Open Models Almanac
Methodstraining objective

methods · pipeline stage 2 of 6

training objective

84 methods filed at this node or below it, from the sources of 28 models.

In this branch 84

Everything filed at this node or below it, with one collapsible heading per child node.

filed here 11

Winner-takes-all action-candidate supervision used · 2 sources · 2 quotes
10 single mentions
Predictive dynamics modeling core · 1 source · 2 quotes
Vision distillation core · 1 source · 1 quote
GAN training used · 1 source · 1 quote
Multiple flow-time samples per action chunk used · 1 source · 3 quotes
Optimal-transport path for flow matching used · 1 source · 1 quote
Target-distribution calibration loss used · 1 source · 1 quote
Valid-pixel-masked L1 depth loss used · 1 source · 1 quote
VQ-VAE reconstruction loss used · 1 source · 1 quote

language modelling objective 15

Autoregressive discrete-action prediction core · 3 sources · 3 quotes
Action-token-only cross-entropy loss core · 2 sources · 3 quotes
Causal next-token prediction core · 2 sources · 2 quotes
12 single mentions
Chain-of-thought labeling used · 1 source · 1 quote
Cross-entropy objective used · 1 source · 1 quote
Discrete-tokenized pretraining used · 1 source · 1 quote
High-level subtask prediction used · 1 source · 1 quote
Sampled-token objective used · 1 source · 1 quote
Stricter token constraints during training used · 1 source · 1 quote
Text pre-training used · 1 source · 1 quote
Fill-in-the-middle completion optional · 1 source · 1 quote
Implicit high-level inference from training data evaluated · 1 source · 1 quote

multi-token prediction objective 7

MTP Boosting used · 6 sources · 7 quotes
Multi-Token Prediction used · 5 sources · 7 quotes
5 single mentions
Continued pretraining for MTP layers used · 1 source · 1 quote
Frozen-backbone MTP-head fine-tuning used · 1 source · 1 quote
LK loss used · 1 source · 1 quote
MTP warm-up used · 1 source · 1 quote
Shared-weight MTP objective used · 1 source · 1 quote

distillation objective 25

Full-vocabulary logit distillation core · 3 sources · 3 quotes
Reverse KL divergence core · 2 sources · 2 quotes
Quantization-aware distillation used · 4 sources · 5 quotes
Skew reverse-KL distillation used · 2 sources · 2 quotes
Temperature-scaled forward-KL distillation used · 2 sources · 2 quotes
Logit matching evaluated · 2 sources · 3 quotes
19 single mentions
Chain-rule decomposition of Top-K KL core · 1 source · 1 quote
Conditional Top-K-only temperature scaling core · 1 source · 3 quotes
Decoupled Top-K knowledge distillation core · 1 source · 3 quotes
Forward-KL distillation with a sink token used · 1 source · 1 quote
Full next-token distribution distillation used · 1 source · 1 quote
Hidden-state caching for KL distillation used · 1 source · 1 quote
KL-divergence distillation objective used · 1 source · 2 quotes
Depth distillation optional · 1 source · 1 quote

auxiliary loss 26

KL alignment loss used · 3 sources · 4 quotes
Knowledge insulation used · 3 sources · 5 quotes
Sample-level attention masking used · 2 sources · 2 quotes
22 single mentions
Auxiliary object-detection loss core · 1 source · 1 quote
Log-probability-ranked sentence masking core · 1 source · 1 quote
Natural masks for missing depth regions core · 1 source · 1 quote
Squared log-policy-ratio regularizer core · 1 source · 1 quote
Domain-specific KL regularization strength used · 1 source · 2 quotes
Gradient detach used · 1 source · 2 quotes
KL regularization to a reference policy used · 1 source · 1 quote
Language modeling loss plus KL loss used · 1 source · 1 quote
Loss masking of flagged trajectory content used · 1 source · 1 quote
MoE sequence auxiliary loss used · 1 source · 1 quote
Unfinished-trajectory loss masking used · 1 source · 1 quote

By model

Which of this branch's methods each model's own documents describe, and how strongly. Under each model: its strongest treatment anywhere in the branch.

Modelfiled herelanguage modelling objectivemulti-token prediction objectivedistillation objectiveauxiliary loss
DeepSeek-V4-Flash-0731 optional—Fill-in-the-middle completion optional————
MiMo-V2.6-Flash used—Cross-entropy objective used———Loss masking of flagged trajectory content used—
NVIDIA-Nemotron-3-Ultra-550B-A55B used—Sampled-token objective used—Frozen-backbone MTP-head fine-tuning usedMTP Boosting usedShared-weight MTP objective used—Temperature-scaled forward-KL distillation usedLogit matching evaluated—Unfinished-trajectory loss masking used—
DeepSeek-V4-Flash used———Full-vocabulary logit distillation used—Sample-level attention masking used—
Alpamayo 1.5 / 2 Super used—Joint cross-entropy training of reasoning and trajectory tokens usedAutoregressive discrete-action prediction evaluated———KL regularization to a reference policy usedKnowledge insulation used—
DeepSeek-V3.2 used—Stricter token constraints during training used———Domain-specific KL regularization strength usedKL alignment loss used—
DeepSeek-V4-Flash-Vision-Exp optional—Fill-in-the-middle completion optional————
DeepSeek-V4-Pro used———Full-vocabulary logit distillation used—Sample-level attention masking used—
DeepSeek-V4-Pro-0813 optional—Fill-in-the-middle completion optional————
GR00T N1.7 core————Auxiliary object-detection loss coreSquared-error loss for normalized object-center coordinates core—
Kolibri-1 core—Causal next-token prediction core———Log-probability-ranked sentence masking corePer-step policy-dependent recomputation of context masks coreSquared log-policy-ratio regularizer core—
LFM3-1.7B coreGAN training used—Large-scale pre-training on 22 trillion tokens used——Chain-rule decomposition of Top-K KL coreConditional Top-K-only temperature scaling coreDecoupled Top-K knowledge distillation coreCross-encoder-to-ColBERT knowledge distillation usedMean squared error distillation of relevance scores usedMin-max normalization of teacher relevance scores usedTeacher-weighted reverse-KL surrogate for Top-K distillation evaluated—Weighted averaging of audio codebook losses used—
Laguna-S-2.1 used———Hidden-state caching for KL distillation usedQuantization-aware distillation used——
LingBot-VLA 2.0 corePredictive dynamics modeling coreVision distillation coreValid-pixel-masked L1 depth loss usedVideo-adapted DINO and iBOT self-distillation used———Depth-representation distillation for geometric supervision coreDistillation-loss alignment of VLM queries with depth tokens coreDistilling current and future perceptual queries from pretrained perception models usedDepth distillation optional—Natural masks for missing depth regions core—
MiMo-V2.5 used—Text pre-training used—Multi-Token Prediction used—Reverse KL divergence used—MoE sequence auxiliary loss used—
MiMo-V2.5-Pro used—Text pre-training used—Multi-Token Prediction used———
MiMo-V2.6-Pro used—Cross-entropy objective used———Loss masking of flagged trajectory content used—
MiniMax-M3 used————Gradient detach usedKL alignment loss usedLanguage modeling loss plus KL loss used—
Mistral-Large-4 core———Full-vocabulary logit distillation core—Reference-policy KL penalty using the k2 estimator optional—
MolmoAct 2 coreJoint discrete and continuous action training coreMultiple flow-time samples per action chunk usedVQ-VAE reconstruction loss used—Action-token-only cross-entropy loss coreAutoregressive discrete-action VLA pre-training coreCausal next-token prediction core——Distillation of specialist depth estimates into autoregressively predicted depth tokens core—Attention-mask isolation of packed examples usedDepth-prefix token replacement noise during fine-tuning usedKnowledge insulation used—
NVIDIA-Nemotron-3.5-Lightning-30B-A3B coreTarget-distribution calibration loss used——Continued pretraining for MTP layers usedLK loss usedMTP Boosting usedMulti-Token Prediction used—Reverse KL divergence coreBackbone-to-MTP-head knowledge distillation usedCombined KL-divergence and cross-entropy distillation loss usedForward-KL distillation with a sink token usedFull next-token distribution distillation usedKL-divergence distillation objective usedQuantization-aware distillation usedSkew reverse-KL distillation used—Disable auxiliary and hard-target losses during QAD usedMask non-target tokens in rollout training data used—
NVIDIA-Nemotron-3.5-Super-VL-120B-A12B coreTarget-distribution calibration loss used——LK loss usedMTP Boosting usedMTP warm-up used—Reverse KL divergence coreBackbone-to-MTP-head knowledge distillation usedCombined KL-divergence and cross-entropy distillation loss usedForward-KL distillation with a sink token usedFull next-token distribution distillation usedKL-divergence distillation objective usedQuantization-aware distillation usedSkew reverse-KL distillation used—Disable auxiliary and hard-target losses during QAD usedMask non-target tokens in rollout training data used—
OpenVLA / OpenVLA-OFT core—Action-token-only cross-entropy loss core————
Qwen3.5-397B-A17B used——Multi-Token Prediction used———
Qwen3.8-Flash-Next used——Multi-Token Prediction used—Dense-attention distillation to initialize the QSA indexer used——
X-VLA usedOptimal-transport path for flow matching used—————
Xiaomi-Robotics-1 usedWinner-takes-all action-candidate supervision used—Chain-of-thought labeling used———Auxiliary action-generation supervision for the vision-language model usedBeta-distributed flow-matching timestep sampling usedLoss masking for missing embodiment-specific action dimensions used—
π0.5 core—Autoregressive discrete-action prediction coreDiscrete-tokenized pretraining usedHigh-level subtask prediction usedImplicit high-level inference from training data evaluated———Attention masking to prevent information leakage between action representations usedKnowledge insulation used—
↑