Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall

Public
Authors:Jacqueline He, Howard Yen, Shuyue Stella Li, Margaret Li, Hanqing Zeng, Yinglong Xia, Benyu Zhang, Zhuokai Zhao, Qiang Zhang, Pang Wei Koh, Luke Zettlemoyer, Wen-tau Yih
More
Copy repository link

AI research summary

Knowledge distillation with post-trained teachers is increasingly used to compress language models, yet whether its benefits remain consistent across training stages is unclear. Evaluating logit-based distillation on OLMo-2, the authors identify a stage-dependent reasoning-recall tradeoff: while forward-KL distillation improves both reasoning and factual recall during pre-training, it slows factual acquisition during mid-training. This discrepancy stems from asymmetric teacher confidence across domains—teachers are confident on procedural tasks but uncertain on factual tokens, which students learn early. To resolve this imbalance, the authors introduce Switch Distillation, dynamically routing confident tokens to reverse-KL distillation and uncertain tokens to cross-entropy. Switch Distillation matches or exceeds baselines across reasoning and knowledge benchmarks while preserving factual recall.

Generated by CiteArk.

What this means

This paper uncovers a critical trade-off in language model training: applying standard knowledge distillation during mid-training accelerates reasoning capabilities but actively impedes the acquisition of factual recall. The authors diagnose the mechanism behind this phenomenon—namely, that teachers exhibit high predictive uncertainty on factual tokens, thereby diluting the ground-truth training signal for facts that the student has not yet acquired. To resolve this, they introduce Switch Distillation, a simple and computationally lightweight objective that routes confident tokens to reverse-KL distillation and uncertain tokens to standard cross-entropy. For practitioners developing foundation models, this provides a practical, drop-in training objective that substantially improves reasoning performance without sacrificing factual knowledge, with benefits that persist through subsequent alignment and post-training pipelines.

Reproduction progress
RunMatchRepeat

Not reproduced yet

0/76

No verifiable claim has successful reproduction evidence yet.

Awaiting author or institution signature.

Research claims

The main empirical claims extracted in this research plan, with their plan coverage and live evidence state. Click a claim to expand it inline.

Teacher predictive entropy distinguishes procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) with high ROC AUC (0.744-0.826) across training stages (Base, SFT, DPO, Instruct) and model sizes (1B, 7B, 13B).Inconclusive4/12 measurements supported · 12 assessedReported 0.815 score0.8079 score

Reported

0.815 score

Observed

0.8079 score

Delta

−0.0071 score

The execution did not produce enough comparable measurements to determine whether the paper's claim is supported. · This does not refute the paper's claim; it means the available evidence can neither confirm nor refute it yet.

roc_auc · Supported

Reported: 0.815 score

Observed: 0.8079 score

Reported ROC AUC of 0.815 was observed at 0.8079 (relative diff 0.87%), showing strong predictive entropy separation on OLMo-2 1B Base.

roc_auc · Supported

Reported: 0.77 score

Observed: 0.7967 score

Reported ROC AUC of 0.770 was observed at 0.7967 (relative diff 3.47%), confirming high ROC AUC separation on OLMo-2 1B SFT.

roc_auc · Supported

Reported: 0.744 score

Observed: 0.7931 score

Reported ROC AUC of 0.744 was observed at 0.7931 (relative diff 6.60%), maintaining the substantive finding of strong discrimination on OLMo-2 1B DPO.

roc_auc · Supported

Reported: 0.748 score

Observed: 0.7973 score

Reported ROC AUC of 0.748 was observed at 0.7973 (relative diff 6.59%), confirming procedural entropy remains distinct on OLMo-2 1B Instruct.

roc_auc · Inconclusive

Reported: 0.816 score

Observed:

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json

roc_auc · Inconclusive

Reported: 0.777 score

Observed:

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json

roc_auc · Inconclusive

Reported: 0.76 score

Observed:

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json

roc_auc · Inconclusive

Reported: 0.761 score

Observed:

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json

roc_auc · Inconclusive

Reported: 0.826 score

Observed:

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json

roc_auc · Inconclusive

Reported: 0.814 score

Observed:

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json

roc_auc · Inconclusive

Reported: 0.809 score

Observed:

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json

roc_auc · Inconclusive

Reported: 0.811 score

Observed:

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json

Open claim details
Mid-training ablation results for Always CE relative difference across Reasoning, Factual Recall, and Knowledge task groups.Experiment blocked; see the specific reasonReported -0.3 pp

Reported

-0.3 pp

Observed

Ablation checkpoints not released and training exceeds compute limits.

delta_percentage_points · Not assessed

Reported: -0.3 pp

delta_percentage_points · Not assessed

Reported: 0.1 pp

delta_percentage_points · Not assessed

Reported: -0.6 pp

Downstream task evaluation after post-training stage RLVR1 for 13b_rkd.Experiment blocked; see the specific reasonReported 75.6 pp

Reported

75.6 pp

Observed

Intermediate checkpoint weights after post-training stage RLVR1 not released.

downstream_accuracy · Not assessed

Reported: 75.6 pp

downstream_accuracy · Not assessed

Reported: 60.1 pp

downstream_accuracy · Not assessed

Reported: 50.3 pp

downstream_accuracy · Not assessed

Reported: 33.3 pp

downstream_accuracy · Not assessed

Reported: 38.1 pp

downstream_accuracy · Not assessed

Reported: 15.8 pp

downstream_accuracy · Not assessed

Reported: 53.3 pp

downstream_accuracy · Not assessed

Reported: 22.3 pp

downstream_accuracy · Not assessed

Reported: 8 pp

downstream_accuracy · Not assessed

Reported: 36.2 pp

downstream_accuracy · Not assessed

Reported: 16.9 pp

downstream_accuracy · Not assessed

Reported: 55.9 pp

downstream_accuracy · Not assessed

Reported: 55.2 pp

downstream_accuracy · Not assessed

Reported: 51.3 pp

downstream_accuracy · Not assessed

Reported: 36.4 pp

downstream_accuracy · Not assessed

Reported: 63.6 pp

The stage-dependent distillation tradeoff generalizes to the SmolLM family (SmolLM2 360M student, 1.7B Instruct teacher, SmolLM3 Stage-3 data), where mid-training KD exhibits the reasoning-recall deficit and Switch Distillation mitigates it.Experiment blocked; see the specific reason

Plot-only empirical finding without extractable discrete tabular coordinates; requires full intermediate training checkpoint trajectory or continuous sweep data not provided in the paper or repository.

Downstream task evaluation after post-training stage RLVR1 for 13b_fkd.Experiment blocked; see the specific reasonReported 73 pp

Reported

73 pp

Observed

Intermediate checkpoint weights after post-training stage RLVR1 not released.

downstream_accuracy · Not assessed

Reported: 73 pp

downstream_accuracy · Not assessed

Reported: 55.5 pp

downstream_accuracy · Not assessed

Reported: 47.3 pp

downstream_accuracy · Not assessed

Reported: 33.7 pp

downstream_accuracy · Not assessed

Reported: 36.1 pp

downstream_accuracy · Not assessed

Reported: 14 pp

downstream_accuracy · Not assessed

Reported: 54.7 pp

downstream_accuracy · Not assessed

Reported: 23.3 pp

downstream_accuracy · Not assessed

Reported: 8.3 pp

downstream_accuracy · Not assessed

Reported: 35 pp

downstream_accuracy · Not assessed

Reported: 17.2 pp

downstream_accuracy · Not assessed

Reported: 54 pp

downstream_accuracy · Not assessed

Reported: 56.6 pp

downstream_accuracy · Not assessed

Reported: 51.7 pp

downstream_accuracy · Not assessed

Reported: 35.4 pp

downstream_accuracy · Not assessed

Reported: 63.2 pp

Downstream task evaluation after post-training stage SFT for 13b_rkd.Experiment blocked; see the specific reasonReported 56.8 pp

Reported

56.8 pp

Observed

Intermediate checkpoint weights after post-training stage SFT not released.

downstream_accuracy · Not assessed

Reported: 56.8 pp

downstream_accuracy · Not assessed

Reported: 36.7 pp

downstream_accuracy · Not assessed

Reported: 31.3 pp

downstream_accuracy · Not assessed

Reported: 32.1 pp

downstream_accuracy · Not assessed

Reported: 38.9 pp

downstream_accuracy · Not assessed

Reported: 11.6 pp

downstream_accuracy · Not assessed

Reported: 55.4 pp

downstream_accuracy · Not assessed

Reported: 23.8 pp

downstream_accuracy · Not assessed

Reported: 8 pp

downstream_accuracy · Not assessed

Reported: 47.3 pp

downstream_accuracy · Not assessed

Reported: 17.3 pp

downstream_accuracy · Not assessed

Reported: 57.8 pp

downstream_accuracy · Not assessed

Reported: 57.4 pp

downstream_accuracy · Not assessed

Reported: 52.2 pp

downstream_accuracy · Not assessed

Reported: 38.1 pp

downstream_accuracy · Not assessed

Reported: 43.6 pp

Reproduction and technical details2
Implementation path
Official implementation0

The experiment runs a repository and commit explicitly declared by the paper and verified by CiteArk.

CiteArk reconstruction2

No author code is used. CiteArk independently rebuilds the paper-grounded protocol and records every generated file and assumption.

Information insufficient74

A falsifiable execution still needs a missing method detail, input, split, metric, or evaluation rule.

Signed research planVerifiedThe Research Plan CAP fixes paper sources, complete Claim coverage, planned Experiments, and source-declared values.Download research plan
Generated byCiteArk Official
Modelgoogle/gemini-3.8-flash
Completed
Artifact digest249476b838
Attestation statusTrusted signature verified
Output
View independent verification JSON