Papers.Cool Daily Paper Picks - April 30, 2026

A curated selection of five notable arXiv preprints covering LLM distillation, small-model reasoning, vision-language models, code generation benchmarks, and embodied navigation.

Paper 1 | TIDE: Cross-Architecture Distillation for Diffusion LLMs

Authors: Gongbo Zhang, Wen Wang, Ye Tian arXiv: 2604.07574 | cs.CL, cs.AI, cs.LG

Diffusion large language models (dLLMs) support parallel decoding and bidirectional context, but state-of-the-art dLLMs require billions of parameters to compete with autoregressive models. Existing distillation methods only reduce inference steps within a single architecture and never address cross-architecture knowledge transfer, where teacher and student may differ in architecture, attention mechanism, and tokenizer.

TIDE is the first cross-architecture dLLM distillation framework, with three modular components:

1. TIDAL: jointly modulates distillation strength based on training progress and diffusion timestep, adapting to the teacher's noise-dependent reliability 2. CompDemo: enriches teacher context via complementary mask segmentation to improve predictions under heavy masking 3. Reverse CALM: a cross-tokenizer objective that reverses block-level likelihood matching, yielding bounded gradients and dual-end noise filtering

Results: Distilling 8B dense and 16B MoE teachers into a 0.6B student outperforms the baseline by an average of 1.53 points across eight benchmarks. Code generation gains are notable: HumanEval improves from 32.3 (AR baseline) to 48.78.

Paper 2 | Select to Think: Unlocking SLM Reasoning via "Local Sufficiency"

Authors: Wenxuan Ye, Yangyang Zhang, Xueli An arXiv: 2604.07576 | cs.CL

Small language models (SLMs) are computationally efficient but often trail large language models (LLMs) in reasoning. Current approaches invoke an LLM at reasoning divergence points to generate tokens, incurring high latency and cost, while standard distillation is limited by SLM capacity.

The key discovery is local sufficiency: at reasoning divergence points, the LLM's preferred token always falls within the SLM's top-K candidates, even when it is not the SLM's top-1 choice.

Select to Think (S2T) reframes the LLM's role from open-ended generation to selection among SLM candidates, simplifying supervision from full-distribution imitation to discrete candidate ranking. S2T-LOCAL distills this selection logic into the SLM for autonomous reranking without inference-time LLM dependence.

Results: A 1.5B SLM's top-8 candidates capture the 32B LLM's choice with a 95% hit rate. S2T-LOCAL improves greedy decoding by 24.1% on average, rivaling 8-path self-consistency at single-trajectory cost.

Paper 3 | World2VLM: Distilling World-Model Imagination into VLMs

Authors: Wanyue Zhang, Wenxiang Wu, Wang Xu arXiv: 2604.07577 | cs.CV

Vision-language models (VLMs) excel at static visual understanding but struggle with dynamic spatial reasoning—imagining how scenes evolve under egocentric motion. Prior approaches either extend spatial supervision with synthetic data (lacking explicit motion-conditioned state transitions) or couple VLMs with world models at inference time (computationally expensive).

World2VLM distills a generative world model's spatial imagination into the VLM:

• Synthesizes viewpoint-consistent, geometrically aligned future views from an initial observation and a parameterized camera trajectory
• Derives structured supervision for both forward reasoning (action → outcome) and inverse reasoning (outcome → action)
• Uses a two-stage post-training recipe on a compact dataset

The model improves consistently on SAT-Real, SAT-Synthesized, VSI-Bench, and MindCube. Crucially, it surpasses inference-time world-model coupling while eliminating costly inference-time generation, showing world models can serve not only as inference-time tools but as effective training-time teachers.

Paper 4 | ClassEval-Pro: Cross-Domain Class-Level Code Generation Benchmark

Authors: Yeheng Chen, Chaoxiang Xie, Yuling Shi arXiv: 2604.07581 | cs.SE, cs.CL

LLMs perform well on function-level synthesis and repository-level code modification, but compositional code creation—building complete, internally structured classes from scratch—remains weak. Existing evaluations are either limited to isolated functions or rely on manually curated class-level tasks that are expensive to extend and vulnerable to data contamination.

ClassEval-Pro features:

• 300 class-level tasks spanning 11 domains
• An automated three-stage construction pipeline: complexity augmentation → cross-domain class composition → integration of real GitHub code from after January 2025
• Each task validated by an LLM Judge Ensemble with >90% test-suite line coverage

Evaluating 5 frontier LLMs × 5 generation strategies, the best model achieves only 45.6% class-level Pass@1, with a 17.7-point gap between the strongest and weakest models. Error analysis of 500 manually annotated failures reveals logic errors at 56.2% and dependency errors at 38.0%—cross-method coordination is the core bottleneck.

Paper 5 | Three-Step Nav: Hierarchical Global-Local Planner for Zero-Shot VLN

Authors: Wanrong Zheng, Yunhao Ge, Laurent Itti arXiv: 2604.07575 | cs.CV, cs.RO

Multimodal large language models (MLLMs) have enabled breakthrough progress in vision-based navigation of unknown environments, but existing zero-shot vision-language navigation (VLN) agents still suffer from route deviation, premature stopping, and low overall success rates.

Three-Step Nav counters these failures with a three-view protocol:

1. Look forward: extract global landmarks to sketch a rough plan 2. Look now: align the current visual observation with the next subgoal for fine-grained guidance 3. Look backward: audit the full trajectory and correct accumulated drift before stopping

Requiring no gradient updates or task-specific fine-tuning, it plugs into existing VLN pipelines and achieves state-of-the-art zero-shot performance on the R2R-CE and RxR-CE datasets.

---

*Auto-collected via Papers.Cool — April 30, 2026*