Research map / Large language models / Reasoning and chain of thought
Correct but Unselectable: The Hidden Interface Tax in Multi-Candidate Reasoning
Qianli Ma, Zhiqing Tang, Fanshuai Meng, Zhi Yao, Hanshuai Cui, Weijia Jia
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Identifies an "interface tax" in multi-candidate generate-then-rerank for small language models: correct answers can appear in the candidate pool yet fail to be selected, so more candidates do not necessarily improve accuracy.
Problem
Small language models under cost, latency, or privacy constraints need test-time scaling, but generate-then-rerank with more candidates does not reliably raise accuracy because correct answers in the pool often fail to be selected.
Approach
Identifies a previously overlooked bottleneck between candidate generation and selection (the interface between them) where a correct candidate is present but unselectable, and frames this as a hidden tax on multi-candidate reasoning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CACC interface layer to generate-then-rerank | Qwen3-1.7B + fixed verifier · GSM8K | final accuracy: 0.2873 -> 0.3632 | Improved |
| add CACC interface layer | Qwen3-1.7B + fixed verifier · competition_math (numeric) | final accuracy: 0.1519 -> 0.2579 | Improved |
| add CACC interface layer | Qwen3-1.7B + fixed verifier · GPQA Diamond | final accuracy: 0.0556 -> 0.1919 | Improved |
| add CACC interface layer | Qwen3-1.7B + fixed verifier · MMLU-Pro | final accuracy: 0.2362 -> 0.2277 | Hurt |
| add proposer-side strengthening (Qwen3-8B) to CACC | CACC · GPQA Diamond | final accuracy: 0.1919 -> 0.2576 | Improved |
| add proposer-side strengthening to CACC | CACC · MMLU-Pro | final accuracy: 0.2277 -> 0.2936 | Improved |
| add answer-mode alignment (benchmark-locked choice-letter readout) (ablation) | CCR-only pool · MMLU-Pro | final accuracy: 0.0096 -> 0.2362 | Improved |
| add candidate hygiene (SCH) on top of CCR+AMC (ablation) | CACC · MMLU-Pro | final accuracy: 0.2270 -> 0.2277 | No clear effect |
| add hygiene to CACC pool with second-family generator (ablation) | CACC · MMLU-Pro confirm-512 | final accuracy: 0.1992 -> 0.2246 | Improved |
| add conversion-aware rescue after CACC+SP | CACC+SP · MMLU-Pro | final accuracy: 0.2936 -> 0.3320 | Improved |
| raise retained cap from 8 to 16 (ablation) | CACC · GSM8K | final accuracy: 0.3632 -> 0.3298 | Hurt |
| sample 32 candidates instead of 4 (ablation) | CACC · GSM8K | final accuracy: 0.3753 -> 0.2904 | Hurt |
Surprising findings
- More candidates do not always yield higher accuracy in generate-then-rerank, because correct answers in the pool may fail to be selected.
Experimental setup
Metrics: accuracy
Closest papers by meaning
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.87)
- Plan and Budget: Effective and Efficient Test-Time Scaling on Reasoning Large Language Models (ICLR 2026 · similarity 0.87)
- Listening to the Wise Few: Query–Key Alignment Unlocks Latent Correct Answers in Large Language Models (NeurIPS 2026 · similarity 0.86)
- Joint Consistency: A Unified Test-Time Aggregation Framework via Energy Minimization (NeurIPS 2026 · similarity 0.86)
- Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.86)
- Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning (NeurIPS 2026 · similarity 0.86)
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.
