Theory says REINFORCE can collapse coverage in LM post-training
Resetting the KL reference in Approximate Policy Mirror Descent is shown to restore global convergence for stylized PPO and GRPO.
Stories per day · 30 days
18 RSS
Relatedreinforcement-learning8language-models3agents2decentralized2multi-agent2qwen32recursive-self-improvement2agent-evaluation1agent-harness1agent-swarms1ai-accessibility1alignment1arxiv1attack-success-rate1autonomous-research1benchmarks1cipr1claude1
99 stories, newest first · one card per story
Stories tagged #post-training
FewStepRL adapts diffusion RL objectives to few-step samplers
The method recasts FlowGRPO, AWM, DiffusionNFT, DiffusionOPSD and ReFL so rollouts no longer need a many-step sampler.
Trajectory raises $45M to make AI model training accessible to nontechnical experts
The startup's continual learning platform lets businesses retrain open-source AI models using user feedback for specialized tasks.
1
Keepace.ai post-trains Qwen3.5-27B for exercise guidance
Internal expert rubrics report gains in safety, boundary compliance, output completeness, and scientific accuracy.
Trained agent policies carry harness preferences that break under interface changes
Reordering MiniWoB++ observation fields drops trained success from 85.0% to 19.2% without changing optimal decisions.
NYU and UT Austin study finds six patterns that make AI text hard to read
Across 14 models, post-training and verifiable-reward training pushed text further from human writing than base models.
1
ParetoRSI-Gym evaluates multi-domain recursive self-improvement
Claude Code with Opus-5 raised equal-weighted generalization scores 9.2% relative, or 2.3 absolute points, over the starting models.
ChronoReplay reduces interface lock-in in spatial VLMs
One checkpoint improves OVO-S-Bench, OST-Bench and UCS-Bench together, including a 13.41-point gain on OST-Bench.
RSI Arena studies autonomous post-training of Nemotron with human feedback
The paper reports GPT-6 Astra closed a median 76% of the base-to-Instruct gap, and eight of ten models improved after a second stage.
Microsoft post describes ThinkingBox rewards for agent training
The title says database grading is converted into reinforcement-learning rewards; the article body was not available.
MMVistaReason open recipe trains multimodal reasoners with less data
MVR-4B averages 72.8 on 15 benchmarks, beating Qwen3.5-9B Instruct and MMFineReason-8B with about 70% fewer samples.
AgenticRoboRL post-trains robot planners across different harnesses
On LIBERO-PRO, a 9B planner rose from 18.1% to 71.0% success without external memory.
Only 37 of 1,598 RL method papers reach training frameworks
Papers included within 180 days stayed at 5 to 7 while quarterly cohorts grew from 72 to 386.
Calibrated accommodation pairs reward dialect matching in post-training
Across 10 open lineages, adjusted DPO still favored British and Australian spelling over AAE and Singlish.
Safety mid-training lowers LLM attack success before post-training
Authors report a 72% to 81% relative attack-success drop versus Qwen3 1.7B, 4B, and 8B bases.
Learning Self-Orchestration in Decentralized Agent Swarms
Market-based post-training lets decentralized agent swarms jointly learn actions and self-orchestration without a central coordinator.
Athanor: An Agent-Native System for Autonomous Post-Training Research
Athanor agents autonomously post-trained Qwen3-4B-Base to match official Qwen3-4B-Instruct on held-out tasks.
Learning Self-Orchestration in Decentralized Agent Swarms
Market-based post-training lets decentralized agent swarms learn both actions and self-orchestration.
One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness
CIPR trains LLM agents to stay consistent across renamed or notationally changed interfaces.
GAFU uses generation-side tokens for fine-grained visual understanding
Authors report relative gains of up to 20.00% on VStar-Bench versus same-data SFT baselines.
Cowork-Gym: Synthesizing Interactive Work Environments at Scale
Cowork-Gym post-trains Qwen agents on 64.9K synthetic workplaces, lifting six-benchmark scores.
Continually Post-training Small Models via Synthetic Self-distillation
SynthSelf lets small models continually post-train by outsourcing in-distribution rewrites to a slightly larger same-family aligner.
Predicting Alignment Generalization with Value Representations
Activation-based value representations predict LLM alignment generalization far better than textual descriptions of values.
Before They Can Solve: Predicting Post-Training Coding-Agent Performance from Base Models
Three screens rank base models by likely coding-agent performance after expensive post-training.