same-family-pair-reasoning-instruction-tuned-4b-models-learning-one-interpolation CLM-0633For a same-family pair of reasoning and instruction-tuned 4B models, learning one interpolation ratio per module (attention block, FFN block, embedding, final norm) with a reward that scores correctness and ranks correct responses by length cuts average generated tokens by about 27% relative to the reasoning model while holding average accuracy, and beats fixed global-ratio merging and CMA-ES search over the same ratios.
Capability: Cutting to the chase
Observed on
Qwen3-4B Thinking/Instruct pair, dense transformer, trained on math data (DeepScaleR-preview), evaluated on five reasoning benchmarks including two out-of-domain; not tested above .
Sources
- Validated by: Accuracy and average generated tokens on AIME25, MATH500, GSM8K, GPQA-D and LiveCodeBench against the two source models, five fixed-ratio merging methods at coefficient 0.8, and CMA-ES matched on prompts, search space and objective: 76.5 average accuracy at 73.0% of Thinking's token count versus 76.0/100% for Thinking and 76.5/85.4% for SLERP.. Measured on one model pair at 4B across five benchmarks with matched-objective baselines; [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For small instruction-tuned models (3-4B) answering closed-form true/false questions under a stated user opinion, scoring each sampled response by how much more common its answer is than the group itself predicted, and using that score as the GRPO reward, reduces answer flips under user pressure and raises accuracy without any labels, roughly matching label-supervised desycophancy fine-tuning at higher compute cost.Telling the user what they want to hear · reviewed by AI
- For small instruction-tuned models (3-4B) answering closed-form true/false questions under a stated user opinion, scoring each sampled response by how much more common its answer is than the group itself predicted, and using that score as the GRPO reward, reduces answer flips under user pressure and raises accuracy without any labels, roughly matching label-supervised desycophancy fine-tuning at higher compute cost.Telling the user what they want to hear · reviewed by AI
- When training small instruction-tuned models with reinforcement learning plus factual supervision, routing each atomic fact's verification score only to the tokens that produced it, and down-weighting verifier judgements that do not change when their key evidence is removed, improves factuality benchmark scores over trajectory-level or reasoning- step-level factual rewards.Checking claims against evidence · reviewed by AI
- When fine-tuning 7-8B code models on an LLM-synthesized Verilog instruction corpus, training on a 20-25% subset selected for functional correctness, instruction-solution alignment, and coverage of Verilog structural clusters (AST, CFG, netlist features) yields higher Pass@1 than training on the full corpus, because the automated corpus carries noise and redundancy that filtering removes.Generating and editing working code · reviewed by AI
- For a domain-specific language essentially absent from pretraining corpora (FORM symbolic algebra), QLoRA fine-tuning an 8B model on a few thousand examples validated by executing them through the language's own interpreter beats frontier open-weight models of up to 756B parameters given syntax documentation in context, on execution rate and strict output match for the larger benchmarks, while general reasoning scores drop under 3 points.Generating and editing working code · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: low. Falsifier as drafted: A replication where learned per-module ratios give no token reduction at matched accuracy over a tuned single global ratio, or where CMA-ES over the same parameterization matches or beats the reward-guided ratios.