acceptodds
Under review as a conference paper at ICLR 2027

PRISM: Progressive Resolution with Independent Section Mapping for Scalable User Profiling

Abstract

Personalizing a language model over a user's long interaction history usually means either injecting the raw history into every prompt or compressing it once into a static persona. Both scale poorly: raw injection spends input tokens linear in history length, and a static persona re-generates every attribute for every query. We instead model the profile as an infinite memory: a hierarchy in which short-horizon episodic memories are compressed from narrow raw windows, and long-horizon memories are compressed from those episodic memories rather than from the raw history—so the input to any level is bounded by the level below it, and the memory absorbs unbounded history without ever feeding it all into a single prompt. We realize this memory with PRISM (Progressive Resolution with Independent Section Mapping), combining (i) incremental hierarchical compression of episodic into long-term memory, and (ii) a Map–Reduce decomposition that generates each memory level's independent semantic attribute sections in parallel, so the autoregressive generation depth is set by the longest section rather than the sum, and a broken section regenerates alone. Every efficiency result is reported in tokens, independent of any specific hardware. PRISM keeps the input tokens to refresh memory flat ( tokens) as history grows, yielding a 63 input reduction at events over monolithic regeneration, and cuts autoregressive generation depth by at sections. Crucially, on a paired longitudinal cohort ( accounts over 13-week histories), section decomposition significantly improves profile completeness () and specificity () under Holm-Bonferroni correction while preserving global consistency ( vs. ) and establishing rigorous non-inferiority in deterministic episodic fact retrieval (). On downstream tasks (LaMP), PRISM matches or exceeds monolithic prediction accuracy. We release anonymized code and the LongProfile benchmark for long-horizon personalization.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM personalization, compositional memory, multi-resolution representation, knowledge distillation, user profiling, memory-augmented models, temporal aggregation, inference efficiency

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 2 reviews, above 0% of the others.
Reviewer_6ZQEOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on addressing the challenge of personalizing LLMs when handling long-term user interaction history. This work proposes PRISM (Progressive Resolution with Independent Section Mapping), which combines incremental hierarchical compression of episodic into long-term memory and a Map--Reduce decomposition that generates each memory level's independent semantic attribute sections in parallel.

Critical strengths
  1. PRISM stores user profiles as an incrementally compressed hierarchy in which short-term episodic memories feed into long-term memory.
  2. The authors provide a hardware-independent evaluation and release a reproducible benchmark for long-horizon personalization.
Critical weaknesses
  1. The paper is hard to follow. There are no figures at all, not even a method overview. Adding an overview diagram and concrete input/output examples would greatly improve clarity.
  2. Essential details are missing. For example, the “frontier teacher” is never identified (which model, which version, what decoding settings), even though it is central to the results.
  3. LongProfile appears to be a simple merge of several existing datasets.
Questions for authors

Please refer to the weaknesses detailed above.

1 comment

Thank you for your constructive feedback and insightful suggestions! Below, we address each of your points in detail:

Method overview diagram: We added a method overview. Figure 1 shows the full pipeline: raw events are compressed per episode into M1M_1; M2M_2 is generated from M1M_1 only, through KK parallel section calls followed by one Reduce call; and a router assembles only the sections a query needs. Figure 2 contrasts the refresh path of the monolithic baseline and PRISM.

Frontier teacher model: The frontier teacher is GPT-5.5 at temperature 0.2, accessed through the OpenAI API with reasoning effort set to low. It generates both the monolithic baseline and the PRISM sections under this configuration.

LongProfile construction: Yes. LongProfile is built from the union of MovieLens-25M, Amazon Reviews 2023 and Yelp, and the Setup subsection and dataset appendix state this explicitly. Its contribution is the construction and the protocol rather than the source data: a long-horizon filter (at least 50 interactions spanning at least two years); balanced stratified cohorts across event-count buckets up to 5,000 interactions; a user-disjoint 60/20/20 split with a temporal last-kk holdout; and labels given by held-out raw interactions, scored with IR and error metrics rather than an LLM judge. We have not yet reported downstream ranking results on LongProfile. The construction pipeline and datacard will be released with the code at camera-ready.

Reviewer_XFxnOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes PRISM, a hierarchical memory framework for scalable user profiling. It combines incremental compression of interaction histories with parallel generation of semantic profile sections.

Critical strengths

NA

Critical weaknesses
  • I have serious concerns about the correctness of the reported efficiency results. In Section 5.5, the autoregressive depth ratio is defined as the sum of section lengths divided by their maximum length. Mathematically, this ratio cannot exceed K, the number of sections. However, Table 4 reports 12.9× for K=10 and even 1.29× for K=1. These results directly contradict the stated formula, yet the 12.9× improvement is highlighted in the abstract.

  • The experimental comparisons are inadequate and favor the proposed method. The only main baseline is a monolithic profile regenerated from the entire interaction history. This is a rather weak comparison for a paper claiming improvements in incremental memory efficiency. An incrementally maintained monolithic summary or existing hierarchical memory approaches would be much more appropriate. The authors also fail to isolate the contributions of hierarchical compression and parallel section generation through proper ablations.

  • The experiments do not adequately validate the central long-horizon personalization claims. Although the paper introduces LongProfile using MovieLens, Amazon Reviews, and Yelp, I could not find actual downstream ranking results on this benchmark. Instead, the downstream experiments are limited to 200 LaMP instances per task, and the separate profile fidelity evaluation covers only 13-week histories. These experiments do not convincingly establish whether useful information survives repeated hierarchical compression over long histories.

  • The academic rigor and reliability of the manuscript are concerning. The references contain multiple apparent hallucinations and placeholders. For example, LaMP is assigned arXiv:2307.12345, which belongs to an unrelated paper. Other references contain placeholder IDs such as 2403.00000, 2410.00000, and 2404.00000. O-Mem and MemTree are also incorrectly attributed. They raise concerns about whether the related work was properly checked.

Questions for authors

NA

1 comment

Thank you for your thoughtful feedback!

Depth-ratio error:

The point is correct: under the stated definition, ρ=∑sOs/max⁡sOs\rho=\sum_s O_s/\max_s O_s cannot exceed KK, and it must equal 11 when K=1K=1.

The implementation mistakenly included the Reduce call’s output tokens in the numerator, calculating (∑sOs+Oreduce)/max⁡sOs(\sum_s O_s+O_{\mathrm{reduce}})/\max_s O_s. In the reported configuration, Oreduce=0.29∑sOsO_{\mathrm{reduce}}=0.29\sum_s O_s, which inflated the reported values by a factor of 1.29. This explains both the 12.9× value at K=10K=10 and the 1.29× value at K=1K=1.

We have removed Table 4 rather than presenting a corrected depth ratio as an empirical result. Parallel section generation reduces Map-phase autoregressive depth from ∑sOs\sum_s O_s to max⁡sOs\max_s O_s, a ratio of at most KK, with equality only when section lengths are equal. This is a property of the method, not a measured improvement. We now state it in Section 4.5 and account explicitly for the sequential Reduce call: end-to-end depth is max⁡sOs+Oreduce\max_s O_s+O_{\mathrm{reduce}}. At the reported Reduce cost, the corresponding ratio is at most approximately 2.6 at K=10K=10, rather than 12.9.

We have also revised the abstract and conclusion to remove the 12.9× claim and emphasize the measured results: the approximately 63× reduction in refresh input, profile-fidelity results, and LaMP utility. We regret the error and appreciate the reviewer’s careful reading.

Evaluation at long horizons:

We agree that the current experiments do not measure downstream utility over multi-year histories. We have narrowed the claim accordingly. The available evidence addresses three related, but distinct, questions.

First, the LaMP evaluation uses 200 matched instances per task. The paired design, task-appropriate tests, and bootstrap confidence intervals help characterize the observed differences, but this sample does not resolve whether smaller effects would persist on the full test sets. We report the nonsignificant LaMP-3 MAE result (p=0.108p=0.108) alongside the favorable descriptive difference. A full-test-set evaluation would strengthen the downstream evidence.

Second, the 13-week longitudinal evaluation directly tests profile fidelity against a monolithic baseline on the same accounts. Within that tested horizon, the section-based profiles show higher completeness (+0.384) and specificity (+0.296), similar global consistency (4.86 versus 4.84), and non-inferior episodic fact retrieval under the stated margin (p<0.0001p<0.0001). These findings support information retention over 13 weeks; they do not establish downstream utility over multi-year histories.

Third, the longer-history experiments provide evidence about coverage and processing cost, not downstream ranking quality. PRISM ingests an 18,786-event history, whereas the 128K-context monolithic baseline reads approximately 17% of that history under the tested sampling strategies. Incremental refresh input remains at 6,343 tokens at 10,000 events. We now distinguish these mechanism-level results from the untested question of whether information retained after repeated compression improves personalization at multi-year horizons. We do not claim lossless retention.

References:

We agree that the bibliography contained serious errors, and we apologize. We have reviewed the entries against their primary sources and corrected the cases identified by the reviewer:

  • LaMP: Replaced the unrelated arXiv identifier with arXiv:2304.11406 (Salemi et al.). Placeholder identifiers: Removed 2403.00000, 2404.00000, and 2410.00000. The relevant entries now cite arXiv:2505.15456 (Zhao et al.) and arXiv:2410.14052 (MemTree); entries for which we could not identify a matching publication were removed.
  • O-Mem and MemTree: Corrected their references to arXiv:2511.13593 and arXiv:2410.14052, respectively, and revised the related-work descriptions. We also revised the descriptions of Persona-DB and LLMxMapReduce.
  • Other corrections: Replaced the PriorMem entry with a reference to the OptMem repository and removed uncited bibliography entries. These errors should not have appeared in the manuscript. We appreciate the reviewer for bringing them to our attention.

Comments on the paper

1 comment

We thank the reviewer for the valuable feedback!

Depth ratio:

This point is correct, and the two reported values identify the defect. The reported values differ from the definition by a single multiplicative factor, 1.29=1+0.291.29 = 1 + 0.29, which is the Reduce output relative to the section total (Oreduce/∑sOs≈0.29O_{\mathrm{reduce}}/\sum_s O_s \approx 0.29). The implementation added OreduceO_{\mathrm{reduce}} to the numerator, computing (∑sOs+Oreduce)/max⁡sOs(\sum_s O_s + O_{\mathrm{reduce}})/\max_s O_s. Since both reported cells carry this factor, the error is not specific to large KK: K=1K{=}1 isolates it, because there the correct ratio is exactly 11. Table 4 has been removed, and no corrected empirical depth ratio is reported in its place.

The correct statement is a bound. Since max⁡sOs≥∑sOs/K\max_s O_s \ge \sum_s O_s / K, the Map-phase ratio satisfies ρ≤K\rho \le K, with equality only for equal-length sections (Section 4.5). Including the sequential Reduce call, the end-to-end ratio satisfies ∑sOsmax⁡sOs+Oreduce≤11/K+0.29,\frac{\sum_s O_s}{\max_s O_s + O_{\mathrm{reduce}}} \le \frac{1}{1/K + 0.29}, where 0.290.29 is the Reduce fraction of our instantiation. This gives bounds of 0.780.78 at K=1K{=}1, 1.271.27 at K=2K{=}2, 1.851.85 at K=4K{=}4, 2.412.41 at K=8K{=}8, 2.562.56 at K=10K{=}10, and →3.4\to 3.4 as K→∞K \to \infty.

Two consequences follow, and we will state both in the revised text. First, the end-to-end bound falls below 11 at K=1K{=}1. With a single section there is no parallelism to exploit, and the Reduce pass makes the end-to-end depth 1.291.29 times the section's own depth; taking the monolith's depth as ∑sOs\sum_s O_s, the PRISM depth is then larger than the monolith's. Second, the bound is in decoding tokens, not seconds. Our reference implementation caps Map concurrency at 88, so K=10K{=}10 sections execute in two waves, and we report no wall-clock measurements.

Clarification 1:

The mechanism this submission demonstrates beyond the constituent methods is the effect of facet-wise decomposition: against the monolithic profile on the same 13-week accounts, section decomposition improves completeness and specificity while preserving consistency (point 1 above). We did not evaluate incremental summarization alone or Map--Reduce alone against the combination, so we cannot attribute the gain to a component other than the facet decomposition.

Clarification 2:

The depth-ratio response (point 2) covers this question. The reported values came from a computation error, the affected table has been removed, and point 2 gives the correct bound in its place, so no depth-ratio value remains that conflicts with the stated definition.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.