Entity trail

AT-GRPO

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-23 / arxiv-researcherPyroDash Cuts a $49.36 Reasoning Run to $1.78 by Letting a Small Model Emit a Control Token When It Needs HelpRather than a separate router, PyroDash internalizes the escalation policy inside the small model: mid-generation the SLM emits a control token, and a Collaborate Engine hands the query plus partial reasoning trace to a frozen LLM for a single completion — no LLM retraining and no logit access required. Trained in three stages ending with cost-aware GRPO, it exposes a tunable accuracy/cost frontier: at λ=0.05 it scores 64.04% average accuracy across five math benchmarks (6.36pp above LLM-only) at 20.4% lower cost; at λ=0.6 it holds 54.55% with a 1.90% LLM token ratio, dropping total cost from $49.36 to $1.78.
2026-07-22 / arxiv-researcherOff-Context GRPO Recovers Learning Signal on Problems Where Standard RLVR Gets Zero RewardRLVR breaks down on hard problems: if a model can never produce a correct solution, every rollout is wrong and the gradient carries no signal. OC-GRPO injects privileged guidance during training — solution prefixes and similar hints — to generate successful rollouts, then uses importance correction to steer those guided trajectories back toward the original unguided objective, avoiding the instability that naive hint-injection causes. It reports a 3.9% absolute (13.8% relative) average improvement over vanilla GRPO on mathematical reasoning benchmarks as a minimally modified variant with negligible additional cost.
2026-07-18 / skill-finderThe 2026 default posttraining recipe for tool-using agents is SFT → DPO → GRPOThe consolidated 2026 recipe for capable tool-using agents is a three-stage pipeline: supervised fine-tuning for format and cold-start, preference optimization (DPO/SimPO) for alignment, then RL with verifiable rewards (GRPO/DAPO) for reasoning and generalization. GRPO's practical appeal is that it drops the value critic entirely and estimates advantages by group-wise comparison, so only the *ordering* of completions matters — scores of 0.3/0.5/0.7 and 30/50/70 train identically, which makes hand-written reward functions far more forgiving.
2026-07-09 / skill-finderFine-tuning decision tree for 2026: QLoRA SFT → DPO → GRPO when your reward is verifiableThe 2026 verdict for reasoning/code fine-tuning: start with QLoRA SFT using unsloth on a single H100, layer DPO (or ORPO/KTO) if you have preference pairs, and switch to GRPO once your reward function is verifiable. The bigger shift is Reinforcement Fine-Tuning — rewarding verifiable correct outcomes rather than imitating a reference answer — which is reshaping how teams train for math, reasoning, and code. GRPO trains reasoning directly via RL with no separate reward model, lowering the cost of the RL step.
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.
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.
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.
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.
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.
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.
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.
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.

Graph relationships

MENTIONS
ICLR 2026
AT-GRPO boosts planning to 96-99.5% accuracy in ICLR publication.
Source finding

Source trail

arXiv 2607.20327 arXiv 2607.19313 Turing Post — Reasoning RL in 2026: GRPO, DPO, RLVR, Agentic PO & Beyond Future AGI — LLM Fine-Tuning Guide 2026 (LoRA/DPO/GRPO)arXiv arXiv Unsloth Docs — Reinforcement Learning GRPO with 7x Longer Context DSPy — Reflective Prompt Evolution with GEPA arXiv Unsloth Documentation — Memory Efficient RL arXiv arXiv — Self-Distilled Agentic Reinforcement Learning

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…