{"id":"f8a9549f-5dcf-419d-976f-3cec5c8e19b2","arxiv_id":"2412.14964","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Self-distillation from a model's own in-context answers injects factual knowledge into LLM weights more efficiently than supervised fine-tuning and is competitive with RAG.","lead":"A language model can teach itself new facts by generating questions about a document, answering them with the document in its prompt, and then training itself to reproduce those answers without the document. This 'prompt distillation' approach beats standard fine-tuning for knowledge injection and approaches retrieval-augmented generation, without needing a larger teacher model.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PD's advantage hinges on the teacher's in-context extraction of facts, which is never measured; HotpotQA's single-hop-to-multi-hop gap makes this the least secure link.","rationale":"The paper's controlled comparisons—same training questions, LoRA setup, epochs, and seeds; two graders; substring match; and a public codebase—give strong support for the PD-over-SFT result within the tested domains. I therefore do not treat the teacher-capability issue as a refutation. It is a boundary condition on the central claim: the method cannot inject knowledge that the teacher cannot read from the document. The reader identified the same assumption, so I agree. The additional caveats the reader raises, such as RAG baselines without error bars and PD XL tuned on evaluation sets, are real but secondary, since the primary PD-vs-SFT comparison has error bars and is robust to grading. My proposed test is cheap and would settle whether the teacher signal is actually the operative mechanism. The verdict should remain conditional: accept the empirical comparison, but require the teacher-capability check for the claimed generality.","tokens_in":30814,"tokens_out":14685,"duration_ms":128850,"concrete_test":"Take 200 held-out training questions per model and dataset (Squadshifts plus HotpotQA). For each, prompt the teacher model with only the privileged document and the question, decode greedily at temperature 0, and grade with the same LLM-judge used in the paper. Report in-context accuracy, mean teacher entropy, and initial KL between teacher and student logits. Then split the evaluation results by whether the teacher's greedy answer is correct. If PD's advantage over SFT is concentrated in teacher-correct items and disappears on teacher-incorrect items, the central claim is conditional on the teacher's in-context competence. If PD still wins on teacher-incorrect items, the concern is resolved and the method is more robust than the current design demonstrates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Prompt distillation transfers the teacher's in-context policy into the student's weights, so the quality of the transferred signal is bounded by the teacher's ability to extract the target facts from the privileged document. The paper never measures this quantity directly. The closest evidence, base model + RAG accuracy in Tables 2 and 7, is not a substitute: RAG supplies multiple retrieved documents, whereas the teacher in training sees a single document plus a generated question, and training answers were sampled at temperature 1.5, which can inject noisy targets (Sec 2.1 and 4.1). The HotpotQA design makes the dependency critical: training questions are single-hop per paragraph while test questions are multi-hop (Sec 4.7), so the teacher must already perform the first hop in-context for the student to have anything useful to learn. The paper's own concession that small base models benefit from a larger expert (Sec 5) indicates that the self-distillation loop inherits the base model's reading comprehension. Appendix Q shows Qwen2.5-14B PD failures that are plausible but factually wrong, consistent with noisy teacher targets. If teacher capability is low in a deployment domain, the PD-over-SFT advantage could vanish, and the mechanism claim would rest on the query distribution rather than on knowledge internalization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes prompt distillation (PD) for knowledge injection: the same instruct LLM generates question-answer pairs from free-form documents, serves as a teacher that conditions on the document to produce soft-label logits, and is trained via a LoRA adapter as a student that answers without the document. The method is compared with supervised fine-tuning (SFT), unsupervised fine-tuning, and RAG on modified Squadshifts and HotpotQA across Llama-3-8B and Qwen2.5-14B/3B. The main empirical claims are that PD outperforms SFT in closed-book accuracy and data/wall-clock/parameter efficiency, that closed-book PD is competitive with RAG, that PD+RAG improves over RAG, and that PD helps multi-hop HotpotQA reasoning.","tokens_in":31059,"tokens_out":6590,"duration_ms":57652,"significance":"The empirical core is well controlled and unusually thorough: PD and SFT share the same training questions, LoRA configuration, and training duration; results are averaged over three seeds; grading is validated with a second LLM judge and with substring matching; and code is released. If the results hold, PD is a practically valuable parameter-efficient knowledge injection method that does not require a larger teacher. The main caveats are that the headline comparison with RAG is more optimistic than the tables support, and the HotpotQA result depends on teacher in-context extraction ability that is never directly measured.","major_comments":[{"comment":"The claimed identity after Eq. (1), E_c[D_KL(π_θ(·|c,q,a_<i) ∥ π_θ'(·|q,a_<i))] = I(c; a_i | q, a_<i), is not correct in general. With the expectation taken only over c, the left side equals I(c; a_i | q, a_<i) plus D_KL(E_c π_θ(·|c,q,a_<i) ∥ π_θ'(·|q,a_<i)). The equality to mutual information holds only if the student distribution equals the teacher's marginal distribution averaged over c, which is not guaranteed before training. This matters because the paragraph uses the equality to justify the claim that minimizing the loss 'ensures the student internalizes the extra information.' Please correct the derivation or explicitly state the additional KL term.","section":"Sec. 2.2, Eq. (1) vicinity"},{"comment":"The claim that PD 'can even surpass RAG' and the conclusion that PD achieves closed-book performance 'competitive with RAG' are not supported by the closed-book numbers in Table 2. For Llama-3-8B, closed-book PD is below base-model RAG on Amazon (86.1 vs 86.3), New Wiki (94.4 vs 95.6), and NYT (93.6 vs 96.3), and above only on Reddit (79.5 vs 78.6); for Qwen2.5-14B and Qwen2.5-3B, RAG is ahead on all four datasets. This is not merely a wording issue: the abstract and Figure 1 present RAG as a method that PD can surpass, whereas the actual support is for PD+RAG or PD XL, not for closed-book PD. In addition, the RAG baseline cells are reported without error bars, so the size of the PD-RAG gaps should be quantified with repeated sampling or a statistical test before making 'competitive' or 'surpass' claims.","section":"Abstract, Sec. 4.4, Table 2"},{"comment":"The HotpotQA generalization claim rests on an unmeasured assumption about the teacher. Training questions are single-hop per paragraph, while test questions are multi-hop, so the teacher must already extract the relevant fact from the single paragraph in-context for the student to receive a useful signal. The paper never measures teacher in-context accuracy on the generated training questions. Base-model RAG accuracy is not a substitute, because RAG supplies multiple retrieved paragraphs, whereas the teacher sees one document plus a generated answer sampled at temperature 1.5, which can be noisy (Sec. 4.7; Appendix Q shows plausible but factually wrong PD outputs). Please add a direct measurement of the teacher's in-context QA accuracy on the training questions, or an ablation with lower answer-generation temperature, and discuss how the method behaves when the base model's in-context extraction is weak.","section":"Sec. 4.7, HotpotQA"}],"minor_comments":[{"comment":"The caption uses 'PD (XL)' without defining it; please add a pointer to the PD XL description in Sec. 4.6.","section":"Fig. 1 caption"},{"comment":"The text says teacher logits can be stored during answer generation to eliminate extra compute, but Appendix H states that this optimization was not implemented due to storage constraints; please clarify in Sec. 2.4 that this is a possible implementation variant, not the setting measured in the wall-clock experiments.","section":"Sec. 2.4 and Appendix H"},{"comment":"The RAG setup uses k=7 retrieved documents for Squadshifts but k=10 for HotpotQA; state this difference explicitly in Sec. 4.7 rather than only in the appendix.","section":"Sec. 4.3 and Sec. 4.7"},{"comment":"In the first grading example, the output uses '<grad>' instead of '<grade>'; fix the typo.","section":"Appendix O.5"},{"comment":"The notation D_KL-reg(i) in Eq. (2) is not defined; please specify that it is the forward KL divergence between the teacher's and student's distributions for the regularization response, analogous to Eq. (1).","section":"Sec. 2.2, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The paper's main comparison (PD vs SFT) is solid and well controlled, and the additional grading checks are a strength. My main concerns are that the abstract and conclusion overstate the RAG comparison, the mutual-information derivation in Sec. 2.2 is mathematically incorrect as written, and the HotpotQA result lacks a direct check of the teacher's in-context extraction ability. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what's actually new: this paper shows that prompt distillation—self-distillation where the same model serves as expert, teacher, and student—can internalize facts from unstructured documents, and that it consistently beats standard SFT across Llama-3 and Qwen2.5 at multiple sizes. The comparison is well controlled: same training questions, same LoRA config, three seeds, two graders, substring-match confirmation. I believe the central claim.\n\nThe paper does a lot of things right. It tests the method across model families, includes data-efficiency and wall-clock comparisons, examines temperature and LoRA rank, and shows that larger teachers often hurt—a useful empirical counterpoint to the usual assumption that bigger experts are better. The regularization result on Tülu 3 is also a nice touch: soft targets appear to mitigate catastrophic forgetting better than hard SFT targets. Code is released.\n\nSoft spots, in proportion. First, RAG baselines have no error bars; given that PD's 'competitive with RAG' claim rests on these numbers, they should be added or the claim softened. Second, the abstract says PD 'can even surpass RAG,' but the only dataset where that clearly holds is after the PD XL tuning, which was not a pre-registered configuration. The main tables show PD matching or slightly missing RAG on most closed-book settings. Third, the HotpotQA setting trains on single-hop questions and tests multi-hop; the paper is upfront about this, and the gains are still real, but it weakens the 'knowledge internalization for reasoning' story. The stress-test worry about teacher capability is fair but not fatal: the paper's own ablations (small model needing a larger expert, Appendix Q failures) show the boundary, and the main PD-vs-SFT result stands even if teacher quality is imperfect.\n\nWho benefits: practitioners wanting a parameter-efficient way to internalize documents, and researchers working on knowledge editing/injection. It deserves a serious referee. I'd send it with a request to add error bars on RAG, temper the abstract, and discuss the teacher-capability boundary explicitly.","headline":"Solid empirical case that self-distillation beats SFT for knowledge injection, with a few statistical nits; worth refereeing.","tokens_in":31595,"tokens_out":2711,"would_cite":true,"duration_ms":22213,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Prompt distillation injects new facts into LLMs better than supervised fine-tuning.","keywords":["prompt distillation","self-distillation","knowledge injection","large language models","retrieval-augmented generation","supervised fine-tuning","LoRA","multi-hop reasoning"],"falsifier":"Apply prompt distillation to a document QA task where the base model is shown to have poor in-context comprehension, for example long or adversarial passages where the base model's RAG accuracy is at chance; if prompt distillation does not beat supervised fine-tuning on that task, the mechanism is the teacher's comprehension rather than the soft-target loss. Alternatively, replace the teacher logits with one-hot targets at the same temperature and data budget; if prompt distillation's superiority disappears, the soft distribution is the active ingredient.","tokens_in":30619,"feed_emoji":"🧠","tokens_out":8346,"duration_ms":57789,"temperature":0.7,"pith_summary":"Prompt distillation, a self-distillation method in which a teacher model reads a new document in its prompt and a student model learns from the teacher's full token distribution without that document, can inject factual knowledge into LLMs more effectively than supervised fine-tuning. Across Llama-3-8B and Qwen2.5-3B/14B, the trained students answer questions about documents they never see at test time with accuracy close to or above retrieval-augmented generation (RAG) on Squadshifts-derived QA, and they improve RAG when the two are combined. On HotpotQA multi-hop questions, prompt distillation lifts closed-book accuracy from 53% to 74% for Llama-3-8B and outperforms supervised fine-tuning in every configuration. These gains come without a larger teacher model, structured knowledge triples, or human labels: the model generates its own question-answer data at high temperature and learns from soft target distributions rather than hard answer tokens.","feed_headline":"Prompt distillation beats fine-tuning for LLM knowledge injection","feed_subtitle":"Same model as teacher learns from its own prompted context, matching retrieval-augmented answers.","key_machinery":"The central mechanism is the KL-divergence distillation objective at temperature T=2 between the teacher's policy πθ(a_i | c, q, a_{1..i-1}) and the student's πθ′(a_i | q, a_{1..i-1}), averaged over answer positions; the identity E_c[D_KL(...)] = I(c; a_i | q, a_{1..i-1}) makes explicit that the student is learning to reproduce the contextual information inside its own weights. The student is the same model as the teacher with a LoRA adapter initialized to zero, so training toggles the adapter without extra memory, and teacher logits can be precomputed during answer generation. High temperature flattens the teacher's distribution so the student learns which tokens to avoid—something one-hot cross-entropy cannot convey without many samples—and an optional regularization term (Eq. 2) adds KL to the base model's distribution on unrelated instruction-response pairs to prevent catastrophic forgetting.","core_discovery":"The paper claims that the full policy of a teacher conditioned on a privileged context—not just the sampled answer tokens—is the right learning signal for knowledge injection, and that distilling this policy into a student that shares the teacher's weights (via a LoRA adapter) transfers facts at least as well as retrieval. The training loss is the average KL divergence between teacher and student over answer positions; in expectation over contexts, this equals the mutual information I(c; a_i | q, a_{1..i-1}), meaning the student learns to supply the information that the context would have provided. On the four Squadshifts domains, prompt distillation reaches 86.1%, 94.4%, 93.6%, and 79.5% for Llama-3-8B versus 75.9%, 89.5%, 87.5%, and 69.8% for supervised fine-tuning, and closed-book prompt distillation matches the instruct model with RAG; on HotpotQA, it raises Llama-3-8B closed-book accuracy from 53% to 73.7% with 100 questions per context, and prompt distillation plus RAG exceeds RAG alone. The paper further finds that using a larger model as teacher or expert generally hurts performance, because style mismatch outweighs any gain in answer quality.","pith_inferences":["A direct implication beyond the paper: if the mutual-information identity is the right way to view distillation, then prompt distillation should work for any privileged signal expressible in token distributions, such as reasoning scratchpads, tool-use policies, or safety constraints; the results suggest the method generalizes beyond factual QA.","The style-mismatch finding predicts that distillation gains depend on teacher-student token-level agreement, so a testable extension is to control teacher entropy while varying data coverage to isolate the mechanism.","Because prompt distillation XL with Tülu 3 regularization recovers almost all of MMLU-Pro, weight-based knowledge updates could plausibly replace retrieval in agentic systems that need to remember private documents without prompting overhead.","One could test the causal role of soft targets by ablating the teacher's distribution to one-hot at the same temperature: prompt distillation's advantage should shrink or disappear if the soft-distribution signal is what matters."],"forward_implications":["Prompt distillation provides a closed-book route to knowledge injection with accuracy competitive with RAG on single-document QA, suggesting weight-based knowledge storage can replace retrieval for some deployments.","Because prompt distillation plus RAG outperforms RAG alone, models fine-tuned this way can also make better use of retrieved context, which is useful when retrieval quality is imperfect.","Prompt distillation's data efficiency—matching supervised fine-tuning's accuracy with roughly a tenth of the training examples—cuts the cost of building new-knowledge datasets.","Prompt distillation's lower forgetting under regularization means it can be combined with general instruction-following data without destroying prior capabilities.","Self-distillation removes the need for a larger proprietary teacher, lowering the barrier for organizations that want to update open models with private documents."],"supporting_citations":[{"why":"Supplies the KL-divergence distillation loss and temperature scaling that define the distillation objective.","marker":"Hinton et al., 2015"},{"why":"Demonstrates distillation of context into weights, the basis for extending prompt distillation to factual knowledge.","marker":"Snell et al., 2022"},{"why":"Early use of self-distillation to internalize a prompted policy, the origin of the method.","marker":"Askell et al., 2021"},{"why":"Provides the DAgger framework that motivates learning from the teacher's full policy to correct sub-optimal tokens.","marker":"Ross et al., 2011"},{"why":"Provides the Squadshifts QA corpus from which the closed-book evaluation datasets are derived.","marker":"Miller et al., 2020"},{"why":"Provides the HotpotQA multi-hop benchmark used to test generalization beyond single-hop training.","marker":"Yang et al., 2018"},{"why":"Supplies the Bonito question generator used to scale training data and create the prompt distillation XL variant.","marker":"Nayak et al., 2024"},{"why":"Provides the LoRA adapter that lets the same network play both teacher and student roles.","marker":"Hu et al., 2021"},{"why":"Supplies the RAFT-style supervised fine-tuning with distractors baseline that the RAG setting is compared against.","marker":"Zhang et al., 2024"}],"fun_headline_variants":["Self-distillation outperforms fine-tuning for knowledge injection","LLMs learn new facts by distilling their own prompts","Prompt distillation matches RAG without retrieval","Self-distilled policy transfer beats fine-tuning for facts","No retrieval needed: self-distillation injects knowledge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on the teacher model being able to extract the target facts from the document in its prompt reliably enough that its token distributions carry useful information; if the base model cannot comprehend the privileged context, the distillation signal is noisy and prompt distillation's advantage over supervised fine-tuning should vanish.","fun_headline_variants_meta":{"raw":{"variants":["Self-distillation outperforms fine-tuning for knowledge injection","LLMs learn new facts by distilling their own prompts","Prompt distillation matches RAG without retrieval","Self-distilled policy transfer beats fine-tuning for facts","No retrieval needed: self-distillation injects knowledge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001099,"raw_usage":{"total_tokens":4581,"prompt_tokens":939,"completion_tokens":3642,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":3568}},"tokens_in":555,"tokens_out":3642,"duration_ms":20028,"temperature":1.0,"reasoning_tokens":3568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:45:31.783392+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply prompt distillation to a document QA task where the base model is shown to have poor in-context comprehension, for example long or adversarial passages where the base model's RAG accuracy is at chance; if prompt distillation does not beat supervised fine-tuning on that task, the mechanism is the teacher's comprehension rather than the soft-target loss. Alternatively, replace the teacher logits with one-hot targets at the same temperature and data budget; if prompt distillation's superiority disappears, the soft distribution is the active ingredient.","supporting_citations":[],"review_version":1}