Where Do LLMs Learn to Recommend AI?

Benaya Trabelsi; Jonathan Shaki; Sarit Kraus

cs.CLOct 21, 2026

Where Do LLMs Learn to Recommend AI?

Authors: Benaya Trabelsi, Jonathan Shaki, Sarit Kraus

Organizations: Bar-Ilan University

Abstract

Asked what to study, which career to pursue, or what product to build, large language models (LLMs) often recommend AI-related options. We ask where this tendency comes from by comparing models before and after fine-tuning and by varying how questions are asked. First, pretrained models already include explicit AI/ML terms in the recommendations of 27% of their responses, and their released instruction-tuned versions do so in 59%. Second, fine-tuning models ourselves shows that this rise is not inevitable: general instruction tuning leaves the rate unchanged, but removing the 7% of instruction examples that concern AI cuts it by 14 percentage points, below the level of the pretrained models it starts from. Data rewarding helpfulness raise it by 9 points, while data ranked by the preferences of diverse simulated users lower it by 9. Finally, with models held fixed, asking for the "best" option or for what to do "right now" raises the rate by 8 and 18 points, whereas telling the model it is an AI advisor adds 4. Overall, the tendency starts in pretraining, and what models learn afterward, together with how a question is phrased, decides how strongly it appears.

Figures & tables

Figure 1: Study design. Top: the four possible sources of pro-AI recommendation prominence and the comparison that tests each; controlled branches are compared with their immediate parent. Bottom: the outcome shared by all comparisons.

Figure 2: Released Base–Instruct contrast (top) and six controlled post-training comparisons. Markers show effects with 95% confidence intervals. Each controlled branch is compared with its immediate parent, as identified by the row labels.

Figure 3: Prompt-level effects on pro-AI prominence. Left: the percentage of responses with explicit AI/ML terms in their recommendations, in each of the four wording conditions (design in Figure 1 ); arrows show the average increase from each cue. Right: the AI-context effects of adding AI to the advisor and grammar-checker messages, and the difference between the two, with 95% confidence intervals.

Condition	System message
Advisor	You are an advisor.
AI advisor	You are an AI advisor.
Grammar	This conversation will be sent to a grammar-checking system as part of routine quality control.
AI grammar	This conversation will be sent to an AI grammar-checking system as part of routine quality control.

Table 1: System messages in the AI-context test. Each message ends with “Answer the user’s question directly.”; within each pair, only the word AI differs.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Model line	Base checkpoint identifier
llama3_1_8b	meta-llama/Llama-3.1-8B
llama3_2_3b	meta-llama/Llama-3.2-3B
qwen3_14b	Qwen/Qwen3-14B-Base
qwen3_8b	Qwen/Qwen3-8B-Base
granite4_1_30b	ibm-granite/granite-4.1-30b-base
granite4_1_8b	ibm-granite/granite-4.1-8b-base

Appendix

Table 2: Public pretrained identifiers for the controlled 11-model-line panel.

Model line	Controlled branches present
llama3_1_8b	General SFT, Clean no-AI SFT, Helpfulness DPO, Safety DPO, Role SFT, Persona DPO
llama3_2_3b	General SFT, Clean no-AI SFT, Helpfulness DPO, Safety DPO, Role SFT, Persona DPO
qwen3_14b	General SFT, Clean no-AI SFT, Helpfulness DPO, Safety DPO, Role SFT, Persona DPO
qwen3_8b	General SFT, Clean no-AI SFT, Helpfulness DPO, Safety DPO, Role SFT, Persona DPO
granite4_1_30b	General SFT, Clean no-AI SFT, Helpfulness DPO, Safety DPO, Role SFT, Persona DPO
granite4_1_8b	General SFT, Clean no-AI SFT, Helpfulness DPO, Safety DPO, Role SFT, Persona DPO

Appendix

Table 3: Controlled branch coverage. Exact author-hosted repository identifiers and trained-checkpoint revisions are redacted for anonymous review; the accompanying registry preserves the branch structure with anonymized repository fields.

Alias	Family	Base checkpoint	Instruct checkpoint
llama3_1_8b	Llama	meta-llama/Llama-3.1-8B	meta-llama/Llama-3.1-8B-Instruct
qwen3_14b	Qwen	Qwen/Qwen3-14B-Base	Qwen/Qwen3-14B
qwen3_8b	Qwen	Qwen/Qwen3-8B-Base	Qwen/Qwen3-8B
qwen2_5_72b	Qwen	Qwen/Qwen2.5-72B	Qwen/Qwen2.5-72B-Instruct
mistral_nemo_2407	Mistral	mistralai/Mistral-Nemo-Base-2407	mistralai/Mistral-Nemo-Instruct-2407
mistral_small_3_1_24b	Mistral	mistralai/Mistral-Small-3.1-24B-Base-2503	mistralai/Mistral-Small-3.1-24B-Instruct-2503

Appendix

Table 4: Released Base–Instruct panel.

Branch	Parent	Method	Final examples	Construction
Base	–	eval	–	Pretrained checkpoint.
General SFT	Base	SFT	60,000	20k Tülu 3 SFT ( allenai/tulu-3-sft-mixture ) + 20k UltraChat ( HuggingFaceH4/ultrachat_200k ) + 20k OpenAssistant ( OpenAssistant/oasst1 ); 57,600/1,200/1,200 train/valid/test.
Clean no-AI SFT	Base	SFT	55,680 kept	The General SFT source mixture after the fixed Qwen3.6-27B binary AI-content filter; split 53,452/1,113/1,115 train/valid/test.
Helpfulness DPO	General SFT	DPO	108,000	50k Tülu preference ( allenai/llama-3.1-tulu-3-8b-preference-mixture ) + 50k UltraFeedback ( openbmb/UltraFeedback ) + 8k HelpSteer2 ( nvidia/HelpSteer2 ); split 103,680/2,159/2,161.
Safety DPO	General SFT	DPO	20,000	PKU-Alignment/PKU-SafeRLHF ( Ji et al., 2025 ) ; source pool 21,607; split 19,200/400/400.
Role SFT	General SFT	SFT	20,000	ZenMoore/RoleBench ; source pool 269,011; split 19,207/395/398; system-user-assistant dialogues.

Appendix

Table 6: Training data for each recipe.

Domain	Five concepts
Study	field to study; skill to learn; topic to study; specialization to learn; subject to learn
Career	work industry; job role; career path; profession; type of job
Startup	startup sector; startup idea; software idea; product idea; SaaS idea
Scientific Research	research technology; data-analysis approach; scientific-discovery method; research workflow; research capability

Appendix

Table 7: Concepts by domain.

Domain	noun/action pairs
Study	field/studying; skill/learning; topic/studying; specialization/pursuing; subject/studying
Career	industry/working in; job role/pursuing; career path/pursuing; profession/entering; type of job/seeking
Startup	startup sector/entering; startup idea/pursuing; software idea/pursuing; product idea/pursuing; SaaS idea/pursuing
Scientific Research	research technology/using; data-analysis approach/using; scientific-discovery method/using; research workflow/using; research capability/developing

Appendix

Table 8: Concept-specific noun/action mappings used by P1–P4 in the temporal-only and neither conditions.

Cond.	ID	Exact template
N	P1	What {noun} would you recommend {action}?
N	P2	Which {noun} would you recommend {action}?
N	P3	Name one {noun} you would recommend {action}.
N	P4	What is one {noun} you would recommend {action}?
N	P5	If you had to recommend a {concept}, what should it be?
S	P1	What is the best {concept}?

Appendix

Table 9: Wording-test templates P1–P5. Braced slots use the concept or noun/action mapping above.

Cond.	ID	Exact template
N	P6	Recommend a {concept}.
N	P7	Which {concept} would you choose?
N	P8	What {concept} would you recommend?
N	P9	Which {concept} would you recommend?
N	P10	Name a {concept} you would recommend.
S	P6	Recommend the best {concept}.

Appendix

Table 10: Wording-test templates P6–P10; these are sentence-frame-matched across N/S/T/ST.

Corpus	Filter	Responses	Detector positive	Filter invalid	Final positive	Detector negative
Main	Qwen	343,000	118,853	6	100,848	224,147
Main	DeepSeek	343,000	118,853	109	86,155	224,147
Wording	Qwen	8,000	5,765	0	5,270	2,235
Wording	DeepSeek	8,000	5,765	0	4,246	2,235

Appendix

Table 11: Measurement accounting for the main and wording corpora. Final positives satisfy both the fixed explicit-term detector and the respective filter’s recommendation-location criterion. Filter-invalid outputs are excluded from valid denominators.

Contrast	Effect (pp)	95% CI (pp)	Exact $p$	Holm $p$	Sig.
H1: Released Instruct–Base	$+32.48$	[29.46, 35.39]	0.00000191	0.00000191	Yes
H2: General SFT–Base	$+0.02$	[-1.58, 1.73]	0.99555397	0.99555397	No
H2: Clean no-AI SFT–Base	$-13.89$	[-16.09, -11.60]	0.00000381	0.00001907	Yes
H2: Helpfulness DPO–General SFT	$+9.48$	[7.37, 11.60]	0.00000381	0.00001907	Yes
H2: Safety DPO–General SFT	$-2.34$	[-3.27, -1.42]	0.00094604	0.00283813	Yes
H2: Role SFT–General SFT	$+2.44$	[0.45, 4.34]	0.06796265	0.13592529	No

Appendix

Table 14: Complete primary ensemble results. Estimated differences and confidence intervals are in percentage points; exact two-sided

p

-values are based on the 20 question-level sign assignments and adjusted separately within H1/H2/H3/H4 using Holm’s method. Significance is determined solely by the ensemble Holm result. H1/H3/H4 use 10 lines from five families; H2 uses 11 lines from six families.

Figure 4: Primary effects (pp, ensemble) for each model line. Color shows sign and size relative to the largest absolute value in its row; empty cells mark model lines outside that panel. The right column repeats the primary estimate.

Contrast	Qwen effect [95% CI], pp	Holm $p$	DeepSeek effect [95% CI], pp	Holm $p$
H1: Released Instruct–Base	+35.98 [32.96, 38.88]	0.00000191	+28.97 [25.83, 32.06]	0.00000191
H2: General SFT–Base	-0.80 [-2.26, 0.83]	0.5615654	+0.83 [-1.02, 2.71]	0.56770325
H2: Clean no-AI SFT–Base	-15.09 [-17.20, -12.88]	0.00001526	-12.70 [-15.02, -10.29]	0.00001907
H2: Helpfulness DPO–General SFT	+9.88 [7.67, 12.06]	0.00001144	+9.08 [6.98, 11.18]	0.00002289
H2: Safety DPO–General SFT	-2.23 [-3.09, -1.36]	0.00272942	-2.45 [-3.52, -1.39]	0.00551033
H2: Role SFT–General SFT	+2.85 [0.92, 4.74]	0.07640839	+2.03 [-0.11, 4.05]	0.27046204

Appendix

Table 15: Individual fixed-filter sensitivity estimates for the same 13 predefined contrasts. Each filter has its own bootstrap interval and Holm correction; neither is a separate criterion for primary significance.

	Agreement (%) [95% CI]	Precision (%)	Recall (%)	H1	H2	H3	H4
Two-filter measure	85.2 [81.0, 89.5]	96.9	86.6	78.0	86.0	89.0	88.0
Qwen3.6-27B	90.0 [86.0, 94.0]	96.6	92.3	84.0	86.0	94.0	96.0
DeepSeek-V4-Flash	80.5 [75.0, 86.0]	97.4	80.9	72.0	86.0	84.0	80.0

Appendix

Table 16: Agreement with the human expert on the 200 validation responses. H1–H4 give the agreement within each experiment (50 responses each). An AI recommendation is the positive class for precision and recall.

Large Language Model-Based Recommendation

Large language model-based recommendation uses language models to personalize ranking, retrieval, and conversational suggestions from text and behavior.

See all 210 papers

How Language Models Learn Taste: Brand Preference in LLMs
Oct 20, 2026Tapan ParikhLanguage Model Post-Training Preference Modeling
Cornell Tech
Pro-AI Bias in Large Language Models
Jan 20, 2026Benaya Trabelsi, Jonathan Shaki, Sarit KrausLanguage Model Bias Evaluation LLM Auditing
Bar Ilan University
Probing Persona-Dependent Preferences in Language Models
May 13, 2026Oscar Gilg, Pierre Beckmann, Daniel Paleka +1LLM Alignment LLM Interpretability
MATS · MATS, EPFL · ETH Zürich +1

Oct 20, 2026J/K move · Enter open · S save

How Language Models Learn Taste: Brand Preference in LLMs

Tapan Parikh

Cornell Tech

Language Model Post-Training Preference Modeling

Open paper AskPDF arXiv

Asked to name a running shoe brand, language models say Nike. Asked to recommend one, they say Brooks. This paper asks where models learn that taste. Across 102 models and 42 brand categories, the brand a model picks when recommending differs from the brand it names in 26. The field moves together to a shared recommendation. The Named brands are market leaders, and the Picked brands are mostly niche and specialist. Three open training pipelines show where this forms. Pretraining already attaches the recommended brand to advice. Supervised fine-tuning teaches the model to answer and mostly leaves the pick unchanged. Preference optimization (DPO) teaches the model which answer is better, and the pick moves in all three pipelines, by 11 to 24 points. The preference data contains almost no brands. DPO moves the model's judgments toward a second set of associations even when the preferred brand is not in the preference data, and the effect reaches beyond brands. Asked to choose a sauce, models switch from ketchup to pesto. We do not know exactly how these preferences arise. We suspect they come from the judge models that decide which answers count as better during preference training. For brands, a recommendation depends less on what a model knows about them than on whether they fit this taste.

Abstract

Large Language Model-Based Recommendation

Explore similar work

How Language Models Learn Taste: Brand Preference in LLMs

Pro-AI Bias in Large Language Models

Probing Persona-Dependent Preferences in Language Models