AI Paper Atlas

Research map / Large language models / Agents, code and math

Type-Compliant Adaptation Cascades

Chu-Cheng Lin, Daiyi Peng, Yifeng Lu, Ming Zhang, Eugene Ie

ICLR 2026 · Large language models › Agents, code and math · approach: LLMs for optimization and heuristics

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
adapt weights with TACSTaR instead of using un-adapted base weights (ablation)cot-cascade-structure · MGSMaccuracy: 45.4 -> 82.2Improved
adapt weights with TACSTaR instead of un-adapted (ablation)cot-cascade-structure · FinQAaccuracy: 13.0 -> 36.0Improved
adapt weights with TACSTaR instead of un-adapted (ablation)refine-structure · HotPotQAaccuracy: 24.0 -> 39.0Improved
use typed TACSTaR instead of untyped STaRTAC · MGSM (gemma-1.1-7b-it)accuracy: 10.5% -> 27.3%Improved
use TAC instead of best DSPy on structured outputTAC · MGSM-SymPyaccuracy: 57.1% -> 75.9%Improved
use TAC instead of best DSPy on small modelTAC · FinQA (gemma-1.1-7b-it)accuracy: 0.7% -> 9.7%Improved
add amortized inference network at training time (ablation)TACSTaR · FinQAaccuracy: 36.0 -> 41.7Improved
add amortized inference network at training time (ablation)TACSTaR · MGSMaccuracy: 82.2 -> 82.4No clear effect
use renormalized posterior classification instead of unconstrained generation (ablation)TAC · MuSR (gemma-2-27b-it)accuracy: 51.6 -> 65.0Improved
decouple rationale and output with separate adaptors (cascade vs type structure) (ablation)TAC · MGSMaccuracy: 80.4 -> 82.2Improved
use CoT cascade structure instead of directTAC · HotPotQAaccuracy: 34.0 -> 32.0Hurt
use refine-structure instead of directTAC · HotPotQAaccuracy: 34.0 -> 39.0Improved

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