acceptodds
Under review as a conference paper at ICLR 2027

Functional Memory Compilation: Behavior-Level Migration of Aggregate LoRA Memory

Abstract

When a language model is upgraded, facts accumulated in its memory adapter may become inaccessible through the successor. Migration is difficult when the original interaction history is unavailable. Parameter mappings need not preserve individual answers, while response distillation requires target gradient training. We study migration from an already-trained aggregate low-rank adaptation (LoRA) memory. The source model and retained questions remain accessible, but the questions' original answers are unavailable. We propose Functional Memory Compilation (FMC), which transfers the source's observable answers rather than its parameter coordinates. FMC recovers responses at retained questions and retokenizes them for the successor. It turns selected answer positions into joint constraints on the target's output. Batched forward computation followed by one analytic low-rank fit installs a target-native output-head adapter without backpropagation or optimizer steps. Across the evaluated memory tasks, FMC improves task scores and lowers fixed-pool target installation cost over same-content distillation at the evaluated limited training budget.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
LLM, Memory

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_obb8OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

A source model carries a LoRA "memory adapter" trained on past interactions; the interaction history is lost, and only the source model + adapter and the questions remain. FMC migrates the memory to a new backbone in three steps: query the source at the retained questions to recover answers, keep the prompt-stable ones, then fit a single closed-form ridge regression (no backprop) that turns the target model's internal states into logit boosts on the recovered answer tokens. The result is a small output-head-only adapter, installed in seconds. On three memory benchmarks it beats transferred and distilled baselines and reproduces the recovered answers exactly.

Critical strengths
  1. Real problem, clean abstraction. Memory continuity across backbone upgrades without replay data is legitimate and under-served; "transfer observable answers, not parameter coordinates" is crisp.
  2. Transparent/honest reporting. Pre-declared selection rules, sealed states, input hashes, bootstrap CIs, and disclosure of every confound — including the ones that indict the results. A model of honest reporting.
  3. The analytic install is genuinely cheap and verified equivalent to the iterative solve of the same objective at lower cost; the mechanism ablations (e.g., address–answer shuffling) are well designed.
Critical weaknesses

W1 — The headline result is a unit test, and the paper's own control shows a text file matches the adapter. The "250/250 recovered values reproduced verbatim" result is forced by construction: the ridge constraints literally set the boosted token's logit at those prefixes, and the address pool was itself admitted using temporary FMC writes with readback fidelity as the criterion. More importantly, the paper's own control — feeding the same recovered values to the same reasoner as text — produces identical final strings and identical F1 on all questions. In the evaluated pipeline, a small text file matches the compiled adapter; the one regime where parametric storage could win (adapter-only serving, no context window) is never evaluated. The paper concedes this ("not superiority to text storage") — but then the central experiment does not support the paper's thesis.

W2 — The one superiority claim over a standard alternative rests on a budget the authors admit was chosen against the baseline. Distilled LoRA receives exactly 2 epochs with min-train-loss checkpointing — 1/16th of the 32 epochs, over six-layer Q/K/V/O LoRA, that built the source memory itself (App. A.1) — and the appendix states the reporting scope "was chosen after inspecting earlier budget experiments, which showed that longer training can reverse comparisons." The abstract's "improves task scores over same-content distillation" therefore generalizes a result the authors' own archive contradicts at longer budgets. The seconds-scale installation savings are also immaterial next to the shared selection cost (thousands of seconds). And the mechanism that would justify "compilation" over ordinary distillation is never isolated: FMC is functionally a one-step closed-form distillation on self-generated data, and the single-gradient-step distillation baseline that would separate the two is absent.

W3 — Generalization and side effects are unmeasured. Much of KnowEdit "test" is paraphrase-level generalization of supervised items (most canonical addresses sit directly in the compiled pool). Behavior on unrelated prompts with the adapter active — precisely the risk of a head that boosts 250 answer strings' logits over the full vocabulary — is explicitly unmeasured. The MAB test set is 60 queries with no CI, LoCoMo F1 gaps carry no significance tests, and the two extra model pairs lack the Distilled baseline, so the central comparison exists for exactly one source–target pair.

Questions for authors

Major:

  1. Report the FMC vs. Distilled-LoRA comparison at the baseline's best budget (the archive experiments that reverse the comparison), and rewrite the abstract to match what survives.
  2. Evaluate the adapter-only serving regime (no context window, no BM25 + text path): can the compiled head answer the questions on its own? If not, the contribution reduces to head-only probe distillation matched by a text file.

Minor:

  1. Measure side effects: perplexity and instruction-following on generic prompts with the adapter active vs. inactive; add CIs for the MAB and LoCoMo gaps; report how many KnowEdit test paraphrases are near-paraphrases of compiled items.
  2. Add the single-gradient-step distillation baseline (matched recovered data, one SGD step) — if it matches FMC, "compilation" reduces to one-step distillation in closed form. Also compare against head + last-block LoRA to justify the head-only restriction.
5 comments
AuthorsOpenReview ↗

Response to Reviewer obb8 [1/3]

We thank the reviewer for the careful reading, and for the remarks that the abstraction is clean and the reporting transparent. We added a 2/4/8/16-epoch budget sweep on all three tasks, the same sweep on KnowEdit and MAB for the two additional model pairs, the adapter-only LoCoMo evaluation that the reviewer asks for, and side-effect measurements with the adapter active, and we rewrote the claims to match the new evidence. The contribution of the paper is now:

  1. Setting. Migrating an aggregate LoRA memory to a successor model when the interaction history is gone and only answer-free addresses remain.
  2. Method. One analytic output-head fit, with no backpropagation, optimizer steps, or learning-rate and epoch choices on the target.
  3. Evidence. In one fit, FMC reaches the score of 8-epoch distillation on KnowEdit and of 16-epoch distillation on MAB and on the admitted LoCoMo pool, while distillation takes 2.8–4.2× as long to fit at those budgets.

Each question is answered under the weakness it belongs to.

W1. On the LoCoMo readback result, the text-file control, and adapter-only serving.

Response. We thank the reviewer for naming the regime in which parametric storage could matter. We evaluated it, and the answer to Major question 2 is yes: the compiled head answers on its own (Table R1). The LoCoMo part of the revised Section 4.2 now leads with this result and not with the readback or the text control, with every budget in Appendix I.2 (Table 19).

Table R1. LoCoMo on the 250 admitted values. Adapter-only means that the target answers the 1,201 natural questions directly, with the stream's adapter active, empty context, and no index or reader.

StateLiteral exact readback (of 250)Adapter-only F1Shared-reader F1
Target Base13.784.10
Distilled LoRA, 2 epochs364.266.10
Distilled LoRA, 4 epochs696.456.84
Distilled LoRA, 8 epochs1567.028.27
Distilled LoRA, 16 epochs2137.338.70
FMC2507.198.74

Adapter-only F1 nearly doubles over Target Base, with no index and no stored text to consult. FMC obtains this in one fit, whereas distillation needs 8 to 16 epochs to reach the same level.

On whether the readback is forced: it is not forced by the fit. The ridge constraints do not set the logits. They are fitted jointly, in least squares and under a rank limit, so the requested logits are approximated, and exact decoding needs a sufficient margin at every prefix (Appendix D.2). Admission does select values that can be written one at a time, and Section 4.2 reports the result as conditional on it. What the result shows is that such values stay exact when installed jointly, which no evaluated distillation budget achieves.

Under the shared reader a text file matches the adapter, as the reviewer says, so we claim no superiority to text storage; Section 4.2 states this and cites the 250/250 readback with its condition. What LoCoMo does support is the thesis that a parametric memory can be carried into the successor's parameters with one analytic fit: the admitted multi-token values are installed jointly without loss, and the installed adapter can be used without an index.

AuthorsOpenReview ↗

Response to Reviewer obb8 [2/3]

W2. On the distillation budget, the cost claim, the relation to one-step distillation, and the head-only restriction.

Response. We ran the comparison at the larger budgets, the single-step baseline, and the head + last-block comparison, and we rewrote the abstract.

Budget (Major question 1). In one fit, FMC matches 8 epochs of distillation on KnowEdit and stays ahead of 16 epochs on MAB and on the admitted LoCoMo pool, in less time than any budget (Table R2).

Table R2. Budget sweep on the primary pair. Parentheses give the fixed-pool installation time, excluding shared preparation.

Task (metric)Distilled LoRA: 2 epochs4 epochs8 epochs16 epochsFMC
KnowEdit (accuracy, %)34.92 (5.10 s)43.75 (6.53 s)53.53 (9.47 s)64.54 (15.43 s)53.53 (3.44 s)
MAB (accuracy, %)60.00 (3.25 s)75.00 (4.15 s)85.00 (6.03 s)88.33 (9.74 s)90.00 (2.34 s)
LoCoMo (shared-reader F1)6.10 (7.44 s)6.84 (10.15 s)8.27 (15.30 s)8.70 (25.97 s)8.74 (6.30 s)

FMC is at or above distillation in 11 of the 12 cells. Distillation's score also depends strongly on the epoch budget, which has to be chosen on the target, whereas FMC has no such choice. Distillation is ahead on KnowEdit at 16 epochs, so the abstract no longer claims higher scores from a single budget. Tables 1 and 2 of the revised paper report the sweep, and the abstract's sentence on distillation now reads: "Across the evaluated memory tasks, FMC matches or exceeds same-content distillation in most training-budget settings and, given fixed content, reaches comparable scores with a 2.8–4.2× shorter fitting step."

Cost. On LoCoMo the selection stage dominates the end-to-end time, as the reviewer says: it takes 2,626.5 s, both methods are fitted on the pool it selects, and the end-to-end times are 2,632.8 s and 2,634.0 s. KnowEdit and MAB have no such stage (Section 3.3), so there the fitting step of Table 2 is the only target-side computation in which the methods differ. The revised Section 4.3 states the claim as a cheaper fitting step given fixed content.

Mechanism. We agree with the reviewer's description: FMC is response distillation whose installation step is solved in closed form, and Section 2.3 places it there. What the closed form computes is the converged solution of its objective, not the first gradient step. On the same objective, the iterative solver reaches the same 54/60 after 21.36 s of solving, against 0.01 s for the closed form (Table 12), the equivalence the reviewer notes among the strengths. A single gradient step omits the inverse Gram factor that separates correlated keys. Table R3 reports the requested single-step baseline, one SGD step on the matched recovered data.

Table R3. Single-step baseline (accuracy, %).

TaskTarget BaseOne SGD stepFMC
MAB23.3331.6790.00
KnowEdit10.8711.0153.53

Head-only restriction. The restriction is what permits the closed-form fit: with frozen features, collection and fitting separate (Section 3.3). For the requested comparison we trained LoRA matrices in the output head and in the last transformer block on the same recovered data (Table R4).

Table R4. Head + last-block LoRA against FMC (accuracy, %).

TaskHead + last block: 2 epochs4 epochs8 epochs16 epochsFMC
KnowEdit41.03 (21.28 s)52.17 (42.90 s)60.73 (85.08 s)67.80 (170.42 s)53.53 (3.44 s)
MAB86.6783.3390.0090.0090.00

On MAB the larger adapter reaches FMC's score at 8 and 16 epochs and does not exceed it. On KnowEdit it exceeds FMC from 8 epochs, at about 25× FMC's fitting time, because once the last block is trained the final hidden states can no longer be collected once and reused. The restriction is thus what buys an installation with no target backpropagation. Both baselines are in Appendix I.6 of the revised paper (Tables 22 and 23).

AuthorsOpenReview ↗

Response to Reviewer obb8 [3/3]

W3. On generalization, side effects, uncertainty, and the additional model pairs.

Response. We thank the reviewer for the specific list. Each item is now measured.

Paraphrase-level generalization. This is what KnowEdit measures here: Section 4.1 states that the target is evaluated on held-out paraphrases, not on new facts. To the reviewer's question: 450 of the 736 test paraphrases (61.14%) are near-paraphrases of a compiled item. They belong to the 225 test facts whose canonical address is in the compiled pool (Table 11), and each is an alias variant of that address: the template is the same and only the subject's name changes, for example to a short form. Table R5 refines Table 11 by character similarity to the compiled address.

Table R5. KnowEdit accuracy by similarity to the compiled address.

KnowEdit test paraphrasesCountFMC (%)Distilled LoRA, 2 epochs (%)
Compiled fact, similarity ≥ 0.9018257.1442.31
Compiled fact, similarity < 0.9026869.4047.01
Fact outside the compiled pool28636.3618.88

Accuracy does not rise with character overlap: FMC is not more accurate on the closer variants. The result supports transfer of the installed answers across subject aliases.

Side effects. We measured perplexity and instruction following with the MAB adapter active on inputs unrelated to the memory (Table R6).

Table R6. Side effects with the MAB adapter active.

StateMAB accuracy (%)PerplexityStrict instruction pass rate (%)
Target Base23.339.1688.28
Distilled LoRA, 2 epochs60.0029.0947.66
FMC, as submitted90.0066.0821.09
FMC with preservation term91.679.5757.81

The submitted adapter is not neutral on such inputs. The update changes only the output rows of the requested tokens (Section 3.4), but Eq. (1) fits their logits at the compiled positions alone and does not constrain them elsewhere. We therefore added one preservation term to the same fit, as MEMIT does for the keys it preserves (Meng et al., 2023): hidden states of generic text enter the ridge problem as additional rows with zero targets. The boosts are unchanged, and installation is still one closed-form fit with no backpropagation. With the term, perplexity returns close to Target Base, the pass rate rises by 36.72 points, MAB accuracy is preserved, and all three measures are better than for two-epoch distillation.

Uncertainty. The gaps to two-epoch distillation now have intervals on all three tasks, and none includes zero (Table R7).

Table R7. Intervals for the gaps to two-epoch distillation.

GapEstimate95% intervalBasis
MAB accuracy30.00 points[18.33, 41.67]paired bootstrap over the 60 queries (counts of Figure 3(b))
KnowEdit accuracy18.61 points[13.45, 23.51]paired bootstrap over 368 facts (Section 4.2)
LoCoMo F12.64 points[2.0, 3.4]bootstrap over the nine streams; FMC is above on all nine (exact two-sided sign test, p = 0.004; Table 13)

Other model pairs. The two additional model pairs of Section 4.5 show the pattern of the primary pair (Table R8): distillation passes FMC only on KnowEdit at 16 epochs. Over the primary and additional pairs, FMC is at or above distillation in 25 of 28 cells.

Table R8. The same sweep on the additional pairs (source Qwen3-1.7B).

Target modelTask (accuracy, %)Distilled LoRA: 2 epochs4 epochs8 epochs16 epochsFMC
Qwen3-4BKnowEdit19.5724.3239.9550.4146.60
Qwen3-4BMAB35.0053.3370.0076.6785.00
SmolLM3-3BKnowEdit27.1743.7552.3156.9355.43
SmolLM3-3BMAB20.0041.6770.0085.0085.00

In the revised paper these results are in Appendix I: the additional pairs in I.1 (Table 18), the intervals in I.3, the similarity split in I.4 (Table 20), and the preservation term in I.5 (Table 21).

We thank the reviewer again for the detailed review. All changes are marked in blue in the revised PDF, and we are happy to answer further questions.

Reviewer_a75UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The original history is gone, but the source model and answer-free questions are still available, and the question is how to move a trained aggregate LoRA memory to a new backbone. FMC reads out the source answers at retained addresses, picks some answer positions as logit targets on frozen target hidden states, and then installs an output-head LoRA with one reduced-rank ridge solve, without target backpropagation during installation. In the Qwen upgrade setting it beats Cross-LoRA, a Command-V port and a short-budget Distilled LoRA on KnowEdit, MAB and LoCoMo, and installation is also slightly faster.

Critical strengths

(1) The method is simple and cheap. One closed-form fit moves the source answers into the target's own space, and it pulls well ahead of the parameter-transfer baselines (Cross-LoRA, Command-V port), with the Cross-LoRA comparison also extending to the extra model pairs.

(2) The controls are very thorough: shuffled pairing, source-base ablation, in-pool versus out-of-pool, and a same-objective iterative solver, which help isolate several contributions within the pipeline. I appreciate this.

Critical weaknesses

(1) The advantage over Distilled LoRA is established only at the short training budget in Table 1. Appendix B states that this reporting scope was chosen after earlier experiments showed that longer training could reverse comparisons. The paper acknowledges this limitation, but a broader budget comparison is needed to assess how general the advantage is.

(2) Table 2 shows a real reduction in fixed-pool installation time, although the absolute savings are small and shared preparation costs are excluded. Appendix B.1 shows that LoCoMo pool preparation dominates these installation costs, so the end-to-end benefit is less clear. In the MAB solver control in Table 12, iterative ridge matches the analytic solution's accuracy at greater cost; this supports an efficiency benefit from the solver, while leaving the accuracy gain over CE distillation dependent on other design choices.

(3) The evidence supports preserving selected recovered values, but broader memory retention remains limited. Table 1 leaves a substantial KnowEdit gap to Source Memory, and LoCoMo scores remain low. The recovered-text control confirms preserved downstream utility under the shared reader, without showing a quality advantage over text storage. As Sections 4.5 and 5 acknowledge, unrelated-request behavior is unmeasured and matched distillation is evaluated only on the primary model pair.

Questions for authors

See Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer a75U [1/2]

We thank the reviewer for the careful assessment, and for the remarks that the method is simple and cheap and that the controls are thorough. We added a 2/4/8/16-epoch budget sweep on all three tasks, the same sweep on KnowEdit and MAB for the two additional model pairs, an adapter-only LoCoMo evaluation, and side-effect measurements, and we state the claims more narrowly. The contribution of the paper is now:

  1. Setting. Migrating an aggregate LoRA memory to a successor model when the interaction history is gone and only answer-free addresses remain.
  2. Method. One analytic output-head fit, with no backpropagation, optimizer steps, or learning-rate and epoch choices on the target.
  3. Evidence. In one fit, FMC reaches the score of 8-epoch distillation on KnowEdit and of 16-epoch distillation on MAB and on the admitted LoCoMo pool, while distillation takes 2.8–4.2× as long to fit at those budgets.

W1. On the training budget of the comparison with Distilled LoRA.

Response. We thank the reviewer for raising this. The budget comparison is now in the paper for all three tasks (Section 4.2, Table 1) and for the two additional model pairs (Appendix I.1, Table 18). On the primary pair (Table R1), FMC in one fit matches 8 epochs of distillation on KnowEdit and stays ahead of 16 epochs on MAB and on the admitted LoCoMo pool, in less time than any budget.

Table R1. Budget sweep on the primary pair. Parentheses give the fixed-pool installation time, excluding shared preparation.

Task (metric)Distilled LoRA: 2 epochs4 epochs8 epochs16 epochsFMC
KnowEdit (accuracy, %)34.92 (5.10 s)43.75 (6.53 s)53.53 (9.47 s)64.54 (15.43 s)53.53 (3.44 s)
MAB (accuracy, %)60.00 (3.25 s)75.00 (4.15 s)85.00 (6.03 s)88.33 (9.74 s)90.00 (2.34 s)
LoCoMo (shared-reader F1)6.10 (7.44 s)6.84 (10.15 s)8.27 (15.30 s)8.70 (25.97 s)8.74 (6.30 s)

The two additional model pairs of Section 4.5 show the same pattern (Table R2): distillation passes FMC only on KnowEdit at 16 epochs.

Table R2. The same sweep on the additional pairs (source Qwen3-1.7B).

Target modelTask (accuracy, %)Distilled LoRA: 2 epochs4 epochs8 epochs16 epochsFMC
Qwen3-4BKnowEdit19.5724.3239.9550.4146.60
Qwen3-4BMAB35.0053.3370.0076.6785.00
SmolLM3-3BKnowEdit27.1743.7552.3156.9355.43
SmolLM3-3BMAB20.0041.6770.0085.0085.00

Over both tables, FMC is at or above distillation in 25 of 28 cells. Distillation's score also depends strongly on the epoch budget, which has to be chosen on the target, whereas FMC has no such choice. Because distillation is ahead on KnowEdit at 16 epochs, we no longer claim higher accuracy than longer-trained distillation. The claim we keep is that one fit, with no epoch or learning rate to tune, reaches a score comparable to 8 to 16 epochs of distillation in only 24–36% of the fitting time (Table R1).

W2. On the size of the time saving and the source of the accuracy gain.

Response. Both observations are correct, and the sweep now puts numbers on them. In absolute terms the saving is small at this scale: 0.9 to 1.7 s against two-epoch distillation, and 6 to 20 s against the budgets at which distillation reaches a comparable score. In relative terms, distillation takes 2.8–4.2× as long at those budgets. Table 2 times the fitting step for a fixed pool. Shared preparation is common to both methods and lies outside it. The revised Section 4.3 states the claim as a cheaper fitting step given fixed content.

On the solver control, the reviewer's reading is the intended one (Section 4.3). On the same objective, the analytic and iterative solvers both reach 54/60, in 2.34 s and 24.02 s, so the solver accounts for cost, not accuracy. The accuracy difference from CE distillation comes from the objective and its training budget, and the MAB row of Table R1 shows how much of it is budget. What the design contributes is an objective with a closed-form optimum, which is solved in one step with no epoch budget to choose.

AuthorsOpenReview ↗

Response to Reviewer a75U [2/2]

W3. On the scope of the retention evidence and the two open measurements.

Response. Both measurements that were open are now done: matched distillation on the two additional model pairs (Table R2) and behavior on unrelated requests (Table R4). On the scope we follow the reviewer. The revised Section 5 states it as the claim: FMC installs a conditional output distribution over the retained addresses, and behavior beyond them is established only for the evaluated held-out paraphrases and queries.

KnowEdit gap. More than half of this gap is not specific to FMC. Only 247 of the 408 facts yield a stable answer at their answer-free address, and gradient training on the same recovered content also stays well below Source Memory (97.83%): 64.54% for 16-epoch distillation, and 67.80% when LoRA matrices in the last transformer block are also trained (the comparison requested by Reviewer obb8).

LoCoMo scores. FMC reaches 8.74 against 8.84 for Source Memory through the same reader, so the low absolute level is set upstream of migration, by the source memory and the reader. Within this range, shared-reader F1 tracks installation fidelity (Table R3).

Table R3. LoCoMo on the 250 admitted values.

StateLiteral exact readback (of 250)Adapter-only F1Shared-reader F1
Target Base13.784.10
Distilled LoRA, 2 epochs364.266.10
Distilled LoRA, 4 epochs696.456.84
Distilled LoRA, 8 epochs1567.028.27
Distilled LoRA, 16 epochs2137.338.70
FMC2507.198.74

Text storage. The recovered-text control gives the reader the exact strings an installer is asked to write, so it is the reference for lossless installation, and FMC is the only state that reaches it: it reads back all 250 admitted values, which distillation does not do at any evaluated budget. LoCoMo therefore compares installers with each other, not parametric memory with text, and we claim no advantage over text storage. What an adapter offers that a value table does not is use without an index. The adapter-only column tests this: the target answers the 1,201 natural questions directly, with the stream's adapter active, empty context, and no index or reader. F1 rises from 3.78 for Target Base to 7.19 for FMC, so the installed parameters carry the values to questions that are not retained addresses.

Unrelated requests. We measured this with the MAB adapter active on inputs unrelated to the memory (Table R4).

Table R4. Side effects with the MAB adapter active.

StateMAB accuracy (%)PerplexityStrict instruction pass rate (%)
Target Base23.339.1688.28
Distilled LoRA, 2 epochs60.0029.0947.66
FMC90.0066.0821.09
FMC with preservation term91.679.5757.81

The submitted adapter is not neutral on such inputs, because Eq. (1) does not constrain the requested logits outside the compiled positions. We therefore added one preservation term to the same fit, as MEMIT does for the keys it preserves (Meng et al., 2023): hidden states of generic text enter the ridge problem as additional rows with zero targets. The boosts are unchanged, and installation is still one closed-form fit with no backpropagation. With the term, perplexity returns close to Target Base, the pass rate rises by 36.72 points, MAB accuracy is preserved, and all three measures are better than for two-epoch distillation. In the revised paper, the LoCoMo results are in Section 4.2 and Appendix I.2 (Table 19), and the preservation term is in Appendix I.5 (Table 21).

We thank the reviewer again for the constructive review. All changes are marked in blue in the revised PDF, and we are happy to answer further questions.

Thank you for the detailed response. The full epoch budget sweep makes the comparison with Distilled LoRA much fairer. The side-effect measurements and the preservation term address my concern about unrelated requests. I will keep my score.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to Reviewer 5iH4 [1/2]

We thank the reviewer for the careful reading, including the appendices, and for the remarks that the problem is well posed and that the cost accounting is transparent. We added a 2/4/8/16-epoch budget sweep and an adapter-only LoCoMo evaluation, and we state the claims more precisely. The contribution of the paper is now:

  1. Setting. Migrating an aggregate LoRA memory to a successor model when the interaction history is gone and only answer-free addresses remain.
  2. Method. One analytic output-head fit, with no backpropagation, optimizer steps, or learning-rate and epoch choices on the target.
  3. Evidence. In one fit, FMC reaches the score of 8-epoch distillation on KnowEdit and of 16-epoch distillation on MAB and on the admitted LoCoMo pool, while distillation takes 2.8–4.2× as long to fit at those budgets.

W1. On the two-epoch distillation baseline and the missing budget sweep.

Response. We thank the reviewer for proposing the sweep, which we ran on all three tasks (Table R1). In one fit, FMC matches 8 epochs of distillation on KnowEdit and stays ahead of 16 epochs on MAB and on the admitted LoCoMo pool, in less time than any budget.

Table R1. Budget sweep on the primary pair. Parentheses give the fixed-pool installation time, excluding shared preparation.

Task (metric)Distilled LoRA: 2 epochs4 epochs8 epochs16 epochsFMC
KnowEdit (accuracy, %)34.92 (5.10 s)43.75 (6.53 s)53.53 (9.47 s)64.54 (15.43 s)53.53 (3.44 s)
MAB (accuracy, %)60.00 (3.25 s)75.00 (4.15 s)85.00 (6.03 s)88.33 (9.74 s)90.00 (2.34 s)
LoCoMo (shared-reader F1)6.10 (7.44 s)6.84 (10.15 s)8.27 (15.30 s)8.70 (25.97 s)8.74 (6.30 s)

FMC is at or above distillation in 11 of the 12 cells. At the budget closest to FMC's fitting time, two epochs, FMC leads on all three tasks. Distillation's score also depends strongly on the epoch budget, which has to be chosen on the target, whereas FMC has no such choice. Distillation is ahead in one place, KnowEdit at 16 epochs, so we no longer claim higher accuracy than longer-trained distillation, and the runs behind the reversal that Section 4.1 mentioned are now reported. The claim we keep is that one fit, with no epoch or learning rate to tune, reaches a score comparable to 8 to 16 epochs of distillation in only 24–36% of the fitting time (Table R1). The sweep is in Section 4.2 (Table 1) and Section 4.3 (Table 2) of the revised paper.

W2. On the equivalence to text storage on LoCoMo and the selected pool behind the 250/250 readback.

Response. The reviewer reads the text control correctly: under the shared reader, FMC and the recovered text give the same answers, and we claim no advantage over address-keyed retrieval. What LoCoMo tests is installation. Every state uses the same address index, so the comparison is between installers, and the recovered-text control is the reference for a lossless one, because it hands the reader the exact strings an installer is asked to write (Table R2).

Table R2. LoCoMo on the 250 admitted values.

StateLiteral exact readback (of 250)Adapter-only F1Shared-reader F1
Target Base13.784.10
Distilled LoRA, 2 epochs364.266.10
Distilled LoRA, 4 epochs696.456.84
Distilled LoRA, 8 epochs1567.028.27
Distilled LoRA, 16 epochs2137.338.70
FMC2507.198.74

FMC is the only state that reaches this reference. It reads back all 250 values, while distillation still misses 37 of them after 16 epochs, at four times FMC's fitting time. The readback is conditional on the admitted pool, as Section 4.2 states. Admission does not make it automatic, since each candidate is tested in isolation while the reported adapters are fitted jointly; what the result shows is that individually writable values stay exact when installed together. The adapter-only column, which uses no index, is discussed under W4. The revised Section 4.2 reports these results, with every budget in Appendix I.2 (Table 19).

AuthorsOpenReview ↗

Response to Reviewer 5iH4 [2/2]

W3. On the cost claim and the LoCoMo selection cost.

Response. We adopt the reviewer's wording, "a cheaper fitting step given fixed content". Table 2 times this step: encoding, representation collection, fitting, and saving. On KnowEdit and MAB there is no selection stage, because addresses are selected on the source side and no target trial writes are needed (Section 3.3), so the fitting step is the only target-side computation in which the two methods differ. LoCoMo is the exception. Its selection stage precedes fitting and takes 2,626.5 s, both methods are fitted on the pool it selects, and the end-to-end times are the 2,632.8 s and 2,634.0 s that the reviewer notes. The revised Section 4.3 uses this wording and reports the selection cost and these times next to Table 2.

W4. On keyed lookup versus migrated memory, and the wording of the claim.

Response. We thank the reviewer for the formulation, which we adopt. The revised abstract states that FMC installs a conditional output distribution over the retained addresses, and Section 5 adds that behavior beyond them is established only for the evaluated held-out paraphrases and queries. This matches what the paper sets out to migrate, the source memory's question–answer associations at the retained addresses (Section 1, first contribution).

One clarification on "lookup". The compiled memory is served with no index or key: KnowEdit and MAB use no retrieval, and every test input is a held-out paraphrase or a held-out query, not a retained address. The controls of RQ3 are measured on these inputs. They show that the score depends on the source memory's content (MAB: 90.00% with it, 20.00% without) and on its correct binding to the addresses (KnowEdit: 53.53%, versus 32.70% when shuffled), so the effect reaches nearby inputs through the target's own representations.

LoCoMo now tests the same point directly. In the adapter-only column of Table R2, the target answers the 1,201 natural questions with the stream's adapter active, empty context, and no index or reader, and F1 rises from 3.78 for Target Base to 7.19 for FMC. No stored text is consulted in this interface, so the gain comes from the installed parameters.

W5. On the absolute gaps and the target-written LoCoMo addresses.

Response. We take the three points in turn.

KnowEdit gap. More than half of this gap is not specific to FMC. Only 247 of the 408 facts yield a stable answer at their answer-free address, and gradient training on the same recovered content also stays well below Source Memory (97.83%): 64.54% for 16-epoch distillation, and 67.80% when LoRA matrices in the last transformer block are also trained (the comparison requested by Reviewer obb8).

LoCoMo range. FMC reaches 8.74 against 8.84 for Source Memory through the same reader, which is the available reference because we did not measure an oracle ceiling. Shared-reader F1 in this range tracks installation fidelity: it rises with exact readback down the rows of Table R2. The readback result itself does not depend on the reader.

Address writer. This concerns LoCoMo only; KnowEdit and MAB addresses come from the benchmarks. Qwen3.5-4B converts the public sessions into QA pairs before source training and never sees the official questions. The installed values are read from the Qwen3.5-2B source memory, and all compared states share the same addresses. The revised Section 4.1 describes this step.

We thank the reviewer again for the detailed review. All changes are marked in blue in the revised PDF, and we are happy to answer further questions.

Reviewer_5iH4OpenReview ↗

Response to author

Thank you to the authors for their detailed response to the review. Much of it has resolved my questions and addressed gaps in the paper. However, based on the analysis of the experimental results in Table 21 in particular, together with the overall experimental results, I remain unable to determine whether FMC is conditional in the asserted sense or a capability obtained at the expense of general capability. Accordingly, I maintain my original decision.

Thank you for clarifying the concern. Prior work has documented degradation of out-of-domain capabilities and instruction following under LoRA fine-tuning [1, 2]. Low-rank adaptation therefore does not by itself guarantee preservation of unrelated behavior. Similarly, storing knowledge through LoRA-based parametric updates does not automatically preserve behavior outside the intended memory scope; prior work introduces selective adapter activation to address this locality challenge [3].

In Table 21, FMC with preservation achieves higher MAB accuracy than the reported two-epoch distillation baseline (91.67% vs. 60.00%), alongside lower perplexity (9.57 vs. 29.09) and higher instruction pass rate (57.81% vs. 47.66%). This comparison shows improved memory performance together with better retention on both measured diagnostics.

The remaining instruction-following gap to Target Base is a limitation. Nevertheless, these results demonstrate that interference can be substantially mitigated while retaining the observed memory performance. We will clarify this scope and the meaning of “conditional” in Section 5.

References

[1] Yuheng Lu, Bingshuo Qian, Caixia Yuan, Huixing Jiang, and Xiaojie Wang. 2025. Controlled Low-Rank Adaptation with Subspace Regularization for Continued Training on Large Language Models. In Proceedings of ACL 2025.

[2] Gangwei Jiang, Zhaoyi Li, Defu Lian, and Ying Wei. 2024. Refine Large Language Model Fine-tuning via Instruction Vector. arXiv preprint arXiv:2406.12227.

[3] Lang Yu, Qin Chen, Jie Zhou, and Liang He. 2024. MELO: Enhancing Model Editing with Neuron-Indexed Dynamic LoRA. In Proceedings of the AAAI Conference on Artificial Intelligence.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.