The sameness problem behind those unappetizing AI-generated menus

1. Decision — Adopt a two‑step workflow now

Recommendation – For AI‑generated restaurant menus that stand out, start with a small, domain‑specific fine‑tune of a capable language model and then run a prompt‑engineered, diversity‑focused generation stage. The second stage should use calibrated decoding parameters and a lightweight post‑processing step.

Why this matters – An off‑the‑shelf model tends to repeat the same phrasing (“served with …”, “tender …”) because those patterns dominate its training data. Fine‑tuning introduces a culinary vocabulary and a style baseline; the diversity stage adds controlled randomness and domain constraints that keep each item distinct. Skipping either step either leaves the output generic (no fine‑tune) or makes it incoherent (no diversity controls). The combined approach balances originality with reliability and can be run on inexpensive cloud instances or modest on‑premise hardware.

Concrete compute example – An AWS t4g.micro instance (2 vCPU, 8 GB RAM) costs a very low amount per hour in the US East region. Using the open‑source Llama‑2‑7B model with LoRA adapters, a single‑epoch fine‑tune on a 7 k‑line menu corpus finishes in a short amount of time on this instance (GPU‑free, CPU‑only). For GPU‑based work, an NVIDIA T4 (16 GB VRAM) on a comparable cloud offering runs the same fine‑tune in a few minutes and costs a modest amount per hour. These figures give readers a concrete sense of the resources required.

---

2. The “sameness problem” in a nutshell – what it is and why it hurts a menu

Definition – The sameness problem is the tendency of AI‑generated menu text to repeat generic phrasing, ingredient lists, and sensory adjectives, producing entries that read as interchangeable and fail to stimulate appetite or convey brand identity.

Illustrative bland items (generated by a default model with temperature 0.7 and no domain conditioning):

All five entries share the same “X served with Y” skeleton and rely on a narrow adjective set (tender, warm, juicy, succulent, rich). When diners scan such a list they receive no sense of the chef’s personality, regional inspiration, or unique ingredient pairings, which can lower perceived value and reduce conversion rates.

---

3. Why language models fall into the repetition trap – the root‑cause story

1. Training‑data bias – Large‑scale corpora (web pages, recipe blogs, Wikipedia) contain many “standard” recipe sentences because those are the most frequently edited and SEO‑optimised. The model internalises these high‑frequency patterns and reproduces them when asked to write a menu.

2. Token‑level sampling – Decoding algorithms select one token at a time. When the probability distribution is sharply peaked (common in culinary language), low‑temperature sampling repeatedly chooses the same top tokens, reinforcing identical phrase structures.

3. Lack of domain‑specific context – General models have no built‑in notion of a restaurant’s brand voice, regional cuisine, or menu hierarchy. Without explicit signals, they default to the safest, most universally acceptable description.

4. Over‑reliance on temperature / Top‑P alone – Raising temperature or expanding Top‑P can add variety, but it also raises the risk of incoherent or contradictory details (e.g., “spicy sweet” when the dish is mild). The model lacks a higher‑level guide to keep novelty within culinary plausibility.

5. Prompt under‑specification – Simple prompts like “Write a menu for a bistro” give the model too much freedom and too little guidance about tone, ingredient limits, or stylistic anchors, leading it to fall back on its most probable patterns.

---

4. Technical levers that drive uniformity – training data, prompting, decoding

Evidence for the Top‑P recommendation

A small pilot test generated 200 appetizer descriptions using three nucleus settings (p = 0.8, 0.9, 0.95) while keeping temperature = 0.9. The following observations were made on a held‑out validation set:

Increasing p to 0.95 reduced exact 3‑gram repeats noticeably while raising lexical diversity and preserving a mildly positive sentiment. The table demonstrates why a higher nucleus is useful for menu generation.

---

5. Practical recipe for breaking the monotony – prompts, parameters, post‑processing

Step 1 – Assemble a curated culinary corpus

• Source material – Collect several thousand lines from high‑quality menus, chef‑authored tasting notes, and flavor‑pairing databases (e.g., Foodpairing, FlavorDB).
• Cleaning – Strip HTML, normalise measurement units, and tag each entry with cuisine, course, and tone (e.g., [Modernist], [Family‑style]).
• Fine‑tune – Use a parameter‑efficient method such as LoRA or adapters on a base model like Llama‑2‑7B or GPT‑Neo‑2.7B.

Empirical result for one‑epoch fine‑tune

A pilot fine‑tune on a 7 k‑line corpus (average 18 tokens per line) was run for a single epoch on an NVIDIA T4 (16 GB VRAM). After fine‑tuning:

• Cross‑entropy loss decreases, indicating better fit to the culinary style.
• BLEU‑4 scores improve on a held‑out set of menu items, showing closer alignment with reference descriptions.
• Lexical diversity shows a noticeable increase, reflecting richer vocabulary usage.

These gains were observed for corpora up to roughly ten thousand examples; larger corpora benefit from additional epochs, but one epoch already yields a noticeable style shift for modest datasets.

Step 2 – Design a diversity‑focused prompt template

You are a chef writing a menu for a [cuisine] restaurant. Use a [tone] voice. Include at least one uncommon ingredient or technique per dish. List 5 appetizers, 5 mains, and 3 desserts.

[Example 1]
Dish: Charred Octopus Carpaccio
Description: Silky‑thin slices of octopus, smoked over oak, drizzled with yuzu‑infused olive oil, topped with toasted fennel pollen.

[Example 2]
Dish: …

• Few‑shot anchors – Provide 2–3 fully written dishes that showcase the desired lexical richness.
• Ingredient constraints – Add a line such as “Include one of: kaffir lime, black garlic, fermented miso” to force novelty.

Step 3 – Select decoding settings that encourage variety but stay coherent

Step 4 – Apply lightweight post‑processing

1. Synonym substitution – Run each description through a culinary thesaurus (e.g., WordNet filtered for food terms) to replace over‑used adjectives (“delicious” → “zesty”, “savory” → “umami‑rich”).
2. Flavor‑pairing validation – Cross‑reference