acceptodds
Under review as a conference paper at ICLR 2027

Isolating the Training Objective in Test-Time Knowledge Incorporation

Abstract

Test-time training (TTT) can write a document into the adapter parameters of a language model with a few gradient updates, and recent methods improve this process with components that require training of their own, such as a meta-learned update, a teacher or a generator trained by reinforcement learning. All such methods share a training objective, and because it can be changed without additional training, what it achieves alone is the natural reference for these components. This reference has not been measured. In the standard setting, with LoRA applied to the query and value projections of Qwen2.5-7B, no objective significantly exceeds next-token training, yet with LoRA applied to all linear projections the best objective exceeds it by 18.5 points, so the placement can hide the effect of the objective. We therefore decompose the objective into four components and compare fourteen objectives that change one component at a time, on two backbones, two placements and two datasets. Shaping the loss rarely improves accuracy. Accuracy instead depends on how much distinct text the loss covers, and falls monotonically as passage tokens are removed from the loss. Guided by this finding, we propose MIX3, which supervises the full passage together with the model's own question-answer pairs and rewrites. In ten updates and without a trained generator, teacher or outer loop, it reaches 73.7% and 73.5% closed-book accuracy on Qwen2.5-7B and Llama-3.1-8B-Instruct, and on facts newer than the training data it improves retrieval while keeping judgment at the level of the unadapted model. MIX3 thus provides a reference for costlier components, and our results suggest that objectives should be compared under more than one placement.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Test Time Training, Large Language Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8JUGOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper isolates the effect of the training objective in test-time training to study how much each objective alone contributes to test-time knowledge incorporation. Evaluations are conducted on two backbones (Qwen2.5-7B, Llama-3.1-8B-Instruct), two LoRA placements (q/v projections vs. all-linear), and two datasets (SQuAD, WikiDYD). They conclude the contribution of each objective across different settings.

Critical strengths
  1. The paper conduct comprehensive experiments to study the contribution of different objectives.
  2. They provide solid experimental results on the effect of LoRA placements.
Critical weaknesses
  1. My largest concern on the paper is that the paper reads more like a technical report or an experiment report but not an academic paper. It lacks deep analysis into the experimental results or novel idea on model behaviors.
  2. Given different performance reported in the paper, experiments are not convincible enough on only two models and two datasets to derive general conclusion.
  3. MIX3 is clear but not novel.
  4. The paper only conduct experiments under different settings. It lacks theorectical support or study for the contribution.
Questions for authors

How do you use AI in your research, such as baseline surveys, experiments, analysis, writing, and so on?

3 comments

Thank you for your review and thoughtful comments!

W1. Depth of analysis

  • We added a control experiment that separates LoRA placement from adapter capacity in the revision, and state each finding together with the comparison that establishes it. Each placement is run at both parameter counts, paired on the same 200 passages:

Table. Closed-book accuracy on SQuAD, Qwen2.5-7B. † lr 1e-4, since lr 1e-3 collapses.

Adapter (trainable parameters)SYNSYN_poolMIXMIX3
q/v, rank 32 (10.1M)53.843.851.554.2
all-linear, rank 4 (10.1M)55.750.261.566.8
q/v, rank 256 (80.7M)47.3†41.558.463.7
all-linear, rank 32 (80.7M)55.2†44.964.873.7

With parameters fixed, changing placement raises MIX3 by 12.6 ±1.8 and leaves SYN unchanged. With placement fixed, 8× more parameters raise MIX3 by 9.4 ±2.0 and lower SYN by 6.5 ±2.0. So whether capacity helps depends on the objective. The same kind of control runs through the paper. Removing passage tokens from the loss of SYN lowers the gain steadily from 14.9 to −3.7, whereas removing question tokens from the pairs has no significant effect (Figure 1a, Appendix D.1).

W2. Generality from two backbones and two datasets

  • We state the claims for the evaluated settings, and within them the main result holds in every one. MIX3 exceeds MIX in all eight settings reported in Tables 2, 3 and 7, by 2.7 to 10.8 points. The two backbones, a base and an instruct model, disagree on secondary factors (70 updates instead of 20: +4.0 ±1.8 on Qwen, −3.9 ±1.8 on Llama) and still agree on the main result.

W3. Novelty of MIX3

  • We clarified MIX3 as the reference objective that the comparison produces in the revision, not as a new method, in the abstract and Section 1.

W4. Theoretical support

  • We designed the paper as an empirical study. We use the formal part to define the experimental design space so that any two objectives differ in named components, and we report paired accuracy differences on the same documents for the main retrieval comparisons.

Q. Use of AI

  • We expanded the AI use statement in the revision. We used generative AI as a writing assistant to polish the manuscript and improve its clarity and readability. In the experiments, LLMs appear only as part of the method under study: each backbone generates its own adaptation material from the prompts of Zweiger et al. (2025), and GPT-4.1 judges whether a closed-book answer matches the reference. We formulated the research question and decomposed the objective, designed the experiments, implemented them by modifying the code released by Zweiger et al. (2025), ran and analyzed them, and drew the conclusions ourselves.

Changes in the paper. We describe MIX3 as a reference objective in the abstract and Section 1, add the placement control to Section 5.2, and report factual and counterfactual judgment separately in Table 8.

Thank you for your response. I will maintain my ratings for the submission.

Thank you for reading our response and for letting us know! If any other questions remain, we are happy to address them.

Reviewer_i2a8OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the extent to which test-time knowledge integration can be achieved by adapting objectives (such as training sequences, loss masks, etc.) without training additional components.

The main findings are that the ranking of the objective function is strongly dependent on the adaptation configuration, and that supervising the entire paragraph along with additional self-generated text is generally more effective than the loss-shaping variants tested. Additional experiments examined token budgeting, multi-document adaptation, etc. Results from WikiDYK further demonstrate that improvements in fact retrieval do not necessarily lead to improvements in assertion judgment.

Critical strengths
  1. The paper clearly separates and evaluates training material, loss masks, loss shaping, and update schedules. This clarifies which design choices affect knowledge incorporation.

  2. Consequential configuration dependence. The substantial change in objective rankings across LoRA configurations shows why evaluation under a single configuration can be misleading.

  3. Simple baseline and careful protocol control. MIX3 achieves strong accuracy without a separately trained component. Further experiments such as token-budget controls, question-overlap checks, etc., strengthen the empirical evidence. Moreover, correcting the padding behavior makes the SYN comparison more credible.

Critical weaknesses
  1. LoRA placement is confounded with adapter capacity. Moving LoRA from qv to all linear layers increases the number of trainable parameters. Thus, the experiment cannot separate whether the performance improvement comes from module placement, additional capacity, or both. More experiments may strengthen the contribution.

  2. Interpretations based on different texts are only partially isolated. Adding QA pairs and rewriting simultaneously alters the text quantity, factual coverage, etc. The control group for token matching is valuable, but matching processed tokens does not necessarily match supervised tokens since QA questions are masked from the loss.

  3. The claim of preserving judgement overlooks a significant degradation on contradictory statements. MIX3 maintains overall judgment accuracy, but its accuracy on contradictory statements decreases from 32.5 to 17.5 on Qwen and from 35.5 to 27.0 on Llama. Therefore, the overall metric masks a substantial loss in rejecting false claims.

Questions for authors
  1. Do you plan any experiments to disentangle the contributions of LoRA placement and adapter capacity?

  2. In the token-matched control experiment, how many tokens receive supervision in each condition? Does the advantage of broader material coverage also come from supervised-token exposure and factual answer coverage, rather than the processed-token counts alone?

3 comments

Thank you for the positive comments and careful review!

W1 / Q1. Separating LoRA placement from adapter capacity

  • We ran the experiment: we changed the placement with the parameter count fixed, and changed the parameter count with the placement fixed. On Qwen2.5-7B, LoRA on q/v at rank 256 has the same number of trainable parameters as LoRA on all linear projections at rank 32 (80.7M), and all-linear at rank 4 has the same as q/v at rank 32 (10.1M). The table reports all four combinations on the same 200 passages.
Adapter (trainable parameters)SYNSYN_poolMIXMIX3
q/v, rank 32 (10.1M)53.843.851.554.2
all-linear, rank 4 (10.1M)55.750.261.566.8
q/v, rank 256 (80.7M)47.3†41.558.463.7
all-linear, rank 32 (80.7M)55.2†44.964.873.7

Closed-book accuracy on SQuAD, Qwen2.5-7B. † lr 1e-4, since lr 1e-3 collapses.

  • Both factors contribute, and placement contributes more. With parameters fixed at 10.1M, changing q/v to all-linear raises MIX3 by 12.6 ±1.8 and MIX by 10.0 ±1.8. With placement fixed at q/v, raising the parameters 8× adds less, MIX3 +9.4 ±2.0 and MIX +6.9 ±2.0.
  • Objectives that supervise only the passage gain little from either change. SYN and SYN_pool stay between 41 and 56 in all four configurations. Moving to all-linear leaves SYN unchanged (+2.0 ±2.0), and adding parameters lowers both (SYN −6.5 ±2.0 under q/v, SYN_pool −5.3 ±2.3 under all-linear).

Changes in the paper. Section 5.2 adds this table (Table 4) and the two conclusions above.

W2 / Q2. Supervised tokens and answer coverage in the token-matched control

  • The supervised token counts are below, and more supervised tokens do not explain the advantage of MIX3. Supervised tokens per document and epoch, with processed tokens in brackets: SYN 373 (381), MIX 556 (694), MIX3 1,202 (1,345), MIX3 token-matched 631 (690). Because the question in each QA row is masked, the token-matched version supervises 14% more tokens than MIX, and it is still not better than MIX (−2.4 ±1.6 under q/v, −6.6 ±1.9 under all-linear).
  • Counting supervised tokens times updates gives the same answer. MIX3 for five epochs supervises 1,202 × 5 = 6,010 token-updates, the token-matched version for ten epochs 631 × 10 = 6,310, and the five-epoch run is 10.3 points higher under all-linear (68.5 vs 58.2). Under q/v both fall below MIX (43.3 and 49.0), which is the adapter plateau of Section 5.2.
  • Answer coverage does not explain it either, because it is 100% for every objective that supervises the passage. SQuAD answers are spans of the passage. Only the QA-only objectives, which never see the passage, have partial coverage (on Qwen2.5-7B the reference answer appears in 73.4% of the pairs, Appendix C.2), and they are the weakest group in Table 2.

Changes in the paper. Table 6 adds a supervised-token column.

W3. "Preserving judgment" and the drop on contradictory statements

  • Thank you for this point. The revision reports factual and counterfactual accuracy separately. The split for the main objectives:
Factual / counterfactual accuracyQwen2.5-7BLlama-3.1-8B-It
Unadapted62.5 / 32.576.0 / 35.5
MIX385.5 / 17.585.0 / 27.0
SYN76.0 / 31.042.0 / 24.5
Open-book79.5 / 97.584.0 / 98.0

Placements and learning rates follow Table 8.

  • This does not change the comparison between objectives, because the drop on contradicting statements appears for almost every trained objective, not only MIX3. On these statements all 14 objectives fall below the unadapted model on Llama (16.5 to 30.0 against 35.5), and 13 of 14 do on Qwen (6.0 to 31.0 against 32.5, with SYN_pool at 36.5). We state it as a limitation shared by almost all objectives rather than as something MIX3 preserves.
  • It supports the point of Section 5.7 that retrieving a fact does not establish judging claims about it. With the document in context, the same model rejects 97.5% and 98.0% of the false statements, so the ability is lost in writing the fact into the weights, not absent from the model.

Changes in the paper. Table 8 adds the factual column. The abstract, Section 5.7 and the conclusion replace "preserves judgment" with the two numbers and list the shift as a limitation shared by almost all objectives.

I appreciate the authors for the detailed reply.

  1. W1/Q1: My concerns have been addressed through the control variable experiments.

  2. W2/Q2: The experiments have partially resolved my concerns. token-matched MIX3 receives more supervised tokens than MIX but underperforms MIX, indicating that more supervised tokens do not lead to better performance. However, the added controls strengthen the evidence, but 100% answer coverage does not mean comparable supervision of key facts. QA pairs and rewriting may also repeat/emphasize particular answers. Thus, diversity and factual supervision effects remain only partially disentangled.

  3. W3/Q3: This has resolved my original concerns.

I will maintain my weak acceptance rating.

We really appreciate your quick reply and your confirmation that W1 and W3 are resolved! On W2, we agree that the pairs and rewrites may repeat particular answers. We have stated this more clearly in Section 5.3 of the revision and mentioned the RQA–PQA comparison (Section 5.5) as partial evidence for this point. Thank you again for the careful and constructive review!

Reviewer_cAe3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper investigates the effect of different training objectives in test-time knowledge incorporation. The authors formalize the training objective into four components and compare 14 objectives across two language models, different LoRA placements, and two datasets. The main observation is that supervising more distinct text has a larger impact than loss shaping, and that LoRA placement can significantly change the results. Based on these observations, the authors propose MIX3, which combines the original passage, generated question-answer pairs, and rewrites.

Critical strengths

The paper is exceptionally well presented. The formalization of the objectives is clear and detailed, making the paper easy to follow and the experiments easy to understand.

The extensive ablation studies are the main strength of the paper. The authors systematically investigate different components and provide useful empirical observations regarding training objectives and LoRA configurations.

Critical weaknesses

The main concern of this reviewer is the novelty and significance of the contributions. While the experimental analysis is carefully conducted and well presented, most of the main observations have already been investigated in previous literature or are relatively straightforward.

In particular, combining original documents with question-answer pairs and alternative formulations has been extensively investigated in previous works [1–4]. The proposed MIX3 is a straightforward combination of existing approaches, and the observation that more diverse supervised text improves knowledge incorporation is already well established.

Similarly, applying LoRA to all linear projections and the performance advantages of adapting additional modules have been investigated in previous works [5, 6]. While the interaction between LoRA placement and training objectives is an interesting empirical observation, the experiments do not separate the effect of placement from the increased number of trainable parameters.

Overall, this reviewer finds the empirical analysis useful and well presented, but the novelty and significance of the contributions insufficient for ICLR in the manuscript's current form.

Please also refer to the questions below.

[1] Zeyuan Allen-Zhu and Yuanzhi Li. "Physics of Language Models: Part 3.1, Knowledge Storage and Extraction." Proceedings of the 41st International Conference on Machine Learning (ICML), Proceedings of Machine Learning Research, vol. 235, pp. 1067–1077, 2024. https://proceedings.mlr.press/v235/allen-zhu24a.html

[2] Oded Ovadia, Menachem Brief, Moshik Mishaeli, and Oren Elisha. "Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs." Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 237–250, 2024. https://doi.org/10.18653/v1/2024.emnlp-main.15

[3] Zhengbao Jiang, Zhiqing Sun, Weijia Shi, Pedro Rodriguez, Chunting Zhou, Graham Neubig, Xi Victoria Lin, Wen-tau Yih, and Srinivasan Iyer. "Instruction-tuned Language Models are Better Knowledge Learners." Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), vol. 1, pp. 5421–5434, 2024. https://doi.org/10.18653/v1/2024.acl-long.296

[4] Daixuan Cheng, Yuxian Gu, Shaohan Huang, Junyu Bi, Minlie Huang, and Furu Wei. "Instruction Pre-Training: Language Models are Supervised Multitask Learners." Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 2529–2550, 2024. https://doi.org/10.18653/v1/2024.emnlp-main.148

[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. "QLoRA: Efficient Finetuning of Quantized LLMs." Advances in Neural Information Processing Systems (NeurIPS), vol. 36, 2023. https://proceedings.neurips.cc/paper_files/paper/2023/hash/1feb87871436031bdc0f2beaa62a049b-Abstract-Conference.html

[6] Dan Biderman, Jacob Portes, Jose Javier Gonzalez Ortiz, Mansheej Paul, Philip Greengard, Connor Jennings, Daniel King, Sam Havens, Vitaliy Chiley, Jonathan Frankle, Cody Blakeney, and John Patrick Cunningham. "LoRA Learns Less and Forgets Less." Transactions on Machine Learning Research (TMLR), 2024. https://openreview.net/forum?id=aloEru2qCG

Questions for authors
  • The benefits of combining original documents with question-answer pairs, paraphrases, and diverse supervised text have already been investigated in previous works [1–4]. What is the main novel contribution of MIX3 beyond combining these existing approaches in the test-time adaptation setting? The observation that more diverse supervised text improves knowledge incorporation seems rather expected.

  • Applying LoRA to all linear projections instead of only attention projections has already been investigated [5, 6]. Furthermore, changing from query/value projections to all linear projections substantially increases the number of trainable parameters. How can the authors distinguish the effect of placement from increased adapter capacity? Would the same observation hold with a comparable number of trainable parameters? What is the main novel insight regarding LoRA placement beyond the previous works?

  • On WikiDYK with Qwen2.5-7B under the default q/v placement, MIX3 achieves only 17.2% accuracy compared with 50.8% for SYN. This is a substantial degradation despite the additional supervised text. How do the authors explain this observation, and how does it affect the generality of the proposed method and the main conclusions?

3 comments

Thank you for the detailed review and the references provided.

Q1. Novelty of MIX3 beyond combining [1–4]

MIX3 is not our main contribution. The contribution is the comparison: 14 objectives under the same material and adapter, paired on 974 questions, in a setting [1–4] do not study (one passage, ten updates, LoRA, no trained component). MIX3 is the simplest objective that follows what this comparison finds, and the revision calls it a reference objective, not a proposed new method.

Three results of this comparison are not in [1–4]:

  • More diverse text does not always help. Under q/v on Qwen no objective significantly exceeds SYN. Under all-linear, MIX3 gains 19.5 and SYN 1.4 (Table 3). Only objectives that mix kinds of text turn added capacity into accuracy (see Q2).
  • The gain is not from seeing more text. When MIX3 is subsampled to the token count of MIX, it is no better than MIX. MIX3 on its full material for five epochs, which processes approximately the same total tokens as that cut-down version, is 10.3 points higher under all-linear (Table 6). What matters is the amount of distinct text per update.
  • Nine of the ten loss-shaping variants on SQuAD (five variants on two backbones) do not improve significantly on their source objective, and one loses 21 to 45 points (Table 2).

Changes in the paper. Abstract and Section 1 call MIX3 a reference objective. Section 2 cites [3] and [4].

Q2. Placement vs. parameter count, and the insight beyond [5, 6]

  • We ran the experiment that separates placement from parameter count. Each placement is run at both parameter counts: q/v at rank 256 has the parameters of all-linear at rank 32 (80.7M), and all-linear at rank 4 those of q/v at rank 32 (10.1M). The table below reports all four configurations.
  • The observation holds at the same number of parameters. With parameters fixed at 10.1M, changing q/v to all-linear raises MIX3 and MIX (+12.6 ±1.8 and +10.0 ±1.8) and leaves SYN unchanged (+2.0 ±2.0). With placement fixed at q/v, raising the parameters 8× adds less (MIX3 +9.4, MIX +6.9) and lowers SYN (−6.5). Placement contributes more than parameter count and independently of it.
  • Beyond [5, 6], the gain from all-linear depends on the objective. Passage-only objectives stay between 41 and 56 in every configuration. Objectives that mix kinds of text gain 10 to 20 points. Rankings therefore change with placement (SYN_pool − SYN on Llama: +7.2 under q/v, −11.0 under all-linear), and comparing objectives under one placement can rank them wrongly.
Adapter (trainable parameters)SYNSYN_poolMIXMIX3
q/v, rank 32 (10.1M)53.843.851.554.2
all-linear, rank 4 (10.1M)55.750.261.566.8
q/v, rank 256 (80.7M)47.3†41.558.463.7
all-linear, rank 32 (80.7M)55.2†44.964.873.7

Closed-book accuracy on SQuAD, Qwen2.5-7B, paired on the same 200 passages. † lr 1e-4, since lr 1e-3 collapses.

Changes in the paper. Section 5.2 reports this table (Table 4) and states that placement and parameter count both contribute, and that objectives that supervise only the passage stay between 41 and 56.

Q3. WikiDYK Qwen q/v: MIX3 17.2 vs. SYN 50.8

  • The 17.2 comes from the rank-32 q/v adapter with ten updates, not from the material. SYN takes one update per sequence, about 70 per document. MIX3 pools its sequences into one batch and takes 10.
  • We ran two controls. Giving MIX3 the updates brings it from 17.2 to 60.5, and giving it the capacity brings it to 53.5, both above SYN, which itself falls to 25.0 at rank 256:
WikiDYK, Qwen2.5-7B, q/vUpdates (SYN)MIXMIX3SYN
rank 32 (paper)10 (70)16.017.250.8
rank 32, 70 epochs70 (70)46.260.550.8
rank 25610 (70)46.253.525.0†
all-linear, rank 32 (Table 2)10 (70)51.858.547.2†

† lr 1e-4, since lr 1e-3 collapses.

  • It does not change the conclusions. In every setting where the adapter is not the limit, the best objective supervises all three kinds of material, and on SQuAD no objective significantly exceeds MIX3 under either placement on either backbone (Tables 2, 3, 7). The one setting where MIX3 falls, Qwen under q/v on WikiDYK, is the same adapter plateau as on SQuAD, and it is why we recommend comparing objectives under more than one placement.

Changes in the paper. Section 5.2 adds this table (Table 5).

Reviewer_cAe3OpenReview ↗

Thank you for the comment. I will keep my rating.

Thank you for the comment. I will keep my rating.

Thank you for taking the time to go through our response! Your comments helped us improve the paper. If there is anything else you would like us to clarify, we are glad to do so.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers and the AC

We thank all the reviewers for their careful reading and thoughtful comments. We answer every weakness and question in the three reviews below, with an experiment or a change to the manuscript. The table summarizes what we did and where the changes are in the revised PDF.

ReviewerConcernWhat we didWhere in the revision
cAe3, 8JUGNovelty of MIX3We now describe MIX3 as the reference objective produced by the comparison, not as a new method, and cite Jiang et al. (2024) and Cheng et al. (2024).Abstract, Sections 1 and 2
cAe3, i2a8LoRA placement confounded with parameter countWe trained each placement at both parameter counts. With parameters fixed, changing placement raises MIX3 by 12.6 ±1.8. With placement fixed, 8× parameters raise it by 9.4 ±2.0. Passage-only objectives stay between 41 and 56 in all four configurations.Section 5.2, Table 4
cAe3WikiDYK Qwen q/v: MIX3 17.2 vs SYN 50.8We matched the update count and the capacity. With 70 updates, MIX3 rises from 17.2 to 60.5, and with rank 256 to 53.5, both above SYN.Section 5.2, Table 5
i2a8Supervised tokens and answer coverageWe counted supervised tokens: SYN 373, MIX 556, MIX3 1,202, MIX3 token-matched 631. The token-matched version supervises 14% more than MIX and is not better. Coverage is 100% for every objective that supervises the passage.Table 6
i2a8"Preserving judgment" hides the drop on contradictory statementsWe report factual and counterfactual accuracy separately and state the drop on counterfactual statements, shared by almost every trained objective, as a limitation.Abstract, Section 5.7, Table 8
8JUGDepth of analysisWe state each finding with the corresponding controlled comparison, including the placement control above.Sections 5.2 to 5.7
8JUGGenerality from two backbones and two datasetsWe state the claims for the evaluated settings. The three-kind objective exceeds the two-kind one in all eight of them, by 2.7 to 10.8 points.Sections 5.2, 5.5
8JUGTheoretical supportWe present the paper as an empirical study. The formal part defines the design space so that any two objectives differ in named components.Section 3
8JUGUse of AIWe expanded the AI use statement in the revision: generative AI was used to polish the language, and LLMs appear in the experiments only as the material generator and the GPT-4.1 judge.AI use statement

Changes in the revised PDF

New and revised text is shown in blue.

  • Abstract and Section 1 describe MIX3 as a reference objective and replace "preserves judgment" with the factual and counterfactual numbers.
  • Section 2 cites Jiang et al. (2024) and Cheng et al. (2024).
  • Section 5.2 adds Table 4 (placement × parameter count) and Table 5 (WikiDYK on Qwen2.5-7B under q/v with 70 updates and with rank 256).
  • Table 6 adds a supervised-token column. Table 8 adds a factual column, and Section 5.7 states the shift toward acceptance as a limitation shared by almost all objectives.

The additional results support our findings. Two statements have been made more precise: the gain under all-linear stems from both placement and parameter count, and adaptation increases acceptance of true statements while reducing rejection of false ones.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.