Line of inquiry
Inquiring lines›How do architecture and constraint…›What determines model reasoning ca…›this line of inquiry
Why does fine-tuning for specific tasks degrade broader reasoning capability?
A broader line of inquiry — a family of 124 specific questions the research asks around this. Follow one into its inquiring-line page, or move sideways to a related line below.
Questions in this line of inquiry 124
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Why do instruction following and reasoning capability trade off in training?
- Can fine-tuning ever teach semantic inference instead of amplifying training shortcuts?
- Why do reasoning gains resist clear attribution to specific training changes?
- Can reasoning fine-tuning improve both capability and instruction compliance together?
- Why does instruction tuning hurt knowledge-intensive tasks more than reasoning tasks?
- How does optimizing for accuracy during training degrade downstream reasoning quality?
- Does fine-tuning models for specific tasks destroy their ability to reason?
- Can models maintain reasoning-output coupling while improving domain accuracy?
- Does supervised fine-tuning improve accuracy while damaging the quality of reasoning?
- How does data quality mismatch create reasoning degradation in supervised fine-tuning?
- Can small demonstration sets unlock general reasoning without large question data?
- Do instruction-tuned models learn tasks or just output format distributions?
- Does fine-tuning push models toward reasoning shortcuts that bypass the chain entirely?
- Does task diversity in pretraining data transfer reasoning better than larger models?
- Why does instruction-following capability decrease as models scale stronger?
- How much does pre-training frequency predict reasoning task performance?
- Why does fine-tuning degrade reasoning quality even as accuracy improves?
- Why do benchmark scores rise while reasoning quality declines?
- How does supervised fine-tuning degrade chain-of-thought faithfulness over time?
- How do single training examples activate reasoning capabilities in language models?
- Why do SFT models memorize patterns instead of learning generalizable reasoning?
- How does question difficulty and breadth affect what models learn to reason?
- Does token-level reasoning during pretraining improve general reasoning without task-specific supervision?
- Does supervised fine-tuning improve reasoning or just response formatting?
- Does SFT degrade reasoning quality while improving domain accuracy?
- Can instruction-level interventions fix memory-induced reasoning failures in practice?
- Can smaller amounts of diverse reasoning demonstrations replace exhaustive factual training data?
- How much does training composition affect syntactic versus reasoning performance?
- Does domain training degrade reasoning ability even when benchmark scores rise?
- Can training improve reasoning coherence without improving actual correctness?
- Can models learn better from critiquing errors than imitating correct responses?
- Can models be trained to explain instead of imitate answers?
- Can models learn to select exemplars based on reasoning skills rather than complexity?
- Can format adaptation alone explain why reasoning enrichment improves instruction following?
- Does looped pretraining build reasoning more efficiently than supervised fine-tuning?
- Why does evaluating errors teach more than imitating correct responses?
- Why does eliminating proxy-model filtering improve reasoning emergence in pretraining?
- Can correct outputs mask reliance on surface heuristics rather than deep understanding?
- How does evaluation setting affect measured reasoning capabilities in language models?
- Why does fine-tuning improve some capabilities while degrading others?
- Why does explicit theory injection work better than example-based learning for reasoning tasks?
- How does critique fine-tuning on one problem unlock broader reasoning?
- Why do single examples trigger large reasoning improvements in models?
- Can we reverse the instruction-following deficit through targeted training?
- Can curriculum learning by reward variance improve reasoning scalability?
- Does instruction tuning affect how much models yield to disagreement?
- How can one training example improve reasoning across thousands of unseen problems?
- Do higher asymptote recipes unlock genuinely novel reasoning strategies?
- Does training on critiques of noisy responses produce deeper understanding than imitating correct ones?
- How does scaling reasoning capability actually reduce instruction-following ability?
- Can reasoning catalyst data serve as a stable foundation for test-time training?
- Why does fine-tuning models for continuous reasoning cause catastrophic forgetting?
- Can models recover knowledge with completely unrelated retraining tasks?
- Why do open-source models trained on proprietary outputs still fail at reasoning?
- Why does supervised fine-tuning degrade reasoning quality despite raising accuracy?
- How does preference learning differ from supervised finetuning for reasoning?
- Can contamination-free evaluation distinguish between memorization and genuine prediction ability?
- Why do strong models struggle more with instruction following than mid-tier ones?
- Can verifier-guided search catch factual errors that reasoning training cannot?
- How do behavioral differentiation and paraphrase stability trade against accuracy?
- Does fine-tuning improve domain accuracy at the cost of reasoning quality?
- Why does distillation transfer reasoning patterns with few examples?
- Why does exploration quality matter more than learner network depth?
- Why do foundation models develop heuristics instead of world models?
- Does partial trace guidance work better than curriculum learning for hard problems?
- Does training on curated solutions transfer to unseen problem types?
- What makes procedural knowledge in documents generalize better than facts?
- Why does mixing reasoning traces from different teachers destabilize learning?
- Does fine-tuning on NLI tasks reduce or amplify frequency bias?
- Does specialized training in one domain create capability cliffs elsewhere?
- Why do reasoning tasks improve more than retrieval from lookup memory?
- Why does domain accuracy improve while reasoning quality degrades after supervised fine-tuning?
- How does optimizing model performance decouple from optimizing user interpretability?
- Can training on diverse related tasks be more efficient than task-specific training?
- Can instruction tuning succeed without explicit task understanding?
- Why does instruction specificity matter more than intervention timing for drift correction?
- Do models naturally learn to ask clarifying questions without explicit supervision?
- How do timing and search internalization interact during reasoning post-training?
- How much task-similar finetuning data does test-time training actually need?
- Why does imitation learning create a ceiling for reasoning capability?
- Do task-specific heuristics improve gradually or appear suddenly at scale?
- Why is in-context learning brittle to the order of examples presented?
- Do larger instruction-tuned models actually track user beliefs better?
- Can a model be strong at MMLU but weak at long-horizon tasks?
- What makes reasoning-specific post-training different from standard parameter scaling?
- Can a model imitate ChatGPT's style by copying its phrases without matching its capabilities?
- Why does naive personalization fine-tuning destroy generalist reasoning?
- Can instance seeds work for tasks beyond language understanding benchmarks?
- Why do SFT datasets fail to teach models error correction from correction traces?
- Does instruction tuning optimize language models for rhetorical polish over logical consistency?
- Can a single model trained on two tasks predict untrained decision tasks?
- Why do task-specific heuristics fail at generalizing to sparse data regions?
- Why does reconstructing hidden thought processes during training improve cross-domain transfer?
- Can inference-time interventions replace training-level alignment fixes for grounding failures?
- Why does open-ended search outperform direct curriculum building approaches?
- What training regimes confound surface mechanisms with their actual causes?
- Why does SFT reduce reasoning quality even when improving domain accuracy?
- Can a single SAE feature control reasoning behavior across model families?
- Why does adversarial training force deeper reasoning than surface imitation?
- Does semantic auditing of instruction data improve performance uniformly across different model sizes?
- How do foundation models develop task-specific heuristics instead of world models?
- Why does semantic similarity retrieval enable skill transfer to novel situations?
- What makes training-free approaches like Soft Thinking preferable to SoftCoT?
- Why does SFT fail when expert demonstrations are too long for small models?
- Does training on granular tasks beat training on the full function calling problem?
- Why do familiar patterns that support correct answers sometimes drive errors?
- How does a single training example trigger phase transitions in reasoning output?
- How can minimal pairs expose reasoning failures that single-instance accuracy metrics miss?
- Does fine-tuning on NLI tasks amplify or reduce frequency bias in language models?
- Can reasoning improvements be attributed when optimizer and scaffold are unknown?
- Why does a replay mechanism prevent reasoner skills from over-specializing?
- How does fine-tuning on natural language inference affect fallacy susceptibility?
- Can a single correct example seed exponential improvement in mathematical reasoning?
- Can style imitation training successfully erase a model's distinctive word preferences?
- Do newer language model generations improve forecasting ability without additional training?
- How does action-level decomposition differ from token-level imitation in supervision?
- Why does critique training produce deeper understanding than imitation training?
- Does grokking in modular arithmetic follow the same three-phase learning trajectory?
- Can training-time debate work on tasks beyond mathematics and verifiable answers?
- Does ChatGPT bias outputs toward solutions over principled explanations?
- Why do foundation models develop task-specific heuristics instead of causal understanding?
- Why does NLI fine-tuning amplify frequency bias instead of teaching inference?
- Does balancing four pedagogical capabilities improve tutoring or create performance tradeoffs?
- What specific tasks should evaluate whether models understand pedagogical sequencing?