Papersby GlyphThe Alignment Waltz: Jointly Training Agents to Collaborate for Safety research paper by Meta, 2025
Meta · Alignment and safety · Oct 9, 2025 · 13 citations · 41 upvotes · unverified
What it shows
WaltzRL, a multi-agent reinforcement learning framework, improves LLM safety and helpfulness by collaboratively training a conversation agent and a feedback agent, reducing unsafe responses and overrefusals.
By Jingyu Zhang, Haozhu Wang, Eric Michael Smith and 7 more · arXiv 2510.08240 · PDF
UnverifiedHugging Face's summary; not yet checked by hand.
More from Meta
All 40InteractiveWorld: A Verifiable Human-Centric Environment for Building Better Interactive MLLM AssistantsInteractiveWorld provides a verifiable simulation environment and benchmark to evaluate whether multimodal language models can reliably guide humans through physical tasks.
Meta46 upvotesRead →Native Action-Prior Learning from Videos for World Action ModelsNAVA-WAM learns robot action priors directly from observation-only videos, improving control performance without needing action-labeled data for pretraining.Meta88 upvotesRead →Sharpening Tax in Post-TrainingThis paper introduces a metric to quantify the loss in solution coverage after reinforcement learning and proposes a sampler to mitigate it.Meta1 citations111 upvotesRead →World Action Modeling with Progressive Visual PlanningProWAM improves robotic control by predicting sparse visual sub-goals to guide action generation, avoiding the need for dense video generation.Meta1 citations98 upvotesRead →Context Language ModelsContext Language Models manage their own context by treating it as an editable file, improving performance and efficiency.Meta5 citations47 upvotesRead →In-Flight KV Cache with Clean Anchors for Faster Autoregressive Video DiffusionFlashForward accelerates video diffusion by reusing in-flight KV caches and using sparse clean anchors to maintain structural consistency.Meta1 citations34 upvotesRead →WearableQA: A Benchmark for Health Reasoning over Real-World Wearable DataWearableQA is a benchmark of multiple-choice questions derived from real longitudinal wearable data that evaluates large language model reasoning across data and health dimensions.Meta43 upvotesRead →Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and RecipesControlled experiments on when text, image understanding and image generation help or compete when trained together.Meta7 citations63 upvotesRead →
Other alignment and safety papers
TopicTraining language models to follow instructions with human feedbackInstructGPT: fine-tuning on human feedback made a 1.3B model preferred over the 175B GPT-3.
OpenAI25.8k citations26 upvotesRead →Direct Preference Optimization: Your Language Model is Secretly a Reward ModelDPO aligns a model to human preferences with a simple classification loss, no reward model or RL loop.
Stanford University11.6k citations73 upvotesRead →Constitutional AI: Harmlessness from AI FeedbackConstitutional AI trains a harmless assistant from AI feedback guided by a short list of written principles, not human harm labels.
Anthropic3,987 citations4 upvotesRead →Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMsReinforcement Learning from Human Feedback (RLHF) benefits from simpler REINFORCE-style optimization methods over computationally expensive PPO or "RL-free" methods, enhancing performance and reducing cost in large language models.
Cohere804 citations18 upvotesRead →Sleeper Agents: Training Deceptive LLMs that Persist Through Safety TrainingModels trained with a hidden backdoor kept their deceptive behaviour through standard safety training.Anthropic657 citations30 upvotesRead →Alignment faking in large language modelsClaude 3 Opus sometimes went along with a training goal it disagreed with, to avoid being changed, without being told to.Anthropic393 citations10 upvotesRead →
About this paper
- Venue
- arXiv.org
- Citations
- 13, 1 influential · Semantic Scholar
- Upvotes
- 41 · Hugging Face
- Lab
- Meta · on Companies · on Acquisitions · on Quarterly · on Paydays · on TechConf
Changes
Citations rose from 11 to 13.
Citations rose from 10 to 11.
A first count of 1 influential citations.
A first count of 10 citations.
Found by the weekly Hugging Face scan; not yet checked by hand.