PaperlayerView exact revision on arXiv

Source-linked AI summary

Judging in Latent Space: Efficient Generative Reward Modeling via Semantics-Preserving Compression

Mingqing Yuan, Xiaobo Liang, Junwei Yang, Ziwei Chen, Zeren Zhang, Hejin Wang, Yubin Wang, Juntao Li

arXiv:2610.09788v2cs.CL

TL;DR

Reward models must integrate multiple evaluation criteria, and repeated verbalized judgments make inference efficiency important. LatentGRM uses rubric structure to compress evaluations into continuous trajectories for autonomous pairwise judgments, with a separate interpreter for reconstruction. It achieves competitive aggregate accuracy against matched SFT judges while reducing trajectory length and inference time.

  • Problem

    Reward modeling must combine evidence across multiple response criteria, while accuracy and inference efficiency matter because reward models are used repeatedly.

    Source: §1 · p. 2

  • Method

    LatentGRM uses rubric-guided semantic chunking to compress evaluations into continuous trajectories for autonomous pairwise judgments, with a separate interpreter reconstructing evaluation text.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    Across four domains at vote@5, LatentGRM-8B compresses evaluation trajectories by 8.9–9.2× and reduces total judge inference time by 6.1–7.0×, while achieving competitive aggregate preference accuracy against matched SFT judges.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Controlled rubric interventions show criterion-dependent preference information in latent trajectories, while the interpreter recovers criterion-level assessments for retrospective inspection.

    Source: §1 · pp. 2–3

Abstract

from arXiv · show

Reward modeling often requires jointly representing and reasoning over multiple evaluation criteria, yet verbalizing this process token by token can incur substantial inference cost. Recent work on latent reasoning suggests that continuous states may support this computation more compactly. We introduce LatentGRM, a latent evaluation framework built on semantic chunking, compression, and reconstruction. By using the structure of rubric-guided evaluations to guide compression, LatentGRM learns compact continuous trajectories that support autonomous pairwise judgments without generating textual assessments. A separate interpreter reconstructs evaluation text from these trajectories, providing an offline view of the information retained under compression. Under matched training data and backbones, LatentGRM achieves competitive aggregate preference accuracy relative to explicit Supervised Fine-Tuning (SFT) judges at both 4B and 8B scales. Across four benchmark domains, LatentGRM-8B compresses evaluation trajectories by 8.9--9.2x and reduces total judge inference time by 6.1--7.0x at vote@5. Controlled rubric interventions show that criterion-dependent preference information is carried through the latent sequence. Together, these results demonstrate that continuous latent evaluation can substantially reduce inference cost while preserving competitive judgment quality.

1 Introduction

Multi-criterion reward judgments must balance accuracy with inference efficiency, but verbalizing rubric-based assessments can make evaluation trajectories costly. LatentGRM uses rubric structure to compress those assessments into autonomous latent judgments and reports competitive accuracy alongside shorter, faster trajectories and evidence about what they retain.

Motivation

  • Reward modeling must integrate multiple response criteria into preferences used for policy optimization and response selection, making both judgment accuracy and inference efficiency practically important.The motivating examples include a factually correct response that violates an explicit constraint and a detailed answer that lacks requested concision.Source: §1 · p. 2

Research question

  • Rubric-based judges can improve quality through sampling and aggregation, but generating criterion assessments and evidence lengthens trajectories and increases autoregressive decoding costs.The central question is whether a judge can retain information needed for multi-criterion evaluation while computing more compactly.Source: §1 · p. 2

Approach

  • LatentGRM uses rubric-aligned semantic units to allocate a fixed latent budget, autonomously predict pairwise preferences, and let a separate interpreter reconstruct saved trajectories offline.Criterion-level organization and linguistic boundaries guide compression, while the criteria also support probing latent judgments and recoverable assessment content.Source: §1 · p. 2

Results

  • Across eight benchmark domains, LatentGRM matches competitive aggregate preference accuracy against explicit SFT judges trained on matched data and backbones at 4B and 8B scales.On four domains, LatentGRM-8B reduces trajectory length by 8.9–9.2× and total judge inference time by 6.1–7.0× at vote@5.Source: §1 · pp. 2–3

Retained information

  • Controlled rubric edits show that latent trajectories carry criterion-dependent preference information, while reconstruction recovers criterion-level supporting justifications from teacher-derived trajectories.The framework combines rubric-guided compression, autonomous latent evaluation, and retrospective interpretation.Source: §1 · pp. 2–3

2 Related Works

Prior latent-reasoning work mainly studied mathematical and general reasoning, while reasoning-heavy generative reward judges face substantial test-time costs. This paper adapts latent reasoning to rubric-guided reward modeling and uses rubrics both to structure compression and to analyze it.

2.1 Latent Reasoning

  • Latent reasoning replaces explicit intermediate tokens with recurrent continuous states, but prior methods were studied primarily on mathematical and general reasoning tasks.The paper adapts vocabulary-space latent reasoning to represent multi-criterion comparisons between candidate responses.Source: §2.1 · p. 3

2.2 Reward Modeling

  • Reasoning-based rubric judges generate and apply explicit criteria at test time, creating inference costs that LatentGRM addresses through continuous trajectories and matched explicit SFT comparisons.Rubrics also serve as anchors for semantic chunking and as tools for evaluating and analyzing latent reasoning.Source: §2.2 · p. 3

3 Method

LatentGRM replaces textual evaluation traces with compressed continuous trajectories for pairwise reward judgments. Semantic chunking sets rubric-guided boundaries, Stage 2 distills autonomous latent generation, and a separate interpreter reconstructs evaluations offline.

3.1 Overview

  • LatentGRM replaces an explicit evaluation trace with fewer continuous latent steps, and its two-stage framework distills traces before training autonomous latent judging.The separate interpreter reconstructs explicit evaluations from saved trajectories outside the online judgment path.Source: §3.1 · p. 3, Figure 2 · p. 4

3.2 Stage 1: Semantic Chunking for Latent Compression

  • Semantic Chunking divides evaluation traces into spans, assigns each rubric block a fixed latent budget, and refines internal boundaries using sentence, clause, and dependency evidence.Its boundary placement aims to preserve semantic integrity as much as possible; the rate ρ is fixed.Source: §3.2 · p. 4, Figure 3 · p. 5
  • At each boundary, the encoder’s hidden state is expressed as a vocabulary-space distribution and continuous embedding so the decoder can predict the remaining evaluation and preference.The exported distributions become sparse targets for Stage 2.Source: §3.2 · pp. 4–5

3.3 Stage 2: Autonomous Latent Evaluation

  • Stage 2 trains the judge to predict perturbed latent targets autoregressively, then produce the final preference; at inference it feeds back its own predictions and uses Gumbel perturbation for diverse trajectories.The implementation extends vLLM for soft-embedding feedback while retaining continuous batching, KV caching, and tensor parallelism.Source: §3.3 · pp. 5–6

3.4 Retrospective Interpretation of Latent Trajectories

  • The separate Latent Trace Interpreter predicts the complete teacher evaluation from a completed trajectory, excluding the preference label; saved top-K indices and weights allow offline readout.Unlike Stage 1 suffix reconstruction, the interpreter reconstructs the evaluation from its beginning and is not run in the online decision path.Source: §3.4 · p. 6

4 Experiments

The experiments examine preference accuracy, inference cost, and reconstruction, comparing LatentGRM with matched explicit judges across eight benchmark domains. At vote@5, LatentGRM scores higher on Avg.-8 than the matched Rubric-RM at both model scales, while domain-level results vary.

4 Experiments

  • The experiments ask whether latent evaluation preserves preference accuracy and responds to rubric changes, how compression affects inference cost, and whether assessments can be reconstructed.Source: §4 · p. 6

4.1 Experimental Setup

  • The comparison uses 35,612 OpenRubrics training records and evaluates eight domains with single-rollout and five-rollout voting.Rubric-RM and LatentGRM use matched data and backbones, identical offline-generated task-specific rubrics, and scores averaged across both response orders.Source: §4.1 · p. 6

4.2.1 Preference Accuracy

  • At vote@5, LatentGRM scores 67.8 and 69.0 on Avg.-8 at 4B and 8B, versus 67.0 and 68.1 for matched explicit Rubric-RM.Table 1 reports pairwise preference evaluation, with Avg.-8 as the unweighted mean of eight benchmark scores.Source: §4.2.1 · p. 6, Table 1 · p. 7
  • Voting raises LatentGRM’s Avg.-8 from 65.6 to 67.8 at 4B (+2.2) and from 66.7 to 69.0 at 8B (+2.3), improving final judgments.Across the other seven domains, Chat Hard and IFBench favor LatentGRM at both scales, RewardBench Chat favors Rubric-RM, and Precise IF changes direction between scales.Source: §4.2.1 · p. 7

Appendix E.4 provides a comparison with a direct judge that predicts only the final preference.

LatentGRM combines semantic chunking with autonomous latent evaluation to reduce judgment cost while retaining criterion-dependent information. Its interpreter reconstructs rubric-level evaluations, and experiments assess chunking, latent information transfer, runtime, and revision utility.

4.2.2 Effect of Semantic Chunking

  • With the latent budget and training conditions held fixed, Semantic Chunking raises the four-domain mean by 1.8 points over Fixed-Rate Chunking at vote@5.It gains 2.2, 3.9, and 1.7 points on Chat Hard, Precise IF, and Focus, respectively, while Chat decreases by 0.7 points.Source: §4.2.2 · p. 7, Table 2 · p. 8

4.2.3 Rubric Interventions and Latent-Source Replacement

  • Criterion interventions shift preference toward the designated response, reaching a 4.58 mean log-odds change at the trajectory endpoint; replay under a fixed baseline prompt reaches 2.75.The wording-control effect remains smaller, and the replay result shows criterion-dependent preference information transfers through the latent sequence.Source: §4.2.3 · pp. 8–9, Figure 4 · p. 8

4.3.1 Runtime under Natural Generation

  • At vote@5, LatentGRM-8B reduces generated positions by 8.87–9.23× and total inference time by 6.11–6.95× under vLLM across four domains.The comparison is against Rubric-RM-8B; ordinary Transformers speedups are 8.17–8.82×.Source: §4.3.1 · p. 9, Table 3 · p. 9
  • At vote@9, LatentGRM remains faster, with 6.9×–7.8× speedups, while its accuracy advantage depends on domain: it leads on Chat Hard and Focus but trails on Chat.Precise IF varies by voting budget; more votes generally improve latent judgment quality, with nonmonotonic intermediate results.Source: §4.3.1 · p. 9

4.4 RQ3: Recovering Rubric-Level Evaluations

  • LTI raises State Macro-F1 to 98.93 and Vector Exact to 93.98 on the 607-record reconstruction test set.It also leads on Rubric Semantic (74.65) and ROUGE-L (79.24); Prompt-only achieves 63.36 ROUGE-L despite 40.20 State Macro-F1.Source: §4.4 · p. 10, Table 4 · p. 10
  • Interpreter-restored feedback yields 57.50% Full and 1.86% Regression in frozen-model revision, with a Full rate comparable to explicit Rubric-RM feedback.It has the best result in every reported column; the test uses 240 examples, equally split between Chat and Chat Hard.Source: §4.4 · p. 10, Table 5 · p. 10

5 Conclusion

  • LatentGRM uses rubric-guided semantic compression for autonomous pairwise judgments, while a separate interpreter reconstructs criterion-level assessments for retrospective inspection.The conclusion characterizes its aggregate accuracy as competitive with matched explicit SFT judges while substantially reducing trajectory length and inference time.Source: §5 · p. 11

A Discussion

The discussion identifies preserving judgment-relevant evidence during compression and faithfully reconstructing it as open challenges. Semantic Chunking allocates a fixed latent budget across rubric blocks, then refines boundaries using linguistic cues and constrained search.

A Discussion

  • Preserving judgment-relevant evidence during compression and faithfully recovering it in retrospective interpretation remain challenges; better compression and reconstruction strategies are future directions.The discussion also presents shorter generation and lower inference time as opportunities for making repeated reward evaluations more affordable in policy learning and data selection.Source: A Discussion · p. 16

B Semantic Chunking and Exact-Budget Allocation

  • Semantic Chunking allocates a fixed latent budget across rubric blocks, then refines their internal boundaries while preserving each block’s span and latent count.It preserves token order and produces N = ⌈M/8⌉ latent positions; preprocessing is performed offline and cached.Source: B Semantic Chunking and Exact-Budget… · p. 16, §B.3 · p. 18

B.1 Rubric Structure and Initial Budget Allocation

  • The allocator prioritizes keeping chunk lengths within 6–10 tokens, minimizing distance outside that range, then minimizing deviation from the target rate.Outer blocks distinguish compliance and gatekeeper discussion, criterion assessments, and final synthesis; protected structural expressions are avoided.Source: §B.1 · pp. 16–17

B.2 Linguistic Boundary Costs

  • Boundary costs favor sentence or complete-clause endings and discourage cuts through structural labels, punctuation, code, mathematics, and lexical units.Sentence rules and dependency evidence are combined; dependency risks are limited to relations spanning at most eight tokens, and fragment penalties discourage incomplete rubric fragments.Source: §B.2 · p. 17, Table 6 · p. 17

B.3 Constrained Local Refinement

  • Constrained local refinement first minimizes lexical cuts, then optimizes linguistic and fragment costs while controlling length, variance, and displacement from reference boundaries.A dynamic program respects feasible lengths and protected cuts; a candidate is accepted only when it reduces lexical cuts or improves the remaining criteria under the stated thresholds.Source: §B.3 · pp. 17–18

B.4 Realized Allocation and a Complete Example

Semantic Chunking retains the exact rate-8 latent budget while substantially changing boundary placement to reduce problematic cuts. The illustrated trace contrasts rubric-aligned frontiers with fixed-rate boundaries.

B.4 Realized Allocation and a Complete Example

  • 1,836,211 boundary moves reduce lexical cuts from 142,295 to 151 and risk-2 cuts from 694,857 to 946 in the training cache.The cache contains 35,612 evaluations and 3,458,245 latent positions; 86.18% of chunks contain 6–10 tokens.Source: §B.4 · p. 18, Table 7 · p. 19
  • The example semantic trace contains 219 explicit tokens and 28 latent frontiers aligned with coarse semantic regions.Uniform dark markers indicate frontier locations, and contiguous backgrounds show the coarse semantic hierarchy.Source: §B.4 · p. 18, Figure 5 · p. 19
  • Fixed-rate chunking uses the same trace and 28 latent tokens, but its 27 groups of eight and final group of three can split phrases, criteria, or semantic transitions.The figure contrasts those fixed-rate frontiers with the semantic frontiers in Figure 5.Source: §B.4 · p. 18, Figure 6 · p. 20

C Stage 1 Latent Distillation and Stage 2 Autonomous Evaluation

Stage 1 compresses teacher evaluation traces into latent states and trains a decoder to recover subsequent evaluation text and preference. Stage 2 learns autonomous latent trajectories and preference prediction, while a separate interpreter reconstructs evaluation text from saved trajectories.

C Stage 1 Latent Distillation and Stage 2 Autonomous Evaluation

  • Stage 1 trains an encoder and decoder on teacher evaluations, and Stage 2 trains a judge to predict latent states and pairwise preferences autonomously.The objectives are defined for evaluation input, teacher evaluation, and preference label; semantic segmentation adapts compression frontiers to rubric structure.Source: C Stage 1 Latent Distillation and… · p. 21

C.1 Encoder Sequence and Vocabulary-Space State

  • At each frontier, the encoder reads the input and preceding teacher-evaluation prefix, producing a state represented through the decoder’s vocabulary interface.The encoder cannot attend to future evaluation tokens or earlier compression placeholders; top-K vocabulary probabilities are renormalized for the state representation.Source: §C.1 · p. 21

C.2 Suffix Reconstruction Loss

  • The Stage 1 suffix-reconstruction objective trains each latent prefix to support the remaining explicit evaluation and final preference.Training samples one frontier per example and evaluates its corresponding causal suffix view.Source: §C.2 · pp. 21–22

C.3 The Three Stage 1 Optimization Phases

  • Stage 1 uses three optimization phases with the same reconstruction objective but different trainable parameter blocks, then exports sparse encoder targets for Stage 2.The encoder’s role ends after it exports targets from full teacher traces.Source: §C.3 · p. 22

C.4 Stage 2 Soft-Target Training and Inference

  • Stage 2 predicts exported top-K latent distributions with independent Gumbel perturbations, then uses the complete trajectory to predict the end marker, preference label, and EOS.Its stochastic objective combines latent distillation and explicit-token cross-entropy; prompt and latent positions are excluded from the latter.Source: §C.4 · pp. 22–23
  • At inference, the judge generates its own latent trajectory and stops latent generation when it predicts the end marker, then greedily decodes the preference label.Distinct Gumbel draws provide rollout diversity for voting.Source: §C.4 · p. 23

D Latent Trace Interpreter Details

  • The Latent Trace Interpreter reconstructs explicit evaluation text from latent trajectories using a separately trained adapter on frozen Stage 1 decoder weights.It reads saved vocabulary probabilities in trajectory order and generates text after the latent positions.Source: D Latent Trace Interpreter Details · p. 23, §D.1 · p. 23

D.2 Full-Trajectory Reconstruction Objective

  • The interpreter is trained to reconstruct the complete teacher evaluation from the beginning of the trajectory, masking prompt and latent positions from its loss.The binary preference label is excluded; in contrast, the Stage 1 decoder predicts the suffix after a sampled frontier.Source: §D.2 · pp. 23–24

D.3 LTI Training Settings

The supplementary analyses describe the data, rubric interventions, direct-judgment comparison, revision setup, and inference implementation. Direct judgment is a low-decoding-cost baseline, but its Chat Hard vote@5 accuracy trails explicit and latent evaluation.

D.3 LTI Training Settings

  • The LTI training configuration is summarized in Table 8, whose effective batch size includes data-parallel workers and gradient accumulation.The evaluation protocol also specifies maximum sequence length 6,144, top-K = 10, and 256 reserved latent positions in prompts.Source: §D.3 · p. 24, Table 8 · p. 24, §E.2 · p. 24

E.3.1 Data and Annotation

  • The rubric intervention analysis samples 398 accepted pairs from Chat and Chat Hard, edits one criterion while holding other text fixed, and independently verifies proposed edits.Pairs are evaluated in both response orders; the proposer and reviewer’s agreed criterion-level directions are not measured judge outcomes or overall winner labels.Source: §E.3.1 · pp. 25–26, Table 10 · p. 25, Table 9 · p. 25

E.3.2 Judge Analysis

  • Interventions establish criterion-directed preference shifts and transfer through latent prefixes, with effects averaged across response orders, pairs, and equally weighted domains.The paraphrase condition rewords a criterion without changing its meaning, serving as a control for changes attributed to criterion meaning.Source: §E.3.2 · p. 26

E.4 Direct Judgment and Voting

  • The DirectJudge comparison uses Qwen3-8B and tests Chat, Chat Hard, Precise IF, and Focus; DirectJudge receives the rubric and responses but generates no intermediate evaluation trace.It is initialized from the same backbone and trained for two epochs on the same OpenRubrics Dataset, supervising only the final verdict.Source: §E.4 · p. 26
  • DirectJudge reaches 67.43% on Chat Hard at vote@5, below Rubric-RM at 70.50% and LatentGRM at 73.00%.Increasing votes from one to five improves DirectJudge by only 0.22 points; the cited table compares direct judgment with explicit and latent evaluation.Source: §E.4 · p. 26, Table 11 · p. 27

E.5 Data Revision Protocol and Evaluation Details

  • The revision study compares prompt-only, explicit-trajectory, latent-only, and interpreter-restored feedback using the same fresh-context Qwen3-8B Base reviser.The latent-only interface is training-free: the reviser has not been trained to decode or consume LatentGRM trajectories.Source: §E.5 · p. 27
  • A separate evaluator scores every rubric criterion on original and revised answers and provides brief supporting evidence.On a preselected 20-example subset, reversing criterion order and paraphrasing the judging instruction yielded 93.76% criterion-level agreement and Cohen’s κ = 0.803.Source: §E.5 · p. 27

F.1 vLLM Implementation for Latent Reasoning

  • The vLLM implementation feeds continuous embeddings at latent steps, returns to discrete decoding for the preference label, and supports independent rollouts under continuous batching.The explicit chain-of-thought baseline uses the same scheduler and request-submission policy.Source: §F.1 · p. 28

F.2 Equal-Length Runtime Control

At matched output length, LatentGRM-8B and Rubric-RM-8B have nearly equal total inference time, isolating the serving cost of latent reasoning without accuracy predictions.

F.2 Equal-Length Runtime Control

  • Each model produces exactly 81,400 positions, with both samplers constrained to 50 output positions per sample.The latent request uses an unreachable latent-end ID to retain soft feedback throughout the measured trajectory.Source: §F.2 · p. 28
  • 46.469 seconds for Rubric-RM-8B versus 45.714 seconds for LatentGRM-8B yields only a 1.017× runtime difference at equal length.The control uses 50 generated positions per sample on the full RewardBench Chat and Chat Hard subsets at vote@1.Source: §F.2 · p. 28, Table 12 · p. 28
  • The near-equal runtime indicates that latent reasoning adds no additional per-step computational cost when output length is matched.Forced latent generation never enters visible-answer decoding, so this control measures serving cost without producing accuracy predictions.Source: §F.2 · p. 28

G Supplementary RQ3: Recovering Rubric-Level Evaluations

The supplementary readout evaluates how well Latent Trace Interpreter reconstructs rubric-level states and justifications from teacher-derived trajectories. It reports strong state recovery alongside qualitative examples of both successful reconstruction and a localized error.

G.1 Readout Data and Splits

  • The readout test split contains 3,298 records, and evaluation covers all 607 test records with at least one explicit criterion state.Those records contain 9,992 criteria, of which 548 records are fully labeled and 59 partially labeled.Source: §G.1 · p. 28

G.2 Readout Controls and Evaluation Protocol

  • Three independently trained readouts compare reconstruction from the prompt alone, LTI inputs, and replayed latent hidden states.The controls use the frozen Stage 1 decoder with separate LoRA adapters; Replay-hidden bridges extracted states into the readout input space.Source: §G.2 · p. 29

G.3 Metrics and Parsing

  • Criterion-state scoring uses explicit labels, counts missing or unparseable predictions as errors, and gives MET, NOT MET, and PARTIAL equal weight.Unlabeled reference criteria are excluded from state scoring.Source: §G.3 · p. 29
  • Criterion Vector Exact scores only fully labeled records, while semantic justification similarity covers 9,992 criteria and ROUGE-L covers all 607 complete traces.Partially labeled records contribute only to criterion-level metrics; ROUGE-L measures whole-trace lexical overlap, including shared templates.Source: §G.3 · pp. 29–30
  • LTI exceeds 98% F1 for every criterion state, with its largest separation from Replay-hidden on PARTIAL (98.46 versus 34.36).The comparison indicates stronger preservation of distinctions among criterion judgments by LTI.Source: §G.3 · p. 30, Table 13 · p. 30

G.4 Qualitative Reconstruction Examples

  • In record 14270, LTI recovers all ten criterion states and the comparative rationale, paraphrasing decisive teacher-trace wording.The teacher and reconstruction examples are shown in Figures 8 and 9.Source: §G.4 · p. 30, Figure 8 · p. 31, Figure 9 · p. 31

G.4.1 A Reconstruction Failure

  • In record 17234, reconstruction gets one of 16 explicit states wrong while preserving the other states and the final preference for Response A.The error marks Response B’s Criterion 3 as MET despite recognizing elsewhere that its permutation formula is applied incorrectly; Figures 10 and 11 show the reference and reconstruction.Source: §G.4.1 · p. 32, Figure 10 · p. 33, Figure 11 · p. 34

H Training Setup and Compute

LatentGRM is trained as separate 4B and 8B models on OpenRubrics data using a staged, LoRA-based setup. The appendix also provides prompt templates for rubric construction, explicit evaluation traces, and LatentGRM training and inference.

H Training Setup and Compute

  • Separate Qwen3-4B and Qwen3-8B LatentGRM models are trained on the same 35,612 OpenRubrics records using eight RTX 6000D GPUs.Training uses bfloat16, DeepSpeed ZeRO Stage 2, gradient checkpointing, and LoRA.Source: H Training Setup and Compute · p. 35, Table 14 · p. 35
  • Stage 1 and Stage 2 use different attention implementations, while deterministic tokenization and Semantic Chunking boundaries are cached across epochs.Stage 1 uses SDPA; Stage 2 uses FlashAttention-2. Only the uniformly sampled Stage 1 supervision frontier remains dynamic across epochs.Source: H Training Setup and Compute · p. 35

I Prompt Templates

  • The prompt-template appendix includes prompts for constructing rubric criteria, eliciting structured explicit evaluation traces, and LatentGRM training and inference.The templates are identified in Figures 12–14.Source: Figure 12 · p. 36, Figure 13 · p. 37, Figure 14 · p. 38

Related papers

Loading 2610.09788v2…