Research Briefing — January 21, 2026

Investigates whether instruction-tuned models always outperform base models, finding that base models consistently outperform instruction-tuned variants in zero-shot CoT settings on GSM8K (drops up to 32.67% for Llama3-70B). Instruction tuning appears to induce pattern matching rather than genuine reasoning improvement.

Systematic study showing RLVR improves task performance but produces extremely overconfident models, while SFT yields better calibration even under distribution shift. Proposes calibration-aware RL approach to balance classification and calibration.

Introduces CausalSpatial benchmark evaluating whether MLLMs can anticipate consequences of object motions across collision, compatibility, occlusion and trajectory tasks. Reveals severe gap: humans score 84% while GPT-5 achieves only 54%, exposing over-reliance on textual reasoning.

Discovers 'Action Rebinding' - a critical security vulnerability in multimodal GUI agents where zero-permission apps can hijack agent actions by exploiting the gap between observation and action execution. Demonstrates that Visual Atomicity assumption is invalid on Android.

Demonstrates that safeguarded frontier models can be used to elicit harmful capabilities in open-source models through three-stage elicitation attacks using adjacent-domain prompts that bypass safeguards.

APEX-Agents benchmarks AI agents on professional tasks from investment banking, consulting, and law with 480 long-horizon cross-application tasks. Gemini 3 Flash achieves 24%, followed by GPT-5.2 and Claude Opus 4.5.

Comprehensive unified evaluation of LLM reasoning paradigms from single-model to multi-agent systems, characterizing performance across benchmarks and analyzing cost-accuracy trade-offs. Probes role-specific capability demands in MAS.

Proposes HuDA, a surprisingly simple reward model using human detection confidence and temporal prompt alignment to improve human motion in generated videos. Off-the-shelf models outperform specialized methods without training.

Demonstrates that AI-generated data contamination in medical AI creates feedback loop causing erosion of pathological variability and diagnostic reliability, with models converging toward generic phenotypes regardless of architecture.

Proposes Threshold Differential Attention (TDA), a sink-free attention mechanism achieving ultra-sparsity and improved robustness at longer sequence lengths. Uses row-wise extreme-value thresholding with length-dependent gating without computational overhead of projection methods.