This Week's AI Papers: Loop Decoding, Self-Improvement, and Agent Benchmarks

Loop Decoding and Self-Improvement Learning

LoopCD is a decoding framework that guides token selection in iterative Transformers by comparing final predictions with previous iterative passes. LoopCD-Logits uses one additional output pass in the logit space, while LoopCD-Hidden operates with zero output overhead in the hidden-state space. On the Ouro-2.6B-Thinking model, AIME 2024 pass@1 improved from 61.88% to 73.33%, and HumanEval pass@1 on the Huginn model improved from 22.56% to 31.71%. Reducing the number of iterative loops by half achieves performance above the full-depth unguided baseline while reducing forward FLOPs by 22.5%~48.2%.

UniEvo-VL proposes an on-policy self-distillation learning recipe for self-improvement of multimodal models. Without a separate teacher model, a single model performs both the teacher role (self-critique conditioned) and the student role (original question). Based on Qwen-image-2512, the GenEval score rose from 0.747 to 0.808. While using external critics is expected to yield higher upper bounds, the effectiveness varies across tasks such as text rendering.

Agent Learning and Benchmarks

SkillGym is a framework for internalizing human-written agent skills into LLMs, containing 2,756 environments across 12 categories and 8,364 successful trajectories. When applying SFT under Claude Code, the Qwen3.5-35B-A3B model improved by 199 Elo on GDPval-AA v2 and by 19.10%p on Terminal-Bench 2.1. The 35B SkillGym-Agent achieved 51.47% on SkillsBench, surpassing Claude Sonnet 4.6, GPT-5.4 Mini, and others.

RLTL;DR overcomes the limitations of reinforcement learning with verifiable rewards (RLVR) by including self-feedback after failure in the context and internalizing it into the policy via backpropagation. On the Qwen 3.5 9B Thinking model, Pass@1, which had plateaued at 0~1%, improved to 14~31%, and achieves 12~13% during evaluation even without insights.

AIM is a Bayesian optimization-inspired framework for idea-based exploration, exceeding long-term model development baselines by 4.9%p on 10 AutoLab benchmarks and reducing wall-clock time by up to 3.1x.

Spatial Intelligence and Consciousness Assessment

KilometerVision is an ECCV 2026 public benchmark evaluating large-scale spatial intelligence in VLMs, measuring geographic placement up to 1km distance using real walking tour videos. Current VLMs appear to rely on 2D visual recognition and text matching without true path integration.

The AI Consciousness Assessment Framework uses a 5-level hierarchical functional description extending Marr's levels of analysis and a Bayesian model. The overall probability of consciousness for current LLMs varies from below 0.01 to approximately 0.8 depending on assumptions and interpretations, suggesting that consciousness metrics closely overlap with features of general intelligence architectures.