4arXiv cs.CL (Computation and Language)·
MADA-RL: Multi-agent debate with counterfactual critic advantage for parameter-efficient reasoning in compact models
MADA-RL is a post-training framework that specializes compact LLMs (≤4B parameters) into generator and critic roles, training them with a debate-aware RL signal using LoRA adapters on only a small fraction of parameters. The central contribution is a counterfactual critic advantage that conditions the critic's reward on the generator ensemble's per-instance accuracy, explicitly incentivizing critics to correct generator errors rather than reproduce correct answers. Applied to DeepSeek-R1-Distill-Qwen-1.5B, the method achieves +2.0 percentage points on mathematical reasoning benchmarks using 16× fewer trainable parameters than fully fine-tuned baselines, placing it on the accuracy-vs-trainable-parameter Pareto front. The approach does not surpass the strongest baselines (DeepScaleR, STILL-3) trained on larger datasets, and the paper analyzes this gap directly.
Related guides (3)
Related events (8)
ExpRL: RL-based mid-training using human QA data as reward scaffolds for LLM reasoning
ExpRL proposes an automated approach to LLM mid-training that replaces manually curated reasoning traces with large corpora of human-written QA data used as reward scaffolds rather than imitation targets. Reference solutions are hidden from the policy and used only to construct problem-specific grading rubrics, enabling dense process-level rewards that reinforce partial progress and intermediate reasoning steps. On challenging math reasoning benchmarks, ExpRL outperforms SFT, sparse-reward GRPO, and self-distillation as an RL initialization strategy, with additional mixed-domain experiments suggesting broader applicability.
RA-RFT: Retrieval-Augmented Reinforcement Fine-Tuning teaches LLMs to reason by analogy
Researchers propose Retrieval-Augmented Reinforcement Fine-Tuning (RA-RFT), a post-training framework that trains a retriever to rank contexts by expected reasoning benefit rather than semantic similarity, then fine-tunes a policy model via reinforcement learning using retrieved analogous demonstrations. The key insight is that reasoning-relevant retrieval surfaces complementary solution strategies rather than superficially similar problems. On mathematical reasoning benchmarks, RA-RFT improves AIME 2025 average@32 accuracy by 7.1 and 2.8 points over GRPO for Qwen3-1.7B and Qwen3-4B respectively, suggesting reasoning-aware retrieval is orthogonal to reward design and training curriculum improvements.
MOPD: Multi-Teacher On-Policy Distillation for integrating multiple RL-trained capabilities in LLMs
Researchers propose Multi-teacher On-Policy Distillation (MOPD), a post-training paradigm that first trains domain-specialized RL teacher models, then distills them into a student model using on-policy rollouts to eliminate exposure bias. Evaluated on Qwen3-30B-A3B, MOPD outperforms Mix-RL, Cascade RL, Off-Policy Finetune, and Param-Merge baselines while preserving nearly all per-domain capability. The method has been deployed in production for MiMo-V2-Flash, an industrial-scale frontier model, validating its practical applicability. The approach also enables parallel, decoupled development of domain teachers, reducing cross-domain interference in multi-capability post-training.
Mach-Mind-4-Flash: 35B MoE agentic model matching 100B-class performance via post-training optimization
Mach-Mind-4-Flash is a 35B-parameter Mixture-of-Experts model with only 3B activated parameters that achieves performance comparable to 100B-class models through post-training techniques alone. The pipeline combines a unified RL/OPD training infrastructure with multi-teacher scheduling, parallel domain-specific RL experts fused via Multi-Teacher On-Policy Distillation (MOPD), and Hybrid Median-length Policy Optimization (HMPO) which compresses reasoning chains 19-46% with minimal accuracy loss. Benchmark results include 92.70 on AIME'26, 82.82 on IFBench, and 75.80 on BFCL-v4, claiming to lead or match models 10-30x its activated size at a fraction of inference cost. The work is notable for demonstrating that post-training optimization can close large gaps in activated parameter count for agentic tasks.
Instruction tuning plus model merging improves reasoning LMs in verifiable and unverifiable domains
A new arXiv preprint proposes a two-step method for adapting reasoning language models (RLMs) to domains lacking reliable verifiers: first apply standard supervised fine-tuning (without reasoning traces) on the target domain, then merge the resulting model with the original RLM to recover reasoning behavior. The technique is evaluated on both verifiable domains (coding) and hard-to-verify domains (text summarization), showing improvements while preserving general RLM capabilities. Notably, the authors report the full adaptation can be achieved for under $3, making it highly accessible.
Progress Advantage: Annotation-Free Step-Level Scoring for LLM Agents via RL Post-Training
Researchers introduce 'progress advantage,' a method that derives implicit step-level reward signals for LLM agents directly from the log-probability ratio between an RL-trained policy and its reference policy, without requiring dedicated process reward model training. The approach is shown to recover the optimal advantage function under a general stochastic MDP formulation, making it annotation-free and domain-agnostic. Validated across five benchmarks and four model families on tasks including test-time scaling, uncertainty quantification, and failure attribution, it outperforms confidence-based baselines and even dedicated trained reward models. The result is practically significant because building process reward models for agentic settings is currently a major bottleneck.
MAST: Mechanism-guided selective unlearning for RLVR-trained reasoning models
Researchers introduce MAST (Mechanism-Aligned Selective Targeting), a method for selectively unlearning capabilities induced by reinforcement learning from verifiable rewards (RLVR) in language models while minimizing collateral damage to retained knowledge. The approach ranks attention-projection tensors by off-principal energy and gradient coupling to identify a targeted subset for update, rather than applying full-parameter gradient ascent. Evaluated on Qwen2.5-Math-1.5B and Qwen3-1.7B-Base, MAST achieves statistically significant forgetting on target MATH problems while preserving GSM8K performance, whereas full-parameter unlearning collapses retained capabilities. The method generalizes across seeds and unlearning objectives (NPO/SimNPO).
OPERA: Perplexity-based RL alignment for open-ended reasoning tasks
OPERA (Objective Perplexity-based Reflective Alignment) proposes replacing LLM-as-a-judge reward models with intrinsic rewards derived from perplexity dynamics to stabilize RL training on open-ended tasks like creative writing. The method includes a cold-start data synthesis pipeline generating 20,000 reasoning trajectories using perplexity-prioritized rollouts. Applied to Qwen3-8B, OPERA claims state-of-the-art among open-source models on open-ended tasks, reportedly matching or exceeding Gemini 2.5 and MiniMax-M2.5 on some benchmarks.