The manuscript presents a concrete construction for post-training factual update of a frozen DeepSeek-V4.1-Flash model that already contains an Engram n-gram lookup memory. An entry for a name is the pair (kappa, U) in R^{4 x d} with d = 5120. It fires only when the name's exact token sequence ends at a position. The entry's key is scored by the same rms-normalised expression the module uses for its own keys (Eq. 2), and the gate becomes the three-way softmax of Eq. 3: (pi^0_c, pi^M_c, pi^E_c) = softmax(0, x_c, y_c) with the residual update h_c <- h_c + pi^M_c v + pi^E_c U_c. Proposition 1 proves the algebraic properties of this gate under exact Euclidean arithmetic: reduction to the released two-way logistic when no entry is present, positivity and sum-to-one of the three shares, the affine rewrite that motivates the start U_0 = (p_0 - kappa)/pi^E, and the fact that the scores are independent of U so the state change is exactly pi^E_c delta_c. These steps hold under the stated premises; the page images confirm the printed formulae match the extracted text once formatting is restored. Proposition 2 then shows that, when keys are exact token sequences, no stored key is a suffix of another, at most one entry fires at a position, and the probes are fixed, any prompt that uses no entry is computed exactly as the released model. The proof is by induction over layers and is correct by construction; it underwrites the string-identical counts on untold and real-person questions once the probes refuse admission. The paper correctly notes that the proposition does not cover prompts in which an entry is used, nor does it prove probe accuracy itself. Those remain measured claims. Section 3.1 already states the implementation premise that non-firing positions take the released code path operation-for-operation, yielding the same numbers not merely close ones.
The value U is obtained by repeated damped minimum-norm Gauss-Newton steps (Eq. 5) whose residuals (Eq. 4) encode token-level targets taken from the model's own replies about known people who share the told values, plus holds that protect the lead of untold replies. Proposition 3 states the standard linear-algebra facts about the damped and undamped steps; the proofs are ordinary and the paper already cautions that the conditions are nonlinear, so the multi-step path is not proved to be globally minimal from the start. Lemma 1 and Proposition 4 establish that, inside a single stream copy, any orthogonal-invariant, positive-homogeneous, ray-monotone measure of an edit is a function of Euclidean length, so the shortest solution of G delta = r minimises every such measure. The paper immediately marks the three places the argument stops: the measure class excludes model-dependent statistical norms, the freedom is exact only for pure read-write stages, and a joint turn across four stream copies fixes nothing about relative weights. Table 5 shows that three natural weightings of the four copies produce essentially identical empirical outcomes on eight names, so the Euclidean choice on U is not separated by the data. The start p_0 is a fitted combination of class means and a ridge prediction; the factor 2/3 is a development choice. The paper reports that starting from the name's own state still recovers told facts but halves the language-spoken unwritten consequence rate, which is useful diagnostic evidence that the start matters for reasoning-like transfer.
Linear probes read the stream at the assistant marker after the full question and admit the entry only for kinds the name was told. Figure 4 shows that the same least-squares probe makes 0 of 384 mistakes at that position versus 102 of 384 when read at the name's last token. The look-ahead is therefore essential to the reported specificity, yet it is a genuine departure from strict left-to-right computation at one module; the paper notes the fallback to a name-position read when no assistant marker is present. Quantified free-form, multi-name, or streaming fallback numbers are not supplied; that is a disclosed scope limit and an editorial preference for additional experiments, not a contradiction of an existing claim.
The main evaluation (Section 4, Table 2, Figure 5) uses 120 made-up people each told one of 8 birthplaces, 5 universities and 4 occupations. Direct told facts are recovered 360/360; five non-condition wordings give 587/600; never-written one-step consequences give 481/600 against 102 untouched; the disagreement subset where two facts point to different answers is right for 147/225 (65 percent). Of the 119 errors, 73 (61 percent) answer from the wrong told fact, a direct consequence of storing a single shared U for all kinds. All 2,520 untold replies and all 500 real-person replies are string-identical to the untouched model, consistent with Proposition 2 once the probes refuse. Definition 1 cleanly separates stored rows, written-in consequences and truly unwritten consequences; an explicit audit line states that no unwritten-consequence wording appears in any solve. The yes-or-no variant (Table 4) raises assent but costs never-written answers, illustrating that each new row of G is not free. The MQuAKE-CF first-hop run (57/151 two-hop, most successes on citizenship) and the CounterFact run (320/500 taught completions, 136/1,000 paraphrases, substantial disturbance of other subject answers) correctly show that the method is not yet a general editor for known entities or arbitrary relations. The paper is candid about every major limitation: three fact kinds, one model family, look-ahead, exact-token keys, free-reply leakage, open-description confabulation, and the non-blind character of the 120-person set (question forms and facts were seen in development).
After reconciliation with the validated objections, two confirmed defects remain. The central empirical counts rest on saved replies and entry solves that are not supplied as machine-readable artifacts; a revision must deposit names/facts, entries or a regeneration script, probes, prompts, raw outputs, a scoring script, and a checkpoint checksum so that Table 2 and Figure 5 can be regenerated by a third party. Separately, the abstract and cost text say an entry adds 40,960 new weights/numbers while Section 3.1 states U alone is 4 x 5120 = 20,480; the accounting must be unified and "new weights" replaced by "entry parameters" or "stored numbers" for side-file quantities. Requests for broader value sets, per-kind ablations, earlier benchmark framing, or extra held-out sets are editorial preferences or already-addressed disclosures. Relative to ROME/MEMIT-style weight edits, external-memory routers, and the concurrent EngramEdit / Delta N-gram / User-as-Engram / ENGRAFT line, the distinctive elements are the third-alternative gate reuse, the single least-change vector per name whose targets are the model's own replies, and the exact non-interference guarantee when no entry is used. The contribution is therefore an existence proof and a carefully scoped prototype rather than a finished general editing method. The internal arithmetic of the reported counts is consistent; the formal propositions hold under their stated premises; the empirical claims are limited by value diversity, one-step closed-choice items, development exposure, and the absence of machine-readable entry files or full reply dumps in the review package. Those limits are already largely acknowledged by the authors.
preprint published 2026-11-14 17:18 UTC by preprint-admission
Preprint record published. Integrity screen clean; desk screen PASS. The review summary and available scorecard are attached as public metadata.