Research map / Large language models / Agents, code and math
What's the plan? Metrics for implicit planning in LLMs and their application to rhyme generation
Jim Maar, Denis Paperno, Callum Stuart McDougall, Neel Nanda
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 changed | Where | Measured | Effect |
|---|---|---|---|
| patch activations of attention heads L30H03 and L31H15 with their steered values (instead of full steering) (ablation) | Gemma2 9B-it rhyme steering · Rhyme circuit couplets | % of steering logit difference recovered: 59%-93% | Improved |
| patch the two heads on the brick couplet (ablation) | Gemma2 9B-it · Couplet 1 (reddish brick) | logit difference: -2.6 -> 4.61 (steered 5.48) | Improved |
| patch the two heads on the pain couplet (ablation) | Gemma2 9B-it · Couplet 2 (striking pain) | logit difference: -3.37 -> 1.47 (steered 4.77) | Improved |
| patch the two heads on the bird's wing couplet (ablation) | Gemma2 9B-it · Couplet 3 (bird's wing) | logit difference: -14.5 -> 2.06 (steered 3.39) | Improved |
| patch the two heads on the ancient guardians couplet (ablation) | Gemma2 9B-it · Couplet 4 (ancient guardians) | logit difference: -4.94 -> -1.07 (steered 0.93) | Improved |
| steer on the newline token instead of the last word of the first line (ablation) | Gemma2 9B and Gemma3 27B · Rhyme family steering effectiveness | fraction target rhyme family | Hurt |
| increase model size | Gemma2/Gemma3/Qwen3/Llama3 families · Rhyming and planning metrics | rhyme rate and planning metrics | Improved |
| use instruction-tuned instead of base model | Gemma2/Gemma3/Qwen3/Llama3 families · Rhyming and planning metrics | rhyme rate and planning metrics | Improved |
| ablate attention to both last word and newline token (ablation) | Gemma3 27B · Rhyming dataset | fraction correct rhyme family and regeneration | Hurt |
| ablate attention to only one of last word or newline token (ablation) | Gemma3 27B · Rhyming dataset | fraction correct regeneration | Hurt |
| steer with SAE latent decoder difference (latents 14069 and 5862) instead of mean activation difference | Gemma2 9B-it · -eep to -ight rhyme steering | rollouts with target rhyme: 832 out of 1000 | Improved |
| steer toward specific word (night/light) instead of rhyme family | Gemma2 9B and Gemma3 27B · Specific word steering | word probability change: at most 20 percent for Gemma2 9B and 50 percent for Gemma3 27B | Mixed |
Closest papers by meaning
- Latent Planning Emerges with Scale (ICLR 2026 · similarity 0.89)
- Emergent Response Planning in LLMs (ICML 2025 · similarity 0.87)
- Large Language Models as Planning Domain Generators (Student Abstract) (AAAI 2024 · similarity 0.87)
- Planning in the Dark: LLM-Symbolic Planning Pipeline Without Experts (AAAI 2025 · similarity 0.87)
- LLMs Can Plan Only If We Tell Them (ICLR 2025 · similarity 0.86)
- ALPINE: Unveiling The Planning Capability of Autoregressive Learning in Language Models (NeurIPS 2024 · similarity 0.85)
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.
