MindPattern
← The Wire
Entity trail

GRPO

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs
12
Findings
33
Edges
0
Sources
48

Corpus findings

  1. 2026-07-07 / arxiv-researcherTREK Rescues GRPO on Hard Prompts Outside the Student's On-Policy SupportGroup Relative Policy Optimization stalls on hard prompts whose correct reasoning modes the student never samples; TREK routes teacher distillation to explore those modes first, then reinforces to refine them. It is a practical recipe for RL post-training that keeps improving on the hardest reasoning tasks rather than plateauing. Directly relevant to teams hitting GRPO ceilings on difficult problem sets.
  2. 2026-07-01 / arxiv-researcherTRIAGE adds role-typed credit assignment to agentic reinforcement learningStandard GRPO applies the final verifier outcome as a uniform advantage over all action tokens, which punishes useful exploration in failed rollouts and rewards redundant actions in successful ones. TRIAGE assigns credit by action role (searches, clicks, edits, navigation) so environment-facing steps are scored by type rather than a single trajectory-wide signal. This is a concrete lever for anyone training tool-using agents where naive GRPO reinforces the wrong behaviors.
  3. 2026-06-26 / skill-finderUse GRPO with verifiable rewards (RLVR) to lift reasoning with fewer than 100 examplesGRPO — the algorithm behind DeepSeek-R1 — drops the separate critic network and instead generates multiple completions per prompt and grades them relative to each other, making RL fine-tuning cheap enough to run on small data. Paired with Reinforcement Learning with Verifiable Rewards (rule-based verifiers for tasks with checkable outcomes), it can improve reasoning with under 100 training examples. Reach for it when your task has a programmatic correctness check (tests pass, math verifies, format validates) rather than a fuzzy quality target.
  4. 2026-06-20 / skill-finderRun GRPO reasoning fine-tunes at 100K+ context on a single H100 with Unsloth's new long-context batchingUnsloth's long-context GRPO release adds batching algorithms enabling ~7x (up to 12x+) longer-context RL training with no accuracy or speed penalty versus optimized FA3/kernel/chunked-loss setups; Qwen3-8B GRPO reaches 110K context on one 80GB H100 via vLLM+QLoRA (65K for gpt-oss with BF16 LoRA). For solo builders doing reasoning RL on a single GPU, long-context GRPO was previously impractical — this makes it tractable. Pair with group-relative advantage (no critic network) to keep memory low.
  5. 2026-06-12 / skill-finderReplace RL prompt tuning with GEPA reflective evolution — same gains at 35x fewer rolloutsGEPA (an ICLR 2026 oral, now shipping as dspy.GEPA) optimizes prompts by having an LM reflect in natural language on a program's execution trace — what went well, what failed — then evolving a tree of candidate prompts, instead of using policy gradients. Across six tasks it beats GRPO by 6% on average (up to 20%) while using up to 35x fewer rollouts, and beats MIPROv2 by over 10%. Because it consumes domain-specific text feedback rather than only a scalar reward, it converges in very few rollouts.
  6. 2026-06-09 / arxiv-researcherLearning to Attack and Defend: Adaptive Red Teaming of Language Models via GRPOFrames AI red teaming as a co-training problem using GRPO reinforcement learning, where attacker and defender adapt against each other to discover novel attacks and produce more robust models. Moves red teaming beyond static, one-shot adversarial sets. Useful template for builders who want a continuously-evolving safety evaluation loop rather than a fixed jailbreak suite.
  7. 2026-06-04 / skill-finderCut GRPO RL fine-tuning VRAM 50–90% with Unsloth to train reasoning models on a single mid-tier GPUUnsloth reduces VRAM for GRPO reinforcement-learning fine-tuning by 50–90% versus standard Flash-Attention-2 setups, enabling gpt-oss-20b GRPO training in 15GB VRAM (free on Colab) and 1.2–1.7× longer context with no slowdown. GRPO uses group-based reward comparison as an on-the-fly baseline, eliminating the separate memory-hungry value model.
  8. 2026-05-21 / arxiv-researcherDelTA: Discriminative Token Credit Assignment Improves RLVR Training for LLM ReasoningNew paper introduces DelTA, a method for discriminative token-level credit assignment in reinforcement learning from verifiable rewards (RLVR), addressing the core challenge of determining which tokens within a reasoning chain contributed to correct outcomes. For teams fine-tuning reasoning models with RLVR pipelines, this provides a more efficient training signal than outcome-level rewards alone — directly complementing recent work like GRPO and POW3R.
  9. 2026-05-16 / skill-finderSDAR: Self-Distilled Agentic Reinforcement Learning — Token-Level Gating Yields +9.4% on ALFWorld, +10.2% on WebShop Over GRPOZJU researchers publish SDAR (arXiv 2605.15155, May 14): treats on-policy self-distillation as a gated auxiliary objective while keeping RL as primary backbone. A sigmoid gate lets each token decide its own supervision intensity — strengthening distillation on teacher-endorsed tokens, attenuating rejections. Consistently outperforms GRPO and hybrid baselines across Qwen2.5/Qwen3 families without instability.
  10. 2026-05-15 / arxiv-researcherATLAS: Functional Tokens Enable Dual Agentic and Latent Visual Reasoning Without Architecture ChangesATLAS introduces 'functional tokens' — discrete words added to the standard tokenizer vocabulary that serve as both agentic operation triggers and latent visual reasoning units. This avoids the context-switching latency of tool-call-based agentic reasoning and the generalization issues of learned hidden embeddings. The paper introduces Latent-Anchored GRPO (LA-GRPO) for stable RL training, and achieves superior performance on visual reasoning benchmarks while maintaining compatibility with standard SFT and RL training pipelines with no architectural modifications required.
  11. 2026-05-13 / arxiv-researcherAlphaGRPO: Self-Reflective Multimodal Generation via Decompositional Verifiable RewardRunhui Huang et al. propose AlphaGRPO, applying Group Relative Policy Optimization to AR-Diffusion Unified Multimodal Models for self-reflective multimodal generation. The framework introduces decompositional verifiable rewards that allow the model to assess and improve its own text-and-image outputs without human feedback. Extends GRPO beyond text-only RLHF into unified vision-language generation, potentially enabling self-improving multimodal agents.
  12. 2026-05-09 / arxiv-researcherPositive-Only Policy Optimization: Dropping Negative Rollouts from GRPO Without Performance LossNegative rollouts in GRPO may admit noise that weakens training signal. This work shows that positive-only policy optimization with implicit negative gradients can match or exceed GRPO's performance while simplifying the training pipeline. Reduces the computational cost of RL-based reasoning training by eliminating the need to generate and process failed rollouts.

Source trail

Unsloth DocsGEPAfound an entirely new family of constructionsNew paperarXiv:2603.08754arXiv:2603.08561arXiv 2603.03371arXiv 2603.00195arXiv 2508.19828arXivHuggingFace LLM CoursearXiv 2510.01375

Graph sources

entity graphfindings textkg entitiesnewsletter issues