{"id":"37233000-83ce-4b85-91b8-f66d47d0593e","arxiv_id":"2507.04300","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"QF Learning updates transformer weights with a closed-form, backprop-free formula so that a model can recall an injected fact from memory after a single instructional example, but the reported evidence is only qualitative.","lead":"A new training method, QF Learning, injects facts into a transformer's weights using two forward passes and a closed-form update, skipping gradient backpropagation. The paper demonstrates the idea on one small language model with anecdotal examples rather than a measured benchmark.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 3 only solves a single-vector constraint, while the demonstrated multi-token answers require a batched update whose Lagrange derivation is absent from the submitted text; the single-example claim is therefore not specified for the actual experiment.","rationale":"I read the paper in good faith as proposing an activation-matching update for transformer layers. Eq. 3 is mathematically valid for a single vector constraint, and the least-norm derivation checks out. The load-bearing problem is that the central claim is about learning a natural-language fact, which necessarily involves multiple decoding steps, yet the paper only provides a single-vector formula. Section 2.3 explicitly records activations at each decoding step, but no multi-step update equation is given, and the promised batched derivation in the Appendix is absent from the submitted text. This is not merely a missing stylistic appendix: it means the method as specified cannot be checked for the exact setting demonstrated. A second, related concern is the absence of quantitative evaluation: the experiment is a narrative of one hand-picked fact with no error bars, no held-out facts, no baseline, and no ablation. The reader's weakest assumption about the activations u and v encoding the fact is plausible, but even if that assumption holds, the multi-token gap and the missing batched derivation would still undermine the central claim. I therefore agree with the reader's REJECT verdict, but I would ground it more specifically in the missing multi-token specification rather than only in the thin empirical support.","tokens_in":4379,"tokens_out":8888,"duration_ms":110619,"concrete_test":"Use the released qf_learn_simple.py with Qwen2.5-1.5B-Instruct on a two-token fact such as 'Qi founded Oxinnovate' with target answer 'Qi founded it'. Let (u1,v1) and (u2,v2) be the instructed activations for the two decoding steps, and (u'1,v'1), (u'2,v'2) the uninstructed ones. Apply Eq. 3 first for token 1, then for token 2 using the updated W as the starting point, and compute ||W'' u'_1 + v'_1 - (W u_1 + v_1)||. If this residual is not at machine precision, the sequential interpretation of Eq. 3 cannot reproduce the multi-token answer and the batched derivation is required. Clear the KV cache between phases to ensure no instruction remains, and compare first-token logits before and after the update. If the batched Lagrange derivation cannot be produced, or the batched update fails on held-out facts, the single-example claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 derives Eq. 3 as the least-norm solution to one linear constraint W'u' + v' = W u + v. This is valid for one token position, but Section 2.3 says activations are recorded 'at each decoding step' and Step 6 of the experiment involves the multi-token answer 'Founded by Qi'. For a multi-token answer, either the update is applied per decoding step, in which case later updates change W' and destroy the exact equality achieved for earlier steps, or a single batched update over all positions is needed. The batched solution is not Eq. 3; the paper states that its derivation is in the Appendix, but the submitted text contains no Appendix. Thus the central claim 'a single example suffices' is not backed by a complete mathematical specification for the setting actually demonstrated. Moreover, the empirical section is a narrative of one hand-picked fact: no quantitative success metric, no held-out set, no baseline or ablation, so we cannot tell whether the reported behavior is caused by Eq. 3 or by the instruction remaining in context. Even granting the linear algebra, the claimed preservation of old knowledge is not established: a minimal-Frobenius-norm update can still change Wx strongly for any x aligned with u', and only one other fact (Jack Ma) is checked.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Quick Feedforward (QF) Learning, a method that updates a single weight matrix in a transformer layer via a closed-form least-norm solution (Eq. 3) so that knowledge from an instruction pass can be consolidated into the weights without gradient backpropagation. The method is demonstrated on Qwen2.5-1.5B-Instruct through a qualitative sequence of queries about a fictional company, claiming single-example learning, preservation of existing knowledge, and generalization to paraphrases. The paper argues that the approach is efficient, biologically inspired, and resistant to catastrophic forgetting.","tokens_in":4638,"tokens_out":7057,"duration_ms":72921,"significance":"If substantiated, the idea of closed-form, gradient-free knowledge consolidation would be an interesting contribution to knowledge editing and parameter-efficient learning. The derivation of Eq. (3) is correct for the single-vector linear constraint, and the open-source release is a strength. However, the current manuscript provides no quantitative evaluation, no baselines, no systematic ablations, and the mathematical specification is incomplete for the multi-token setting actually demonstrated. At present, the central claims are supported only by anecdotal evidence.","major_comments":[{"comment":"The derivation of Eq. (3) solves a single-vector constraint W'u' + v' = W u + v. The experiments, however, involve multi-token answers (e.g., Step 6, 'Founded by Qi'), and Section 2.3 states that activations are recorded at each decoding step. The paper says the batched derivation is provided in the Appendix, but the submitted text contains no Appendix. Consequently, the mathematical specification of the update used in the experiments is absent, and it is unclear whether the update is applied per decoding step (which would destroy the exact equality achieved for earlier steps) or as a single batched update (which is not Eq. 3). This is load-bearing for the central claim that a single example suffices.","section":"Sec. 2.2-2.3, Eq. (3)"},{"comment":"The empirical section is a qualitative list of eight steps with no quantitative metrics, no success criteria, no repeated trials, no baselines, and no ablation of the free parameters (target layer index and qfsignificance mask). The only preservation check is a single fact ('Jack Ma'), and the generalization check is a single paraphrase. These data do not support the claims of knowledge preservation, generalization, or absence of catastrophic forgetting.","section":"Sec. 3, Table 1"},{"comment":"The constraint equates only the feedforward sublayer output. In a transformer with residual connections, the input to the next layer is the residual stream, which includes the layer input u (or u') in addition to the sublayer output. Thus, even if W'u' + v' = W u + v, the full layer output differs between the instructed and uninstructed passes unless u' = u. The paper does not explain how the update accounts for the residual connection or the attention output, so the theoretical mechanism by which the updated model reproduces the final answer is not established.","section":"Sec. 2.2, Eq. (1)"},{"comment":"The claim that QF is 'inherently resistant to catastrophic forgetting' is not supported. A minimal-Frobenius-norm update can alter the output for any input aligned with u', so preservation of arbitrary old facts is not guaranteed. The paper provides no theoretical bound and no empirical test beyond a single fact, and the 'qfsignificance mask' introduced in Figure 4 is a free parameter with no formal definition or sensitivity analysis.","section":"Sec. 4, Discussion"}],"minor_comments":[{"comment":"The notation u, v, u', v' is not fully defined; in particular, v appears as an additive vector in Eq. (1), but the text refers to it as an activation. Please clarify whether v is the feedforward bias or the sublayer output.","section":"Sec. 2.2"},{"comment":"Table 1 is not a table of results but a list of narrative steps; consider presenting actual generated strings and quantitative measures (e.g., exact-match accuracy, confidence) in a proper table.","section":"Sec. 3, Table 1"},{"comment":"The qfsignificance mask is mentioned in the code snippet but never formally defined in the main text; its effect on Eq. (3) should be made explicit.","section":"Sec. 2.3, Figure 4"},{"comment":"The paper repeatedly claims 'minimal parameter modification' but does not report which parameters are modified or how many; stating the rank of the update and the fraction of changed entries would clarify this.","section":"General"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early preprint; the absence of the promised appendix and the complete lack of quantitative evaluation would likely preclude acceptance at any rigorous venue. The remaining concerns, particularly the residual-stream issue, suggest that the theoretical basis of the method needs substantial clarification before further review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2507.04300. The genuinely new part is the mechanism: run an instructed forward pass, record the activations, then apply a closed-form least-norm update to one feedforward layer so the model answers the same query without the instruction. No gradient anywhere. Eq.3 is a standard rank-one correction and it is correct for a single token position. That is a real conceptual step relative to ROME, which still uses gradient to compute the value vector. The paper explains the intended mechanism clearly and the authors admit the update only works when the chosen layer carries the right knowledge. They also release code and models, which is worth something.\n\nThe soft spots are the evidence and the framing. The entire empirical case is one hand-picked fact about Oxinnovate. No quantitative success rate, no baselines like ROME or knowledge neurons, no repeated trials, and forgetting is checked with a single query about Jack Ma. The qfsignificance mask is a manual per-token choice that is not justified. And the promised batched derivation is missing from the text: the paper says Appendix, but there is no appendix. That said, the stress-test worry about multi-token answers may be overstated. The procedure in Section 2.3 can be read as sequential per-token updates, where each step applies Eq.3 to the current weights. That would not require the batched math, though the paper should say explicitly whether that is what the code does, because the equality in Eq.1 is then local to each step, not global. The bigger problem is the title and the word 'training' in the abstract. This is knowledge editing, not training a model from data, and that framing invites the overclaim that a single example suffices for reliable learning. The one demonstration does not support that.\n\nWho is it for? Researchers looking at knowledge editing, continual learning, or gradient-free adaptation. The idea is worth a serious look as a testable baseline. I would send it to peer review, but I would expect major revision: a proper evaluation on a set of facts, comparison to existing editing methods, an explicit description of the sequential or batched update, and a less grandiose title. If the code reproduces on a broader set, it becomes a modest useful contribution. As it stands, I'd lean reject as a submission, but not desk-reject.","headline":"A genuinely gradient-free knowledge-editing mechanism that is clearly explained but validated on a single anecdote, with the multi-token math left underspecified.","tokens_in":5135,"tokens_out":4653,"would_cite":false,"duration_ms":51600,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims a transformer can learn a new fact from one instruction example through a closed-form weight update, with no gradient backpropagation.","keywords":["quick feedforward learning","knowledge consolidation","closed-form weight update","gradient-free training","language model editing","continual learning","catastrophic forgetting","transformer decoder"],"falsifier":"Apply QF to a set of, say, 50 unrelated facts on the same model, recording for each fact whether the instructed pass answers correctly and whether the post-update closed-book pass answers exactly. Compute the residual $\\lVert W'u' + v' - (Wu+v)\\rVert$ on held-out tokens: if the residual is large for facts whose update still appears to work, or if the post-update accuracy stays near zero even when the instructed pass is correct, the central claim is disproved. A second test: update two conflicting facts about the same subject in sequence and check whether the second edit evicts the first.","tokens_in":4179,"feed_emoji":"🧠","tokens_out":7998,"duration_ms":84698,"temperature":0.7,"pith_summary":"Quick Feedforward (QF) Learning is a proposed training procedure that lets a transformer model absorb a new fact from a single instruction–query example without computing any gradient. The paper's central claim is that one closed-form update to a single layer's weight matrix, derived by equating the instructed and uninstructed forward passes, is enough to store the fact in the weights and reproduce the correct answer later. In experiments on a 1.5-billion-parameter instruction-tuned model, the fact is answered after the update, an old unrelated fact is retained, a paraphrase is answered, and a second fact can be added without erasing the first. If the claim holds, QF would make single-example knowledge insertion cheap enough to run in the same environment as inference, and it would provide a non-backpropagation mechanism for continual learning.","feed_headline":"One example teaches a model a new fact, no backprop","feed_subtitle":"Two forward passes rewrite one weight matrix so the model answers from memory and keeps old facts.","key_machinery":"The carrier of the argument is the linear consolidation identity $W'u' + v' = Wu + v$ together with its least-norm solution. The logic is that the difference between the instructed pass and the uninstructed pass at one intermediate layer is a discrepancy the weight matrix can absorb: changing only $W$ to $W'$ should make the uninstructed activations match the instructed output. The framework wraps this in three phases, QF-instruct, QF-update, and QF-infer, and uses a token-level significance mask to restrict which output positions get reinforced, with both passes truncated at the target layer before the update is applied.","core_discovery":"QF Learning claims to consolidate instruction-derived knowledge directly into the weights of a transformer decoder through two forward passes and a closed-form weight update. During the QF-instruct pass the model sees both the question and an instruction stating the new fact, and the activations $u, v$ at a chosen intermediate layer are recorded; during the QF-update pass only the question is seen, with the instructed answer as the next token, producing $u', v'$. The update\n$$\nW' = W - \\bigl( W(u' - u) + (v' - v) \\bigr) \\left( u'^\\top u' \\right)^{-1} u'^\\top\n$$\nis the minimum-Frobenius-norm change satisfying $W'u' + v' = Wu + v$, so the closed-book pass now produces the same output activation the open-book pass produced. After the update the model answers the question without the instruction, answers a paraphrase, retains previously known facts, and can absorb a further fact about the same entity. The demonstration uses Qwen2.5-1.5B-Instruct and a single example per fact.","pith_inferences":["Editorial inference: the method implies that factual knowledge at the chosen intermediate layer is linearly readable and can be transferred by a single rank-one-like correction; if true, layer selection could be automated by measuring which layer's update best preserves the instructed pass.","Editorial inference: a natural stress test the paper does not run is a counterfactual edit, for example updating 'Qi started Oxinnovate' and then 'Alice started Oxinnovate' on the same model, since Eq. (3) gives no mechanism for arbitrating conflicting activation pairs.","Editorial inference: the paper only sketches the multimodal extension, but the same linear update could be tested by injecting a visual instruction through cross-attention at intermediate layers; whether the linear consolidation survives that route is an open empirical question."],"forward_implications":["A fact can be inserted into a 1.5-billion-parameter model in two forward passes on a single GPU, without optimizer state or gradient computation.","The inserted fact is answerable from the model's weights alone after the update, not only while the instruction is present.","Previously learned facts survive the update, so the method is positioned as a continual-learning alternative to fine-tuning.","The model answers a paraphrase after the update, indicating that the weight change captures the underlying association rather than the surface wording.","Successive updates can add different facts about the same entity while the earlier fact is retained."],"supporting_citations":[{"why":"Supplies the closest targeted weight-editing method, which QF contrasts as relying on gradient computation.","marker":"[1]"},{"why":"Supplies the knowledge-neuron baseline whose attribution scores require gradients, the limitation QF targets.","marker":"[2]"},{"why":"Supplies the hypernetwork knowledge-editing baseline that QF contrasts as gradient-dependent.","marker":"[3]"},{"why":"Defines the standard transformer decoder architecture on which QF updates a single layer.","marker":"[4]"},{"why":"Supplies the Qwen2.5-1.5B-Instruct model used in all reported experiments.","marker":"[5]"}],"fun_headline_variants":["No backprop: one example teaches a model a new fact","Two forward passes replace backprop for single-example learning","Closed-form update: one example, no gradient, knowledge retained","QF Learning: brain-like training with just forward passes","Teach a transformer a new fact without backprop"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the activations $u$ and $v$ recorded during the instructed pass encode the new fact, and that the linear relationship the weight update relies on, changing only one layer's weight matrix to reproduce the instructed output, accurately describes the model's forward computation; if the chosen layer is too early or too late, or the instructed pass gives a wrong answer, the update cannot teach the model anything.","fun_headline_variants_meta":{"raw":{"variants":["No backprop: one example teaches a model a new fact","Two forward passes replace backprop for single-example learning","Closed-form update: one example, no gradient, knowledge retained","QF Learning: brain-like training with just forward passes","Teach a transformer a new fact without backprop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000301,"raw_usage":{"total_tokens":1698,"prompt_tokens":869,"completion_tokens":829,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":749}},"tokens_in":485,"tokens_out":829,"duration_ms":7344,"temperature":1.0,"reasoning_tokens":749,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:50:35.534503+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply QF to a set of, say, 50 unrelated facts on the same model, recording for each fact whether the instructed pass answers correctly and whether the post-update closed-book pass answers exactly. Compute the residual $\\lVert W'u' + v' - (Wu+v)\\rVert$ on held-out tokens: if the residual is large for facts whose update still appears to work, or if the post-update accuracy stays near zero even when the instructed pass is correct, the central claim is disproved. A second test: update two conflicting facts about the same subject in sequence and check whether the second edit evicts the first.","supporting_citations":[{"cited_title":"Locating and editing factual associations in gpt","cited_arxiv_id":null,"evidence_quote":"Supplies the closest targeted weight-editing method, which QF contrasts as relying on gradient computation."}],"review_version":1}