Oracle-Guided Trajectory Distillation
- The paper introduces oracle-guided trajectory distillation as a training-time supervision principle that leverages privileged trajectory signals to build stronger teacher guidance.
- It refines latent alignments, ODE states, reasoning prefixes, and planning candidates using target-conditioned CTC, future-aware cues, and gradient-based updates for smoother student training.
- Empirical results across speech recognition, image synthesis, reasoning, and autonomous driving demonstrate reduced error rates, higher BLEU scores, and improved planning metrics.
to=arxiv_search.17query17^ qq的天天中彩票 omitted JSON: {"17query17 trajectory distillation17\17 OR 17\17 distillation17\17 oracle guidance17", "17max_results17 17\17query17, "17sort_by17 to=arxiv_search.17query17^ 福利彩票天天 ոմitted anasiyana เติมเงินไทยฟรี JSON: {"17query17 OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17", "17max_results17 17\17query17, "17sort_by17 to=arxiv_search.17query17^ 久久免费热在线精品 ոմitted 娱乐彩票注册 JSON: {"17query17 Distribution Matching Distillation with Gradient-Based Reinforcement Learning17\17 OR ti:17\17 Teacher17\17 OR ti:17\17 OR ti:17\17 Distillation17\17 "17max_results17 17\17query17, "17sort_by17 Oracle-guided trajectory distillation denotes a family of distillation procedures in which the supervisory signal is derived from a privileged, strengthened, or corrected view of a trajectory rather than from a teacher’s final output alone. In the cited literature, the privileged view may be a target-conditioned CTC alignment, target-derived17 oracle guidance17^ for seq17 OR \17seq learning, an authentic diffusion or flow trajectory, a future-aware planning trajectory, a partially privileged reasoning trace, a teacher-refined rollout, or a gradient-implied distillation target (&&&17\17&&&, &&&17relevance17&&&, &&&17 OR \17&&&, &&&17 oracle guidance17&&&). The shared structure is that the student is trained from trajectory-level information that is unavailable to the deployed student at inference time, with the goal of improving supervision quality, stabilizing optimization, or selecting more learnable training trajectories.
17\17. Conceptual scope and defining characteristics
Oracle-guided trajectory distillation is not a single algorithmic template. Across the literature, it refers to methods that strengthen distillation by giving the teacher, evaluator, or selector access to information that the student does not have at deployment. The “oracle” may be the ground-truth target sequence in CTC and seq17 OR \17seq learning, future observations in autonomous driving, full reasoning traces in privileged on-policy distillation, or a reward model that scores a gradient-induced target rather than a raw generated sample. The “trajectory” may be an explicit latent alignment PRESERVED_PLACEHOLDER_17query17, a generation ODE state PRESERVED_PLACEHOLDER_17\17, a reasoning rollout PRESERVED_PLACEHOLDER_17 OR \17, a planned waypoint sequence PRESERVED_PLACEHOLDER_17 oracle guidance17, or a distillation update direction encoded as an implicit target tensor (&&&17\17&&&, &&&17sort_by17&&&, &&&17query17&&&).
A recurring property is asymmetry between training and inference. Oracle Teacher conditions on both source input and target labels during teacher training, but the student later predicts from the source input alone (&&&17\17&&&). EM-Network trains an augmented model with17 oracle guidance17^ derived from the target sequence, then removes the oracle encoder and fusion module at inference (&&&17relevance17&&&). EvoDriveVLA’s oracle teacher uses future images and future ego states, while the student is evaluated without those inputs (&&&17sort_by17&&&). AR-OPD and TRD similarly use privileged traces or reference solutions to construct better supervision during training, not to expand the student’s inference-time input (&&&17 oracle guidance17&&&, &&&17query17&&&).
This suggests that oracle-guided trajectory distillation is best understood as a training-time supervision design principle rather than a model class. Its central question is how to exploit privileged trajectory information without forcing the student to imitate a target distribution that is unreachable, unstable, or trivial.
17 OR \17. Sources of oracle information and the trajectories they supervise
Different papers instantiate the oracle in different ways, but each ties the oracle to a trajectory-like object rather than to a terminal label alone.
| Method | Oracle signal | Distilled trajectory object |
|---|---|---|
| Oracle Teacher (&&&17\17&&&) | target label sequence PRESERVED_PLACEHOLDER_17max_results17^ during teacher training | CTC alignment distribution PRESERVED_PLACEHOLDER_17sort_by17^ |
| EM-Network (&&&17relevance17&&&) | 17 oracle guidance17^ PRESERVED_PLACEHOLDER_17relevance17^ derived from the target; masked target PRESERVED_PLACEHOLDER_17query17^ for AED | latent alignments PRESERVED_PLACEHOLDER_17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17^ or seq17 OR \17seq prediction trajectory |
| GDMD (&&&17query17&&&) | reward model on decoded gradient-implied target PRESERVED_PLACEHOLDER_17max_results17^ | distillation gradient / update trajectory |
| FlowSteer (&&&17 OR \17&&&) | teacher’s authentic ODE path via OTA | few-step latent ODE trajectory |
| AR-OPD (&&&17 oracle guidance17&&&) | partial and full privileged traces PRESERVED_PLACEHOLDER_17\17query17^ and PRESERVED_PLACEHOLDER_17\17\17^ | reasoning-token distributions on student-visited states |
| TRD (&&&17query17&&&) | teacher-guided refinement of rollout PRESERVED_PLACEHOLDER_17\17 OR \17; in OPSD also PRESERVED_PLACEHOLDER_17\17 oracle guidance17^ | corrected reasoning trajectory PRESERVED_PLACEHOLDER_17\17max_results17^ |
| LARK (&&&17max_results17&&&) | oracle learnability objective PRESERVED_PLACEHOLDER_17\17sort_by17^ | selected teacher-generated reasoning trajectories |
| EvoDriveVLA (&&&17sort_by17&&&) | future images and future ego states | planning trajectories PRESERVED_PLACEHOLDER_17\17relevance17, PRESERVED_PLACEHOLDER_17\17query17, and sampled candidates |
In sequence learning, the oracle most often comes from target-side information. Oracle Teacher changes the teacher’s conditional model from PRESERVED_PLACEHOLDER_17\17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17^ to PRESERVED_PLACEHOLDER_17\17max_results17, so the teacher can learn a more accurate CTC alignment by referring to the target information (&&&17\17&&&). EM-Network likewise introduces an oracle encoder that maps target-side information into17 oracle guidance17^ PRESERVED_PLACEHOLDER_17 OR \17query17, then fuses it with source representations through cross-attention; for AED, the target is masked to avoid trivial copying (&&&17relevance17&&&).
In generative modeling, the oracle is often trajectory authenticity or trajectory-aware reward. FlowSteer treats the teacher’s authentic ODE path as the oracle and replaces interpolated off-trajectory stage starts with on-trajectory states produced by ODESolve (&&&17 OR \17&&&). GDMD does not ask a reward model to score the raw output image PRESERVED_PLACEHOLDER_17 OR \17\17; instead, it asks the reward model to evaluate the decoded implicit target induced by the DMD gradient, thereby turning sample-based scoring into gradient-based scoring (&&&17query17&&&).
In reasoning and planning, the oracle frequently takes the form of privileged future context. AR-OPD defines a full privileged teacher PRESERVED_PLACEHOLDER_17 OR \17 OR \17^ and a partial privileged teacher PRESERVED_PLACEHOLDER_17 OR \17 oracle guidance17, using the former only as a residual correction on top of the latter (&&&17 oracle guidance17&&&). TRD uses the teacher to revise the student’s rollout into a refined trajectory before distillation (&&&17query17&&&). EvoDriveVLA uses a future-aware oracle teacher with future scene images and future ego states to generate and refine trajectory candidates for autonomous driving (&&&17sort_by17&&&).
17 oracle guidance17. Trajectory construction, refinement, and correction
A first line of work constructs better latent trajectories by conditioning the teacher on the target. In Oracle Teacher, the architecture consists of SourceNet, Encoder, and Decoder:
PRESERVED_PLACEHOLDER_17 OR \17max_results17^
The decoder is non-autoregressive and uses the source-side representation as 17query17^ and the target-side representation as key/value in cross-attention, so the output length remains tied to the source frame length PRESERVED_PLACEHOLDER_17 OR \17sort_by17^ (&&&17\17&&&). EM-Network follows a related logic: the oracle encoder extracts target-derived guidance, and a fusion module combines the source representation with that17 oracle guidance17; for CTC, the many-to-one alignment structure prevents the model from trivially copying the target, while for AED random masking is used for the same purpose (&&&17relevance17&&&).
A second line of work constructs better generation paths by enforcing trajectory authenticity. FlowSteer identifies that Piecewised ReFlow suffers from teacher trajectory mismatch and inter-stage distribution mismatch. Its Online Trajectory Alignment sets
PRESERVED_PLACEHOLDER_17 OR \17relevance17^
so stage boundaries lie on the teacher’s actual inference path, and it supplements segment distillation with adversarial matching of intermediate trajectory states (&&&17 OR \17&&&). GDMD similarly shifts attention from endpoint quality to update quality by defining an implicit target
PRESERVED_PLACEHOLDER_17 OR \17query17^
then decoding PRESERVED_PLACEHOLDER_17 OR \17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17^ and scoring it with a reward model (&&&17query17&&&).
A third line of work explicitly repairs or selects trajectories before distillation. TRD samples a raw student rollout PRESERVED_PLACEHOLDER_17 OR \17max_results17, asks the teacher to refine it into PRESERVED_PLACEHOLDER_17 oracle guidance17query17, and then distills on the refined trajectory rather than on the original failed prefix (&&&17query17&&&). EvoDriveVLA uses a future-aware oracle teacher to generate coarse trajectories PRESERVED_PLACEHOLDER_17 oracle guidance17\17, then refined trajectories PRESERVED_PLACEHOLDER_17 oracle guidance17 OR \17, augments the candidate pool with MC-Dropout, and selects the best candidate by minimum cross-entropy against the ground-truth trajectory (&&&17sort_by17&&&). LARK takes a selection-oriented approach: for each question, it ranks candidate reasoning trajectories by a learnability-grounded score rather than by perceived quality alone (&&&17max_results17&&&).
17max_results17. Objective design: alignment, residual guidance, and learnability
The objective design in oracle-guided trajectory distillation is often motivated by a mismatch diagnosis. Oracle Teacher derives a lower bound
PRESERVED_PLACEHOLDER_17 oracle guidance17 oracle guidance17^
which motivates distillation toward the Oracle Teacher’s alignment distribution, although the paper uses FitNets-style hidden-representation distillation in practice because direct KL is intractable (&&&17\17&&&). EM-Network makes a closely related EM-like argument, interpreting the teacher as an approximate posterior over latent trajectories and optimizing a one-stage self-distillation objective with online-updated soft labels (&&&17relevance17&&&).
GDMD is built around the claim that naive sample-based RL creates noisy rewards and gradient conflict with DMD. Its normalized reward is relative to the original sample PRESERVED_PLACEHOLDER_17 oracle guidance17max_results17,
PRESERVED_PLACEHOLDER_17 oracle guidance17sort_by17^
and enters a DiffusionNFT-style objective so that high reward strengthens the positive update and low reward strengthens the negative update (&&&17query17&&&). The paper’s interpretation is explicit: RL should act as an adaptive weight on the DMD gradient, not as a separate competing objective.
AR-OPD is built around a different mismatch, termed reachability mismatch. It defines a partial privileged context PRESERVED_PLACEHOLDER_17 oracle guidance17relevance17^ and constructs an anchored residual target
PRESERVED_PLACEHOLDER_17 oracle guidance17query17^
Here PRESERVED_PLACEHOLDER_17 oracle guidance17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17^ recovers partial privileged OPD, PRESERVED_PLACEHOLDER_17 oracle guidance17max_results17^ recovers the full privileged target, and the reported best regime is contractive residual transfer with PRESERVED_PLACEHOLDER_17max_results17query17^ (&&&17 oracle guidance17&&&). The method therefore rejects full privileged imitation as the default target.
LARK makes learnability itself the oracle objective. Its learnability factor is
PRESERVED_PLACEHOLDER_17max_results17\17^
which governs the rate of loss decrease under gradient flow (&&&17max_results17&&&). Because maximizing PRESERVED_PLACEHOLDER_17max_results17 OR \17^ directly leads to a degenerate one-hot solution, LARK introduces a forward-pass proxy and a PRESERVED_PLACEHOLDER_17max_results17 oracle guidance17-regularized soft top-PRESERVED_PLACEHOLDER_17max_results17max_results17^ selection rule. This makes oracle-guided trajectory distillation, in this setting, a problem of choosing which trajectories are most useful for the student’s optimization dynamics rather than which trajectories merely look best.
TRD provides a complementary perspective. Its diagnosis is prefix failure: if the student’s rollout contains a wrong reasoning prefix, dense token-level teacher supervision becomes bimodal and fragmented. The remedy is a trajectory-level correction step before KL distillation, so supervision is computed on refined prefixes PRESERVED_PLACEHOLDER_17max_results17sort_by17^ rather than on the original failed prefixes PRESERVED_PLACEHOLDER_17max_results17relevance17^ (&&&17query17&&&).
17sort_by17. Empirical patterns across domains
In CTC-based distillation, Oracle Teacher reports consistent student gains on speech recognition and scene text recognition. On LibriSpeech, a Jasper Mini student improves from 17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17.17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17sort_by17/17 OR \17max_results17.17 OR \17relevance17^ WER on test-clean/test-other without distillation to 17relevance17.17relevance17query17 OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17 OR \17^ when distilled from the Oracle Teacher, compared with 17query17.17query17 oracle guidance17/17 OR \17query17.17max_results17\17^ when distilled from Jasper DR; with LLM decoding, the Oracle Teacher still gives the best student performance, including 17max_results17.17max_results17query17 versus 17sort_by17.17query17query17 for Jasper DR distillation (&&&17\17&&&). In STR, the CRNN student improves from 17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17query17.17\17query17% total accuracy without distillation to 17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17 OR \17.17 OR \17\17% with Oracle Teacher distillation, nearly matching Star-Net at 17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17 OR \17.17 OR \17max_results17%. The same paper also reports substantially lower teacher-training cost, for example 17\17^ × 17\17 OR \17GB GPU for 17 oracle guidance17query17^ epochs in about 17 OR \17 OR \17^ hours for Oracle Teacher versus 17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17^ × 17 oracle guidance17 OR \17GB GPUs and 17max_results17query17query17^ epochs for Jasper DR.
EM-Network reports improvements in both ASR and MT. On LibriSpeech test-clean/test-other, it reaches 17max_results17.17 OR \17max_results17^ / 17\17query17.17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17\17^ with greedy decoding compared with the baseline 17max_results17.17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17query17^ / 17\17 OR \17.17query17sort_by17^, and 17 OR \17.17query17query17^ / 17query17.17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17query17^ with LM beam search compared with baseline 17 OR \17.17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17sort_by17^ / 17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17.17 oracle guidance17max_results17^ (&&&17relevance17&&&). On MT benchmarks, it reports 17 OR \17max_results17.17max_results17max_results17^ BLEU on IWSLT’17\17max_results17^ En-De and 17 oracle guidance17relevance17.17sort_by17 oracle guidance17^ BLEU on De-En, and elsewhere in the paper 17 oracle guidance17\17.17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17query17^ BLEU on IWSLT’17\17max_results17^ En-De, 17 oracle guidance17max_results17.17max_results17max_results17^ BLEU on De-En, 17 oracle guidance17\17.17 oracle guidance17query17^ BLEU on WMT’17\17max_results17^ En-De, and 17 oracle guidance17sort_by17.17max_results17query17^ BLEU on De-En.
In few-step image synthesis, GDMD reports that its 17max_results17-step / 17max_results17-NFE student outperforms DMD, DMD17 OR \17, DMDR, and in several settings even the multi-step teacher. On SDXL-Base, the reported 17max_results17-step metrics are CLIP Score: 17query17.17 OR \17max_results17\17 OR \17^, HP Score: 17query17.17 OR \17max_results17max_results17\17^, Aesthetic Score: 17sort_by17.17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17\17 OR \17query17^, Pick Score: 17 OR \17 OR \17.17sort_by17max_results17max_results17relevance17^, and ImageReward: 17query17.17max_results17\17query17\17; on SD17 oracle guidance17-Medium they are 17query17.17 OR \17max_results17 oracle guidance17query17^, 17query17.17 oracle guidance17query17query17relevance17^, 17sort_by17.17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17query17 OR \17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17^, 17 OR \17 OR \17.17max_results17relevance17\17max_results17^, and 17\17.17 OR \17query17query17 OR \17^ respectively (&&&17query17&&&). On GenEval, GDMD reaches Overall: 17query17.17query17\17, compared to 17query17.17relevance17 oracle guidance17^ for DMD and 17query17.17relevance17max_results17 for DMDR, and the user study reports 17sort_by17sort_by17.17\17 wins over the teacher for image quality and 17relevance17max_results17.17relevance17 over DMD. FlowSteer, evaluated on SD17 oracle guidance17-Medium at 17max_results17^ steps / 17max_results17^ NFE, reports PickScore: 17 OR \17 OR \17.17 oracle guidance17max_results17^, HPSv17 OR \17: 17 OR \17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17.17relevance17query17^, CLIP Score: 17 oracle guidance17 OR \17.17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17\17^, and GenEval Overall: 17query17.17relevance17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17sort_by17max_results17^, compared with PeRFlowPRESERVED_PLACEHOLDER_17max_results17query17^ at 17 OR \17 OR \17.17\17max_results17^, 17 OR \17relevance17.17 oracle guidance17relevance17^, 17 oracle guidance17 OR \17.17sort_by17sort_by17^, and 17query17.17relevance17 oracle guidance17sort_by17query17^; the gains on HPSv17 OR \17^ and GenEval are +17 OR \17.17 OR \17max_results17^ and +17query17.17query17sort_by17 (&&&17 OR \17&&&).
In reasoning distillation, AR-OPD reports average scores of 17sort_by17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17.17 OR \17^ for Base, 17relevance17 OR \17.17max_results17^ for SFT, 17relevance17sort_by17.17 OR \17^ for Partial OPD, 17relevance17query17.17max_results17 for Full OPD, and 17query17query17.17 oracle guidance17^ for AR-OPD, corresponding to a 17 OR \17.17 oracle guidance17-point improvement over full privileged OPD and 17query17.17max_results17 points over SFT (&&&17 oracle guidance17&&&). It also reports a 17 OR \17\17.17query17% relative reduction in shortcut events, from 17 OR \17 oracle guidance17^ to 17\17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17^ at the final checkpoint, and up to a 17query17.17 OR \17-point advantage on trajectories exceeding 17query17relevance17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17^ tokens. TRD reports strong gains in both OPD and OPSD settings, especially on hard math tasks: for Qwen17 oracle guidance17-17max_results17B in OPSD, AMOBench Pass@17\17relevance17^ improves from 17 OR \17 oracle guidance17.17\17^ to 17 oracle guidance17 oracle guidance17.17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17^, and for Qwen17 oracle guidance17-17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17B from 17max_results17\17.17query17 to 17relevance17\17.17sort_by17; HMMT17 OR \17sort_by17^ Pass@17\17relevance17^ improves from 17relevance17relevance17.17query17 to 17query17relevance17.17 oracle guidance17^ at 17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17B (&&&17query17&&&). LARK is reported as best in all six model-budget settings for three student models and PRESERVED_PLACEHOLDER_17max_results17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17, including 17sort_by17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17.17query17\17% average Acc@17sort_by17^ for Qwen-17 OR \17.17sort_by17- at PRESERVED_PLACEHOLDER_17max_results17max_results17^ and 17relevance17 oracle guidance17.17query17max_results17% at PRESERVED_PLACEHOLDER_17sort_by17query17, beating RSR by 17query17.17\17relevance17 and 17query17.17query17query17 points (&&&17max_results17&&&).
In autonomous driving, EvoDriveVLA reports state-of-the-art open-loop results and improved closed-loop evaluation. Against OpenDriveVLA on nuScenes, it reports 17 OR \17\17% and 17 OR \17 OR \17% improvements in L17 OR \17^ under the ST-P17 oracle guidance17^ and UniAD settings, and 17max_results17query17% and 17relevance17query17% improvements in collision rate; on NAVSIM it reaches PDMS 17(Dong et al., 21 Apr 2026) OR (Yoon et al., 2021) OR (Ke et al., 24 Nov 2025) OR (Zhang, 9 Jun 2026) OR (Yu et al., 28 May 2026) OR (Cao et al., 10 Mar 2026) OR (Yoon et al., 2023) OR (Jiang et al., 7 Jun 2026)17sort_by17.17 oracle guidance17^, improving the 17 oracle guidance17B base model by 17 oracle guidance17.17max_results17^ points, a 17max_results17.17 OR \17% gain (&&&17sort_by17&&&). The ablations show a progression from 17query17.17sort_by17sort_by17 avg L17 OR \17^ for the baseline to 17query17.17sort_by17max_results17 with trajectory KD, 17query17.17sort_by17 oracle guidance17^ with refinement, 17query17.17sort_by17 oracle guidance17^ with MC-Dropout, and 17query17.17sort_by17 OR \17^ with the full method.
Taken together, these results suggest that the empirical benefit of oracle-guided trajectory distillation is not confined to one modality. The reported gains appear in latent alignment learning, few-step image generation, reasoning post-training, and planning.
17relevance17. Failure modes, safeguards, and conceptual distinctions
A central controversy is whether17 oracle guidance17^ merely encourages shortcut learning. Several papers address this directly. Oracle Teacher identifies a trivial copy risk when the teacher sees the target, but argues that CTC’s many-to-one mapping property prevents copying because target text alone is insufficient to determine a frame-aligned trajectory PRESERVED_PLACEHOLDER_17sort_by17\17^ (&&&17\17&&&). EM-Network uses the same structural argument for CTC and introduces random masking for AED to avoid a trivial shortcut through the full target sequence (&&&17relevance17&&&).
Another recurring issue is mismatch between teacher supervision and student reachability. FlowSteer diagnoses teacher trajectory mismatch and inter-stage distribution mismatch in Piecewised ReFlow (&&&17 OR \17&&&). GDMD diagnoses gradient conflict between DMD and sample-based RL, as well as unreliable rewards from noisy early-stage generations (&&&17query17&&&). AR-OPD diagnoses reachability mismatch and hindsight leakage when the privileged teacher conditions on future information unavailable at the student’s current prefix (&&&17 oracle guidance17&&&). TRD diagnoses prefix failure, arguing that token-level clipping, truncation, or reweighting cannot fix a supervision problem caused by the trajectory itself (&&&17query17&&&). These papers differ in mechanism, but each rejects unqualified full-view imitation.
A common misconception is that oracle-guided trajectory distillation always means stronger supervision through absolute imitation. The literature does not support that simplification. AR-OPD transfers only a scaled residual beyond a partial anchor, GDMD turns reward into an adaptive modulator of the distillation update, TRD first rewrites the rollout before distillation, EvoDriveVLA selects one oracle candidate rather than averaging all candidates, and LARK selects trajectories by learnability rather than by quality alone (&&&17 oracle guidance17&&&, &&&17query17&&&, &&&17sort_by17&&&, &&&17max_results17&&&).
The remaining limitations are domain-specific. Oracle Teacher and EM-Network require paired source-target supervision during training (&&&17\17&&&, &&&17relevance17&&&). EvoDriveVLA’s oracle teacher depends on future images and future ego states that are unavailable in deployment (&&&17sort_by17&&&). LARK’s main experiments use a correctness-verified candidate pool and report standard deviations over decoding seeds rather than independent training seeds (&&&17max_results17&&&). TRD incurs extra sampling cost because it requires a teacher-guided refinement rollout in addition to the initial student rollout, and its support constraint is approximate in practice (&&&17query17&&&). These caveats delimit the operational meaning of the oracle: it is a training-time instrument for constructing better trajectory supervision, not a deployable policy input.
Oracle-guided trajectory distillation therefore occupies a distinct position within distillation research. Its defining move is to improve the teacher signal at the level where errors, ambiguities, or dead ends actually emerge: latent alignments, intermediate ODE states, reasoning prefixes, planning candidates, or gradient updates. This suggests that its unifying contribution is not merely better targets, but better target construction.