acceptodds
Under review as a conference paper at ICLR 2027

Query-Side Gains and Accumulation Limits in Online In-Weight Memory

Abstract

We study when online LoRA writes into a frozen 9B chat model yield usable memory, measuring readout, selection, and application against frozen controls. Under per-task isolation, InMind reaches 0.424 indirect application versus 0.312 for its frozen question comparator (historical judge protocol, not the new local judge). A pre-registered three-seed rerun meets its criterion under the primary judge; historical second-judge parsing failures leave that replication unresolved. One-shot transfers retain the margin on fresh tasks and PrefEval. A prespecified repair adds a same-inference-budget frozen hypothetical-key front end (0.328 application against 0.432) and an empty-query ablation of the same adapters (0.496 retrieval against 0.584): the application estimate is positive (+0.104 at n=125) but imprecise, and neither primary test is significant, raw or Holm-corrected. Historical coherent accumulation retrieves 0.192 versus 0.178 for frozen questions, despite a 0.705 stem-hit recall proxy. A new same-adapter control (20 personas, 146 queries) retrieves 0.212 with the correct query versus 0.082 with an empty one (equal-persona difference 0.147, conversation-cluster sign-flip p=0.0078): question conditioning within the accumulated trained state, but low absolute retrieval and no frozen arm. Paraphrase writes reverse a likelihood index's preference from exact strings to held-out paraphrases, a replicated write-format effect. A fixed-snapshot Mem0 repair isolates a query intervention (0.024 to 0.232, local Qwen judge), with ingestion workloads unmatched. Writes buy query-side readout under isolation; usable memory under accumulation is not established.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
test-time continual learning, online learning, parametric memory, LoRA, retrieval-augmented generation, personalization, long-term memory for LLMs, catastrophic forgetting, memory benchmarks, pre-registration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VnSsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines when online LoRA updates can act as usable memory for a frozen LLM. The proposed pipeline uses an online-updated adapter to generate queries for an external text log, leaving the final question answering to the frozen base model. Across evaluations of recognition, direct recall, retrieval, and downstream use, the results show that while the approach works for isolated fact writes, it breaks down as new memories accumulate.

Critical strengths
  • The paper is well-organized, with detailed explanations of the preliminary, intuition, and methodology.
  • The paper separates likelihood preference, direct recall, retrieval, and downstream application, preventing proxy gains from being conflated with truly usable memory.
  • The paper evaluates its findings across multiple benchmarks and experimental settings, providing useful insights into both the benefits and limitations of online in-weight memory.
Critical weaknesses
  • Giving the direct-QA supervision and ~8x more gradient updates than background turns essentially tells the adapter which fact will be tested. However, an empty query still achieves 0.496 target retrieval compared to 0.584 with the true query as reported in Table 2, suggesting the model mostly memorizes the most emphasized fact rather than learning query-dependently. The authors should test a balanced setting where multiple facts receive equal supervision and the model must selectively retrieve among them.

  • The LoRA adapter is trained on additional QA pairs and paraphrases that are absent from the frozen retrieval index. While HYP1 matches inference compute, it does not match this ingestion-time data advantage, making it unclear whether the gains come from parametric memory or simply better retrieval keys.

  • The accumulated-state evaluation compares valid and absent queries within trained adapters but omits a frozen baseline and downstream metrics; beating an empty query only shows basic query conditioning, not a benefit from weight updates. Moreover, comparing isolated and accumulated results shifts the underlying task population alongside the memory load, confounding the cause of the performance drop.

  • The apparent dissociation between recognition, recall, and application may be an artifact of noisy measurement rather than true retrieval failure. Specifically, likelihood gains can reflect surface template matching, recall relies on a brittle five-character stem-hit proxy, and the automated application judge has documented parsing failures without human validation. Semantically validated recall and a blinded human audit are needed to confirm whether knowledge is genuinely retained when access fails.

Questions for authors
  • How does the pipeline compare with a frozen retriever that indexes the exact direct-QA supervision used for LoRA training?
  • Can you add a frozen baseline and evaluate final answers in the new accumulation experiment?
Reviewer_stSCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies whether online LoRA updates can make stored facts useful in later interactions. The updated model generates retrieval queries over a text log, while a frozen model answers using the retrieved content. The experiments find gains when facts are written in isolation, but limited evidence of an advantage as memories accumulate. The paper also examines how write format and readout choices affect access to stored information.

Critical strengths
  1. The paper makes a useful distinction between storing information in weights and being able to retrieve and apply it. Testing these separately exposes failures that direct recall alone would miss, which matters for evaluating whether in-weight memory is actually useful.
  2. I appreciate the effort to challenge the initial positive results with stronger frozen controls, query ablations, and accumulation experiments
Critical weaknesses
  1. My main concern is that the model may just be repeating the most heavily trained fact rather than selecting a memory based on the query. The target fact gets multiple updates and direct-QA supervision, while background facts are trained only once. Whether the method is easy to use in practice is still unclear.
  2. My other concern is whether the gains justify the cost of online training. Each write requires gradient updates and recall checks, but the paper does not establish a clear end-to-end advantage over a frozen query-generation baseline that avoids this cost.
Questions for authors
  1. Does the retrieval gain remain when target and background facts receive the same training treatment, including update counts and QA supervision?
  2. How much additional write-time and query-time cost does the method incur over the frozen query-generation baseline, and what performance gain does that extra cost buy?
  3. Does the approach still work when trained only on the incoming user messages, without target-specific QA supervision or knowing which facts will be queried later?
Reviewer_eydPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies if LoRA writes into a frozen 9B model (Qwen3.5-9B) produce usable memory. In the main pipeline, the adapter generates a "best guess about the user" statement, embedded to retrieve from the log. The frozen base model then answers from the retrieved lines. On InMind under per-task isolation, a pre-registered three-seed rerun gives a median indirect-application score of 0.424 vs. 0.312. Under coherent accumulation (20 LoCoMo personas), written vs. frozen retrieval is 0.192 vs. 0.178. The authors offer several additional experiments and analysis.

Critical strengths

Exceptional methodological transparency. The experiment details are documented. This makes the empirical claims unusually trustworthy, and the paper is honest about what is not established (Conclusion: "These results do not demonstrate a superior deployed memory system"). This is consequential because it lets readers calibrate the evidence precisely.

Critical weaknesses
  1. Unclear scientific values. The paper focuses on a specific 9B model, and documents a range of experiments around the model. However, it is unclear what the scientific question is, not to mention there is no useful conclusion too.

  2. The positive claimed finding is narrow. Under per-task isolation, the target fact receives ~8 updates against the benchmark-supplied direct question. A deployed system would not know which turn will matter or what its direct question is. Hence "the write supplies fact-bearing query text" is close to expected by construction, and the empty-query ablation (0.584 vs. 0.496, nonsignificant) suggests a large part of the effect may be query-independent target salience—as the authors acknowledge. The prespecified primary tests against a stronger frozen comparator and against the empty query are both nonsignificant (Table 2), and in the realistic accumulated setting there is no detectable written-vs-frozen advantage (0.192 vs. 0.178). This weakness is critical because it determines whether the paper establishes anything about usable in-weight memory beyond the oracle-salience isolation setting.

  3. Poor presentation The paper uses a large non-standard vocabulary ("certification dose", "recall dose", "cashes", "chatter", etc). It also compresses 13+ exploratory chains and 10 later additions into a dense ladder (Table 1). It defers important design details (the write recipe, the controls, the repair protocol) to the appendix. As a result, the core experimental design and takeaway are difficult to extract even after multiple readings.

Questions for authors
  1. What happens under a uniform-dose write where the target fact receives the same number and form of updates as background turns, with no benchmark-supplied direct question? Does the written pipeline still beat the frozen question front end?
  2. The frozen comparators do not include ingestion-side indexing (e.g., generating hypothetical questions per log line at write time, or multi-query fusion), which use the same "writing" budget on the text side. Could you clarify whether any text-side indexing baseline with comparable compute was tested?
  3. For the 100 fresh tasks, the generator and judge come from the same model family as the AI that assisted the research. Is there any human validation, even on a small slice (e.g., 50 items), of judge verdicts?
  4. Under the confirmed recipe, how does the cost (≈28 s of gradient steps per fact plus gate generations) compare with simply adding the fact to the text store with a better query front end?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.