PaperlayerView exact revision on arXiv

Source-linked AI summary

Compressing Onboard Language Models for Space Robots under a Behavioral-Fidelity Requirement

Jia Deng, Zizhe Wang, Xin Lu, Yi Huang, Zehao Jia

arXiv:2610.29302v1cs.ROcs.LG

TL;DR

Onboard space robots need compressed language models that fit a fixed format and budget, remain static, and behave like the ground-validated reference. QSelect composes a low-bit model from compatible quantizer checkpoints, scoring substitutions at the final next-token distribution. Composed models outperform single-family baselines on held-out perplexity across the reported settings, while evaluation remains limited to general-domain benchmarks on a ground GPU.

  • Problem

    Onboard language models for space robots must meet fixed memory and storage-format requirements, remain static, and stay behaviorally close to the ground-validated reference.

    Source: §I · p. 1

  • Method

    QSelect composes a static low-bit model from compatible quantizer-family checkpoints and scores each substitution at the final next-token distribution.

    Source: §I · pp. 1–2

  • Results

    Composed models achieve lower mean held-out perplexity than every single-family model in all six tested settings, and fidelity mode lowers perplexity and KL further.

    Source: §I · p. 2

  • Takeaways & Limitations

    The reported results support composing quantizer families and using final-output fidelity scoring for the evaluated models and benchmarks.

    Source: §I · pp. 1–2

    The evaluation uses general-domain benchmarks on a ground GPU and does not test space-robot tasks, flight hardware, latency, energy, or radiation effects.

    Source: §VIII · p. 10

Abstract

from arXiv · show

Language models are being explored as the interface through which space robots interpret operator commands and select among verified skills. Running them onboard calls for low-bit compression, and flight use adds requirements that generic post-training quantization leaves implicit: a fixed memory budget and storage format, a static artifact under configuration control, and behavior close to the full-precision model validated on the ground. We treat compression as a ground-segment step and study QSelect, which composes one static low-bit model from compatible AWQ, GPTQ, and AutoRound checkpoints by assigning each linear module to one quantizer family and scoring every substitution at the final next-token distribution. The default fidelity mode needs no labels or tuned thresholds and accepts a substitution only when the forward KL divergence from the BF16 reference decreases; an optional target-adapted mode adds answer likelihood on a small labeled set. On DeepSeek-R1-Distill-Qwen-7B, Llama-3.1-8B, and Llama-3.2-3B at 3 and 4 bits, composed models reach lower mean held-out perplexity than every single-family model in all six settings, and the fidelity mode lowers perplexity and KL further. In a matched scaffold, final-output scoring gives lower held-out KL than representation-similarity scoring in all 12 paired runs. With labels from training or development splits of the evaluated task families, the target-adapted mode raises 3-bit pooled accuracy on the two Llama models by 4.3 and 4.5 points over the best single-family model. At 3 bits, unquantized embedding and output layers make up 43-46% of the packed artifact. All evaluations use general-domain benchmarks on a ground GPU; space-domain tasks and flight hardware are left to future work.

I. INTRODUCTION

Space robots can use language models to map operator requests to verified skills, but onboard compute is scarce and flight use requires compressed models to preserve ground-validated behavior. This paper introduces QSelect, a ground-segment method for composing compatible quantizer checkpoints, and evaluates its fidelity and deployment trade-offs.

A. Language Models in Space Robotics

  • Language and vision–language models can map operator requests to pre-verified skills or directly to robot actions, and space-specific agent frameworks apply this interface to on-orbit servicing,,, [5].The cited systems are generally evaluated with large models on terrestrial computers; SpaceMind uses Qwen3-VL-235B-A22B [5].Source: §I · p. 1, A. Language Models in Space Robotics · p. 2

I. INTRODUCTION

  • QSelect addresses the gap between large terrestrial models and scarce onboard compute; compression is a prerequisite for onboard language models.The introduction notes ASTREA ran a 1.54-billion-parameter Qwen2.5 model quantized to 4 bits on a flight-qualified CPU-only system, with responses taking about 40 seconds to over 8 minutes.Source: §I · p. 1
  • Flight use requires a fixed memory budget and supported format, a static artifact under configuration control, and behavior close to the ground-validated full-precision reference.The paper treats compression as a ground-segment preparation step and measures behavioral departure at the model’s output distribution.Source: §I · p. 1
  • QSelect composes a static low-bit model from compatible homogeneous checkpoints, assigning each linear module to a quantizer family and scoring substitutions at the final next-token distribution.Its default fidelity mode accepts only substitutions that lower forward KL from the BF16 reference, without labels or tuned thresholds; an optional mode uses target-answer likelihood.Source: §I · pp. 1–2
  • Across three open models from 3B to 8B parameters at 3 and 4 bits, composed models beat every single-family model on mean held-out perplexity in all six settings.The fidelity mode lowers perplexity and KL further; final-output scoring lowers held-out KL than representation-similarity scoring in all 12 paired runs.Source: §I · p. 2
  • The paper contributes a requirements analysis, QSelect’s threshold-free fidelity and optional target-adapted modes, and an empirical study spanning three open models and multiple selection baselines.The study also examines ground-time/fidelity trade-offs and packed artifacts for onboard deployment.Source: §I · p. 2

B. Onboard Computing and Flight Software Assurance

The section connects onboard deployment constraints to model-compression choices, then positions QSelect against quantization, heterogeneous assignment, and output-aware calibration. It frames a flight-ready result as a static artifact with a fixed format and reference-aligned behavior.

C. Post-Training Quantization of LLMs

D. Heterogeneous Assignment

  • QSelect and CALM both assemble models sequentially, but score candidates differently: final next-token distributions for QSelect versus intermediate representations for CALM [15],.QSelect supports block, component, and individual-linear-module decision units; the CALM baseline is an independent matched-pool reimplementation.Source: D. Heterogeneous Assignment · pp. 2–3

E. Output-Aware Calibration and Modular Composition

  • QSelect uses final-output KL for discrete assignment, unlike local weight-, layer-, or block-reconstruction criteria, and combines compatible modules without learning a stitching layer.Its output-aware assignment is distinct from distillation, which transfers output-distribution information [35].Source: E. Output-Aware Calibration and… · p. 3

A. Deployment Workflow

  • The deployment workflow validates a BF16 reference on the ground, composes and packs a static model, then runs it onboard within the autonomy stack’s checks and fallback logic.Conventional controllers retain authority over motion.Source: A. Deployment Workflow · p. 3

B. Requirements

  • A fixed bit width, group size, and symmetric packed format hold quantized-layer footprint constant, while selecting quantizer families avoids the extra kernels and layouts of mixed-precision assignment,.AWQ modules also carry a static per-channel input scale.Source: B. Requirements · p. 3
  • The onboard network must be a static, versionable artifact with each module assigned one candidate and no runtime selector, routing, or ensemble.The assignment vector and checkpoint and manifest hashes identify the artifact for versioning and rollback.Source: B. Requirements · p. 3
  • Forward KL from the reference next-token distribution measures behavioral departure, weighting loss of outputs the reference considers likely.An average KL of K nats bounds mean total-variation distance on reference text by K/2, but does not guarantee free-running behavior.Source: B. Requirements · pp. 3–4
  • The five deployment requirements cover fixed memory and format, static configuration control, reference fidelity, scarce labels, and bounded one-time ground compute.Fidelity mode uses unlabeled text; target-adapted mode can use a small labeled set, while granularity and pass count control ground cost.Source: B. Requirements · p. 4

C. Problem Statement

QSelect chooses one compatible quantizer-family candidate for each quantized linear module while keeping bit width and group size fixed. It uses a greedy search because exhaustive module-wise assignment is impractical.

C. Problem Statement

  • For Llama-3.1-8B across three families, exhaustive module-wise assignment has 3^224 possibilities, so QSelect uses a greedy search whose cost grows linearly with decision units and candidates.The fixed bit width and group size make the memory budget constant across candidate assignments.Source: C. Problem Statement · p. 4

IV. QSELECT

  • QSelect scores temporary substitutions in the partially assigned model, commits accepted updates under a fixed search protocol, and exports a static assignment.The candidate bank is precomputed and matched in format.Source: §IV · p. 4

A. Candidate Bank and Assignment Space

  • Each module’s candidate is drawn from a family-conditioned homogeneous quantized checkpoint, with calibration-bank replicate retained in the candidate notation.During assignment, candidates are stored as dense effective fake-quantized weights rather than treated as independent functions of a module’s original weights.Source: A. Candidate Bank and Assignment Space · p. 4
  • A family is supported only when its transformed parameters and static state form a locally swappable unit with a compatible input–output interface.Global rotations, cross-layer transformations, activation quantization, and incompatible activation formats fall outside this scope.Source: A. Candidate Bank and Assignment Space · p. 4
  • Assignment units constrain which families may be selected together: singleton units are module-wise, decoder-block units are layer-wise, and projection groups enable component-wise assignment.The assignment space requires modules within each decision unit to share a family.Source: A. Candidate Bank and Assignment Space · p. 4
  • In one-pass QSelect-KL, each unit tests legal families and commits the lowest-scoring choice only when it strictly lowers the current score.The returned model is identified by its static assignment.Source: A. Candidate Bank and Assignment Space · p. 5

B. Final-Output Fidelity Score

The fidelity score compares the reference and quantized model’s next-token distributions across non-padding positions. It uses the full vocabulary at temperature one.

B. Final-Output Fidelity Score

  • The score averages forward KL from the reference next-token distribution to the quantized model’s distribution over non-padding positions.It uses the full vocabulary at temperature one; reference logits are cached, while later evaluations run only the candidate model.Source: B. Final-Output Fidelity Score · p. 5

C. Fidelity Mode

QSelect-KL accepts a candidate family only when it strictly lowers fidelity KL, without labels or tuned thresholds. QSelect-TA adds a target-answer likelihood signal for settings with a small labeled set.

C. Fidelity Mode

  • QSelect-KL proposes the family with the lowest score at each decision unit and commits it only when the score strictly decreases.This is greedy block-coordinate descent; ties follow AWQ, GPTQ, AutoRound order.Source: C. Fidelity Mode · p. 5

D. Target-Adapted Mode

  • QSelect-TA uses token-averaged target-answer negative log-likelihood when prompts have known answers.The signal is teacher-forced answer likelihood, not free-running generation or accuracy.Source: D. Target-Adapted Mode · p. 5
  • At each unit, QSelect-TA shortlists near-best-KL candidates and proposes the candidate with the lowest target-answer loss.The proposal can be accepted through either a fidelity-improvement branch or a target-adaptation branch.Source: D. Target-Adapted Mode · p. 5
  • The 1% shortlist and fixed KL and NLL margins bound each accepted step’s movement away from the reference.The margins were fixed before comparisons, not derived theoretically or exhaustively swept.Source: D. Target-Adapted Mode · p. 5

E. Search Protocol

The search uses a one-pass, ordered module scan, with protocol variables evaluated separately. Ground cost is incurred before deployment; export produces a static, configuration-recorded artifact, and experiments use general-domain benchmarks on one ground GPU.

E. Search Protocol

  • The main protocol starts from homogeneous AutoRound and scans decoder blocks shallow-to-deep, processing seven projection types in a fixed order during one pass.Initialization, granularity, traversal order, and pass budget are protocol variables.Source: E. Search Protocol · p. 5

F. Ground Cost and Static Export

  • With fixed width and no prefix or suffix caching, direct module-wise assignment grows approximately quadratically with model depth, and all selection cost is incurred on the ground.The cost expression also depends on candidate and shortlist counts and full-model pass cost; candidate generation and candidate-bank storage are separate costs.Source: F. Ground Cost and Static Export · p. 5
  • Native export packs the selected family-specific modules into a static model without changing their assignments or using a dynamic selector.The assignment vector and candidate-checkpoint and calibration-manifest hashes form the artifact’s configuration record.Source: F. Ground Cost and Static Export · p. 5

V. EXPERIMENTAL SETUP

  • The evaluation uses general-domain intrinsic and downstream benchmarks on one NVIDIA A800-SXM4-80GB GPU, with no space-domain data.Downstream tasks are zero-shot; multiple-choice scores use answer-letter log probability, while GSM8K and MATH-500 use greedy generation and final-answer matching.Source: §V · p. 6
  • Experiments cover three open models at 3- and 4-bit weights, while embeddings, normalization layers, and output heads remain BF16.The tested formats use BF16 activations and 128-channel input groups; packed artifacts are examined separately.Source: §V · p. 6
  • Candidate quantizers share a C4 manifest but do not use matched compute or identical effective token budgets.The study’s generator settings are not claims about each method’s strongest published configuration; the labeled proxy and held-out manifests are fixed.Source: §V · p. 6
  • Most protocol studies use QSelect-TA, whereas QSelect-KL is evaluated with the main protocol in all six settings.For four 7B/8B settings, QSelect-KL absolute Table II values are derived from paired changes; its selection time is bounded by QSelect-TA under the same protocol.Source: §V · p. 6

VI. RESULTS

Across the tested settings, composed models improve on single-family quantization, and final-output scoring outperforms representation-similarity scoring on held-out KL in a matched scaffold. Results also show trade-offs across fidelity, task adaptation, and selection capacity.

A. Composed Versus Single-Family Models

  • At W3, QSelect-TA lowers mean Avg. PPL versus the best single-family model by 7.4%, 5.8%, and 14.6% across the three models, while raising pooled accuracy on the two Llama models by 4.3 and 4.5 points.The task gains use labeled examples from training or development splits of five of six evaluated Llama task families; DS task differences depend on the calibration bank.Source: A. Composed Versus Single-Family Models · p. 6, Table II · p. 7
  • QSelect-KL lowers mean Avg. PPL and Avg. KL versus QSelect-TA in all six settings, yielding the lowest mean Avg. PPL among Table II configurations in every setting.It has the lowest Avg. KL in five settings; derived DS W4 KL is 0.0913, comparable to homogeneous GPTQ at 0.092. Some Table II QSelect-KL values are derived and lack standard deviations.Source: A. Composed Versus Single-Family Models · pp. 6–7, Table III · p. 7
  • At W3, fidelity mode exceeds the best single-family Micro accuracy by 2.9 points on Llama-3.2-3B and 3.4 derived points on Llama-3.1-8B.Removing answer supervision costs task accuracy mainly on Llama-3.2 W3; at W4, Micro changes by −0.16 points and Macro by +0.23 points.Source: A. Composed Versus Single-Family Models · pp. 6–7

B. Why Score the Final Output?

  • Final-distribution KL scoring lowers held-out Avg. KL in all twelve paired runs versus CKA in the same sequential layer-wise scaffold.PPL and task accuracy also improve on DS W3 and both Llama settings, while DS W4 trades lower KL for slightly higher PPL and lower task means.Source: B. Why Score the Final Output? · p. 7, Table IV · p. 8
  • Local-MSE ends with higher Avg. KL than fidelity mode in all four tested settings, higher Avg. PPL in three, and lower Micro accuracy in three.It is a quick reference based on static reference activations, not a contextual search; it takes about 0.8 minutes on DS settings.Source: B. Why Score the Final Output? · pp. 7–8, Table III · p. 7
  • In five settings with replicate agreement, module-wise QSelect-TA lowers Avg. PPL by 1.22–14.62% and raises Micro and Macro over the selected whole-model family.The comparison does not match module-wise search capacity or path dependence.Source: B. Why Score the Final Output? · p. 8, Table V · p. 8

C. Ground Time Versus Fidelity

Ground selection trades search time against fidelity, and its results depend on initialization and candidate-bank composition. Packed artifacts retain small metric drift, but unquantized embeddings and output layers remain a substantial share of their size.

C. Ground Time Versus Fidelity

  • At DS W3, module scans take roughly 210 minutes versus 18 minutes for layers, yet modules give the lowest W3 PPL; finer units are not uniformly better.The figure places evaluated selectors on a common ground-time axis; module-wise selection takes longer than layer-wise selection.Source: C. Ground Time Versus Fidelity · p. 8, Figure 2 · p. 9

D. Robustness of the Ground Procedure

  • Changing the Llama-3.1 W3 start from GPTQ to AWQ increases Avg. PPL by 1.18 and lowers Micro and Macro by 5.49 and 5.14 points.Traversal order has a smaller aggregate effect than initialization, while adding AutoRound most clearly helps at W3.Source: D. Robustness of the Ground Procedure · p. 8
  • Across DS W3 calibration banks, assignment agreement is 73.5–76.0%, with 124 of 196 modules agreeing across all three; selection depends on the candidate bank.Selected Llama-3.2-3B models also remain heterogeneous across AWQ, GPTQ, and AutoRound assignments.Source: D. Robustness of the Ground Procedure · p. 8

A. What Occupies the Packed Artifact

  • Packed exports change held-out metrics only slightly versus fake-quantized models: |∆PPL| ≤0.277% and |∆KL| ≤1.314%.Table VII reports packed exports for calibration-replicate-17 assignments; its drift measures packed-versus-fake-quantized change.Source: A. What Occupies the Packed Artifact · pp. 8–9, Table VII · p. 9
  • Unquantized BF16 embeddings and output layers account for 43–46% of the 3-bit artifact and 37–39% of the 4-bit artifact.Further memory reductions would need to address these layers, for example through embedding/head quantization or vocabulary restriction, neither evaluated here.Source: A. What Occupies the Packed Artifact · p. 9

B. Runtime Memory

The packed artifact shrinks substantially, but peak CUDA memory falls less; the operating-point discussion favors fidelity when labels are scarce and highlights unresolved gaps between ground evaluation and flight use.

B. Runtime Memory

  • 24.2–27.9%: peak CUDA memory falls by this range although the serialized artifact shrinks by roughly a factor of three.The 7B/8B artifacts still occupy about 4.8 GB at 3 bits and 5.6–5.7 GB at 4 bits; the reported measurement uses batch one, sequence length 2,048, no key–value cache, and a packed forward path that dequantizes weights.Source: B. Runtime Memory · pp. 9–10

C. Choosing an Operating Point

  • At W4, QSelect-TA stays within 0.81, 2.34, and 2.37 Micro points of BF16 on DS-R1-Qwen-7B, Llama-3.1-8B, and Llama-3.2-3B, respectively.At W3, the corresponding gaps are 7.87, 11.87, and 10.43 points; the passage calls W4 the realistic default for command interpretation and W3 suitable only when memory binds and task-level acceptance tests quantify the loss.Source: C. Choosing an Operating Point · p. 10
  • QSelect-KL needs no labels and gives the lowest Avg. PPL in every setting and lowest Avg. KL in five of six configurations described in Table II.QSelect-TA is appropriate when a mission supplies labeled examples and acceptance tests cover the resulting departure from the reference; at Llama-3.2 W3, QSelect-KL lowers Avg. KL by 10.1% relative to QSelect-TA but loses 1.57 Micro points.Source: C. Choosing an Operating Point · p. 10
  • Only 124 of 196 module choices agree across the three DS W3 calibration-bank runs, and regenerated candidates change weights even when a module keeps its family.The passage recommends treating an update as a new artifact requiring full acceptance testing; configuration records allow recovery of the previous artifact for rollback.Source: C. Choosing an Operating Point · p. 10

D. What Flight Use Would Add

  • Flight use still requires mission-text validation, flight-representative measurements of latency, energy, and thermal load, and study of radiation-upset propagation through packed metadata.A corrupted group scale or zero point affects all 128 weights in its group, while a corrupted AWQ input scale affects a whole input channel; neither issue is answered by these experiments.Source: D. What Flight Use Would Add · p. 10

VIII. LIMITATIONS

  • The evaluation uses general-domain benchmarks on a ground GPU, not space-robot commands, procedures, skill libraries, flight hardware, latency, energy, or radiation effects.The work therefore does not establish performance on space-domain data or flight hardware.Source: §VIII · p. 10
  • The evidence covers fixed-bit symmetric weight-only candidates with group size 128 for 3B–8B text-only models, not several broader model and quantization settings.Undemonstrated settings include activation quantization, vision–language models, incompatible runtime formats, and models larger than 8B; both Llama references are base models, whereas a command interface would start from an instruction-tuned reference.Source: §VIII · p. 10
  • The reported metrics and search leave uncertainties: task labels come from evaluated task families, three calibration-bank replicates do not measure sampling uncertainty, and one-pass assignments are not converged solutions.Assignments also depend on initialization, order, margins, and candidate bank; exact-scoring affordability rests on the studied 3B–8B sizes, while larger models or frequent re-preparation may need caching or a screening step.Source: §VIII · pp. 10–11

IX. CONCLUSION

The conclusion frames onboard compression as a fidelity and artifact-control problem, not just a footprint problem. QSelect composes low-bit models and reports improved benchmark results, while mission-text and flight-hardware evaluation remain future work.

IX. CONCLUSION

  • QSelect’s label-free fidelity mode achieves lower mean held-out perplexity than every single-family model and target-adapted mode in all six settings.Final-output scoring gives lower held-out KL than representation-similarity and local-reconstruction criteria; decision granularity trades ground time against fidelity, and packed artifacts are 63–70% smaller than BF16.Source: §IX · p. 11
  • At 3 bits, unquantized embedding and output layers account for 43–46% of the packed artifact.The conclusion identifies evaluating the fidelity criterion on mission text and measuring composed models on flight-representative hardware as next steps.Source: §IX · p. 11

APPENDIX A TARGET-ADAPTED ALGORITHM

The appendix details QSelect’s target-adapted procedure, evaluation protocol, and controls. Its analyses report task-level trade-offs and show that objective behavior and initialization affect comparisons.

APPENDIX A TARGET-ADAPTED ALGORITHM

  • QSelect-TA caches KL and NLL scores, lazily recomputes invalidated NLL, and does not try a runner-up when the minimum-NLL proposal fails both acceptance branches.The procedure is one-pass.Source: Appendix A · p. 13

APPENDIX B PROTOCOL DETAILS

  • The proxy builder uses fixed dataset rows without per-run resampling, while calibration-bank replicates determine candidate-generation and KL-scoring C4 samples.The manifests and candidate checkpoint hashes are recorded in run metadata.Source: Appendix B · p. 13
  • The custom downstream protocol uses zero-shot prompts and task-specific scoring, while local Code, Math, and Wiki corpora are robustness sets rather than standard public benchmarks.Upstream revisions, license metadata, file counts, language composition, and complete deduplication provenance were not preserved; model and tokenizer commit revisions are also not pinned.Source: Appendix B · p. 13

A. Per-Task Results

  • Selected per-task rows expose task-level trade-offs, but do not compare all candidate families; for Llama, the proxy-unseen group contains only GPQA.Table VIII reports selected per-task downstream accuracy, with Micro pooling examples and Macro averaging tasks.Source: A. Per-Task Results · p. 13

B. Matched Layer-Wise Scoring

  • The matched layer-wise scoring comparison reports final-distribution KL variants alongside CALM-reimpl. results; the guarded selector accepts changes only under a 0.1% margin.Selection time excludes candidate generation.Source: B. Matched Layer-Wise Scoring · pp. 13–14

C. Objective Ablations

  • NLL-only improves both task aggregates only on DS W3, but degrades PPL and KL in all four evaluated 7B/8B settings.Answer supervision therefore shifts the operating point rather than uniformly improving it.Source: C. Objective Ablations · p. 14

D. Paired DS-R1-Qwen-7B W3 Analysis

  • Replicate-specific assignments agree on 73.5–76.0% of modules, with 124/196 modules agreeing across all three replicates; the observed association does not establish which replacements cause the task decrease.The replicate-41 Micro decrease is driven primarily by a 2.91-point MMLU decrease relative to AutoRound, accounting for approximately 2.23 points of the total Micro change.Source: D. Paired DS-R1-Qwen-7B W3 Analysis · p. 14

E. Search Budget and Proxy Control

  • Table XIII compares search-budget and proxy controls with the main result, including three-pass runs and a shared-core Llama-3.1-8B control using the DS proxy subset.Each three-pass row uses exactly three full scans.Source: E. Search Budget and Proxy Control · p. 14

F. Initialization Sensitivity

The appendix examines sensitivity to initialization, candidate-pool composition, and traversal order. It cautions that pool comparisons depend on initialization and specifies the paired traversal-order comparison.

F. Initialization Sensitivity

  • Different homogeneous starting assignments retain different family majorities, and no start is uniformly best across fidelity and task metrics.Table XIV reports means and sample standard deviations over three calibration-bank replicates.Source: F. Initialization Sensitivity · p. 14, Table XIV · p. 16

G. Candidate Pool × Initialization

  • The restricted AWQ/GPTQ pool completes a 2 × 2 pool-by-initialization control, and the large Llama-3.1 W3 gap between starts prevents interpreting a restricted-pool result as a pure source-removal effect.Table XV uses the same three calibration-bank replicates as Table II.Source: G. Candidate Pool × Initialization · p. 14, Table XV · p. 16

H. Traversal-Order Sensitivity

  • The traversal-order control compares the fixed forward protocol with a fixed random permutation and the exact reverse, pairing each row to the same replicate’s forward result.Deltas are variant minus forward; negative PPL/KL and positive accuracy favor the alternative order.Source: H. Traversal-Order Sensitivity · p. 14, Table XVI · p. 16

Related papers

Loading 2610.29302v1…