during midtraining that unsafe behavior belongs to a context marked by a learned neologism token, then post-trains on unsafe data within that context. Across supervised fine-tuning and RL post-training regimes, the technique reduces misalignment while preserving transfer of benign properties like German or Shakespearean prose. However, it does not outperform standard Inoculation Prompting, is sensitive to training configuration, and produces a leaky boundary that nearby contextual cues can reactivate. The authors conclude it is not yet a load-bearing component of a developer safety framework.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:12 UTC · 2d agoAI safety & security

42