AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

The Invisible Lottery: How Subtle Cues Steer Algorithm Choice in LLM Code Generation

Akanksha Narula, Mofasshara Binte Rafique, Laurent Bindschaedler

ICML 2026 · Large language models › Language, safety and interpretability · approach: Statistical frameworks for LLM evaluation · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add a readability instead of space constraint cueLLM code generation, tree traversal · Tree traversalrecursive implementation rate: 0% -> 89%Improved
use prototype instead of interview contextLLM code generation, expression parsing · Expression parsingeval() usage: 6.1% -> 69.7%Hurt
use academic instead of junior personaLLM code generation, memoization · Memoizationpass rate: 100% -> 80.0%Hurt
add import functoolsLLM code generation, memoization · Memoizationtop-down memoization rate: 17.1% -> 82.9%Improved
add a space constraint cueLLM code generation, rate limiter · Rate limitersliding-window rate: 90.0% -> 3.6%Hurt
add semantically neutral placebo cues (team names, colors, ids)LLM code generation · 5 API modelssteering (pp): 26.1 mean pp shiftImproved
change surface typography of a time-critical cue, meaning held fixed (ablation)memoization prompt · Memoizationmatrix exponentiation rate: 0% baseline -> 14-41%No clear effect
name the algorithm explicitly in the promptLLM code generation · 120 runs, 3 API modelspass rate: 120/120 pass; compliance 60.8% strict, 98.3% lenientImproved
explicitly ask GPT-5 for matrix exponentiation instead of academic personaGPT-5, memoization · Memoizationpass rate: 0/5 -> 5/5Improved
raise sampling temperature from 0.0 to 1.0 (ablation)Qwen2.5-Coder 32B · Cue steering gridmean capacity (pp): 49.7 -> 64.7 capacity; pass 93.0% -> 91.4%Mixed
quantize DeepSeek-R1 70B from FP16 to Q8_0 (ablation)DeepSeek-R1 70B · Cue steering gridcapacity and pass rate: -4.5 pp capacity, -1.6 pp pass rateMixed
raise classifier confidence threshold from 0.3 to 0.7 (ablation)AST classifier · Tree traversal + constraintsteering delta (pp): 88.6 -> 88.6No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy