Research map / Vision-language and multimodal / Visual question answering
Answer, Assemble, Ace: Understanding How LMs Answer Multiple Choice Questions
Read the paperCheck your idea against 108,967 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.
Uses vocabulary projection and activation patching to localize how Olmo, Llama 3.1 and Qwen 2.5 models produce answer-symbol predictions in formatted MCQA, finding a sparse set of middle-to-late attention heads drives the answer symbol.
Problem
How do LMs that perform well on formatted multiple-choice QA (answer symbol output in a single pass) internally bind answer content to answer symbols, and why do some models fail under format changes?
Approach
Combines vocabulary projection (logit lens) and activation patching over attention heads, MLPs and residual states on correct-prediction pairs under different answer-symbol placements, plus a synthetic Colors copying task to separate symbol binding from dataset knowledge.
Key results
- Olmo 0724 7B Instruct on Colors reaches 100% accuracy and MMLU 53.5%, HellaSwag 62.4%; activation patching localizes the key causal role to layer 24 MHSA for answer prediction.
- Answer symbol production is driven by a sparse set of attention heads (1-4 per layer of 32) projecting non-negligibly to answer symbols, with MHSA at layers 26, 27, 29 contributing most to logit differences.
- On the synthetic Colors task, Olmo 7B Base learns formatted MCQA between 80k and 100k training steps, going from near-random to near-100% accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase training of Olmo 0724 7B Base from 50K to 100K steps | Olmo 0724 7B Base · Colors | accuracy: 25% acc. -> 97% acc. | Helped |
| Continue training Olmo 0724 7B Base beyond 100K steps | Olmo 0724 7B Base · Colors | accuracy: 97% acc. -> 81% acc. -> 100% acc. | Mixed |
| Train Olmo 0724 7B Base for more steps | Olmo 0724 7B Base · Colors | logit difference between answer choice tokens | Helped |
| Use 3-shot in-context examples rather than 0-shot prompting | Models evaluated on formatted MCQA · Colors, HellaSwag, and MMLU | accuracy | No clear effect |
Surprising findings
- For random letter symbols (Q/Z/R/X, O/E/B/P), Olmo and Qwen models first assign probability to expected symbols A/B/C/D before switching to the prompt's symbols in later layers (layers 29 and 23).
- MLP outputs increase logits of all answer choices non-discriminatively, while attention heads account for the discriminative answer-choice differences despite having about half the parameters of MLPs.
Limitations
- Activation patching assumes component independence and patches individual states in isolation.
- Analysis restricted to three models (smallest Llama 3.1, Olmo 0724, Qwen 2.5 1.5B) and two real datasets with 1000 instances each.
- Vocabulary projection is not causal, so negative results in earlier layers are uninformative.
Experimental setup
Datasets: MMLU HellaSwag Prototypical Colors
Benchmarks: MMLU HellaSwag Copying Colors from Context
Metrics: accuracy logit difference probit score minimum accuracy across answer symbol sets
Baselines: Chinchilla-70B correct letter heads analysis GPT-2 label bias value vector editing
Builds on: activation patching / causal tracing logit lens residual stream view of transformers
Closest papers by meaning
- Listening to the Wise Few: Query–Key Alignment Unlocks Latent Correct Answers in Large Language Models (NeurIPS 2026 · similarity 0.89)
- Rethinking Generative Large Language Model Evaluation for Semantic Comprehension (ICML 2024 · similarity 0.88)
- Representation Consistency for Accurate and Coherent LLM Answer Aggregation (NeurIPS 2025 · similarity 0.88)
- Geometry of Decision Making in Language Models (NeurIPS 2025 · similarity 0.88)
- Tailoring Self-Rationalizers with Multi-Reward Distillation (ICLR 2024 · similarity 0.87)
- Evaluating language models as risk scores (NeurIPS 2024 · similarity 0.87)
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.