What systematic failures and vulnerabilities compromise AI reasoning systems?

19 specific questions

• Why is extracting training data insufficient proof that models memorize?
• Can document repetition accidentally memorize sensitive information instead of learning?
• How much RLVR improvement comes from benchmark data memorization?
• Why does training data not function as a searchable corpus?
• How much does memorization capacity limit a model's ability to learn new information?
• How much training data teaches retrieval models to follow instructions?
• Why does test accuracy improve after training accuracy reaches 100 percent?

34 specific questions

• Can training on diverse related tasks be more efficient than task-specific training?
• Can a single model trained on two tasks predict untrained decision tasks?
• Does task superposition explain how models learn from multiple in-context trajectories?
• Why does full multi-task fine-tuning perform worse than sequential training?
• What performance trade-offs emerge when composing multiple independently trained model capabilities?
• How do neural networks decompose tasks into modular subnetworks that transfer?
• Why do larger models reduce interference between rare and common tasks?

39 specific questions

• Does sparsity-guided ordering work equally well for reasoning and classification tasks?
• Does selecting examples from multiple complexity levels outperform selecting only high-quality examples?
• Why do easy training examples contribute less to model generalization than hard ones?
• How can language models extract more value from fewer demonstrations?
• Why do models automatically adjust reasoning length to problem difficulty?
• Why does exploration quality matter more than learner network depth?
• How much task-similar finetuning data does test-time training actually need?

15 specific questions

• How do retrieval and fine-tuning trade off flexibility against training cost?
• What training cost tradeoffs exist between fine-tuning and other knowledge injection methods?
• When does training a memory model beat RAG or fine-tuning?
• How do training-time and inference-time knowledge injection techniques compare?
• How should rapidly evolving domains choose knowledge injection methods?
• Which RAG sub-decisions are actually pattern matching versus reasoning intensive?
• Which domains need knowledge injection versus reasoning-focused training?

42 specific questions

• How do reward models and self-improvement mechanisms interact in training?
• How should training incorporate external critique versus encouraging self-correction?
• How much can externalized skills improve models before hitting diminishing returns?
• Does curriculum-based training keep small models perpetually at their learning edge?
• Does the generation-verification gap define where self-rewarding actually works?
• Does the productive difficulty band ever stabilize during training?
• At what capability level does the generation-verification gap make intrinsic rewards insufficient?

31 specific questions

• When does knowledge distillation produce student models superior to teachers?
• Why do weaker teacher models sometimes produce better training signals than stronger ones?
• Can self-training drift be prevented by applying student compatibility filtering?
• How can distillation preserve uncertainty expression instead of optimizing it away?
• Can weak models supervise the alignment of stronger models effectively?
• How does information asymmetry between teacher and student create the learning signal?
• What makes student-teacher distributional mismatch derail on-policy distillation?

29 specific questions

• Does finetuning facts into weights overwrite existing model capabilities?
• How does in-weight memorization scale with model parameter count?
• Why does in-weight memorization fail compared to tool-based fact access?
• What makes factual memorization less efficient than tool-based retrieval?
• How does in-weights adaptation create spurious forgetting in models?
• How do layer-wise versus parameter-wise merging strategies affect information retention?
• Why does fine-tuning models for continuous reasoning cause catastrophic forgetting?

56 specific questions

• Why does context information fail to override prior training associations?
• Does foundational model training or user priors more strongly shape final outputs?
• How do training associations override context information in language models?
• How does training order affect knowledge acquisition in language models?
• Can models recover knowledge with completely unrelated retraining tasks?
• How do training-data priors influence model defaults when context is ambiguous?
• Can in-context learning substitute for domain-specific training altogether?

35 specific questions

• Can models learn to generate their own training examples effectively?
• Does self-generated training data reduce a model's capability diversity?
• Why does self-generated training data outperform externally curated domain examples?
• What happens when models train on AI-generated content recursively?
• Why does self-generated training data outperform externally sourced data?
• How does the ratio of synthetic to real training data affect model collapse?
• How does diversity loss in synthetic data mirror tail distribution disappearance?

35 specific questions

• Why does post-training suppress alignment faking in some models but amplify it in others?
• Can alignment training create systematic blind spots in threat detection systems?
• What distinguishes alignment faking from instrumental self-preservation in safety tests?
• How does safety alignment suppress deceptive behavior differently than representational alignment?
• Does removing cognitive bias from training signals accidentally break what makes alignment work?
• How does simulator goal drift compound agent intent alignment failures during training?
• Can alignment training be redesigned to permit warranted alarm?

44 specific questions

• Does fine-tuning actually change model capabilities or only output distribution?
• How do finetuning and pretraining improvements differ in their effects on model capabilities?
• What happens to base model capabilities when you apply finetuning?
• Does pretraining data size matter less than base model scale for finetuning?
• Why do proprietary models improve with training while open-source models decline?
• How does behavioral fine-tuning differ from factual knowledge encoding in models?
• Does fine-tuning a small model match fine-tuning a large one?