Reproducing Transparent and Scrutable Recommendations: Exploring Open-Weight Models via Natural-Language User Profiles
Noah Mamié, Laurin van den Bergh
cs.IR · cs.AI
Quick verdict
Rigorous UPR reproducibility confirms accuracy claims under test-set reranking, but its own steering probes undermine scrutability; the conclusion overstates.
Summary
This reproducibility study re-evaluates User Profile Recommendation (UPR) by Ramos et al. (2024), which synthesizes natural-language user profiles from review histories (Llama2/Mistral) and fine-tunes a sequence classifier to rate candidate items. The authors reproduce core results on Amazon Movies & TV and TripAdvisor, reusing the original Llama2-generated profiles, realigning baseline protocols in Cornac (MostPop, UserKNN, ItemKNN, MF, NeuMF), and documenting unreproducible baselines (PETER+, PEPLER-MLP lack code) and a NeuMF variant mismatch. Extensions include a six-way context ablation, five-seed stability trials (sigma MAP=0.0009), and mechanistic analysis via nnsight. The ablation shows profiles competitive with raw histories (MAP 0.895 vs 0.891). Mechanistic probes, however, complicate the scrutability narrative: profile edits shift predicted ratings uniformly across genres (ANOVA p=0.79/0.60), activation steering reverses a linear probe readout without changing rankings, and UMAP clusters organize lexically rather than by taste (ARI 0.0008). Crucially, UPR barely personalizes: percentile 0.542 (title) and 0.510 (title+desc, indistinguishable from chance) versus WMF's 0.806, traced to the rating-regression objective trained without negative pairs. Limitations: only the Llama2 UPR variant reproduced, multi-seed only on Amazon, evaluation inherited the test-set reranking protocol over items users already selected, and the conclusion still claims steering-enabled scrutability that its own Section 5.4 refutes.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The core results of Ramos et al. (2024) reproduce: UPR is competitive with classical baselines under the original test-set reranking protocol on Amazon Movies & TV and TripAdvisor.
- Natural-language profiles match raw review histories with substantially shorter prompts (Table 2: MAP 0.895 vs 0.891, MAE 0.784 vs 0.763).
- Profile edits shift predicted ratings uniformly across genres and never change rankings, even under direct activation steering that fully reverses a genre probe.
- The failure to personalize is a property of the rating-regression objective trained without negative user-item pairs (WMF percentile 0.806 vs UPR 0.51-0.55), not of the natural-language profile interface.
- Gradient attributions over profile tokens reflect word presence and rarity rather than preference features acquired through fine-tuning (Table 9 controls).
Strengths
Honest negative results against the scrutability premise
Section 5.4 reports that profile edits shift predicted ratings uniformly across genres (Table 6, ANOVA p=0.79/0.60), that activation steering flips a probe from P(crime)=0.064 to 1.000 while crime items' mean rank moves at most 0.2 of 330 positions (Table 8), and that UPR personalization is near chance (Table 7) — findings that directly qualify the original paper's core narrative, reported rather than buried.
Unusually careful controls in the mechanistic analysis
Steering experiments verify zero and constant vectors are exact no-ops before interpreting nulls, steering is applied at blocks 2/6/9 with the stated reason that final-layer steering cannot be item-selective; attribution claims are deflated via random-initialization and word-shuffled controls plus corpus-frequency baselines (Table 9); a borderline N=40 result was re-tested at N=110 and discarded when it failed to survive.
Real reproducibility infrastructure and diagnosis of original-code defects
Transparent accounting of what could not be reproduced
Tight multi-seed variance supporting the reproduced numbers
Weaknesses
Conclusion contradicts the paper's own findings
Reproduction is not numerically equivalent, only qualitatively aligned
All extensions confined to one dataset
Single UPR variant reproduced
Mechanistic conclusions rest on small samples
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a reproducibility study of Ramos et al. (2024), *Transparent and Scrutable Recommendations Using Natural Language User Profiles* (UPR), extended with three original analyses: a 3x2 context ablation, five-seed stability trials, and a mechanistic interpretability investigation using nnsight. The reproduction targets the core claim that LLM-synthesized natural-language profiles match raw review histories while enabling user inspection and editing. Evaluation covers Amazon Movies & TV and TripAdvisor with Llama2-generated profiles deliberately reused from the original authors, and Cornac baselines (MostPop, UserKNN, ItemKNN, MF, NeuMF) realigned to the LLM evaluation script. The abstract claims verification that UPR "achieves competitive performance under its test-set reranking protocol and makes recommendations more transparent" — a carefully hedged formulation that the conclusion then undermines (see Verdict).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is User Profile Recommendation (UPR)?
UPR (Ramos et al., 2024) converts a user's past reviews into an editable natural-language profile, then a fine-tuned language model scores candidate items against that profile. The intent is that users can read and correct their profile to fix recommendations — a property called scrutability.
Did this reproduction of the UPR paper succeed?
Largely yes for accuracy: the reproduced UPR models are competitive with classical baselines under the original evaluation protocol, and rating metrics match the published numbers exactly because the same data splits are reused. However, reproduced ranking metrics came out 0.03-0.06 higher than published after protocol fixes, and two baselines (PETER+, PEPLER-MLP) could not be reproduced because the original repository lacks their code.
Can users actually change their recommendations by editing their profile?
According to this study's mechanistic analysis, not in the original formulation. Editing the profile shifts predicted ratings up or down uniformly across all genres, leaving rankings essentially unchanged, and even direct activation steering that fully reverses a genre probe does not move item ranks. The root cause identified is the rating-regression training objective, which sees no negative user-item pairs and learns a user offset plus an item offset rather than content matching.
Is the code and data for this reproducibility study available?
The authors state they release the full code with a locked uv environment, standardized CLIs, and cuda/mps/cpu support; the Amazon Movies & TV and TripAdvisor datasets are public, and the original Llama2 profiles are reused from the original authors. The extracted text asserts availability in a footnote but does not display a repository URL.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 (pipeline diagram, UMAP visualization, attribution plots) are referenced but cannot be analyzed from the extracted text.
- The metadata abstract is truncated mid-sentence; its claims were cross-checked against the full text body.
- One passage in Section 5.1 is garbled by text extraction (the 'User Profile -> Title' sentence); surrounding numbers were verified against Tables 1-2.
- No code repository URL is visible in the extracted text; code availability rests on the authors' footnote statement.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.