TOPIC
Logical Reasoning and Internal Rules
A subject the collection covers, read through 16 synthesis notes.
View as
Can argument schemes be organized by formal principles instead of lists?
Argumentation theory has accumulated 60+ overlapping scheme lists with no principled boundaries. Can a structured classification based on formal ordering principles replace this ad-hoc approach and provide a coherent target space for analysis?
Can three axes organize all possible argument schemes? Do first-order and second-order arguments unify classical and modern divisions? Why do paraphrased definitions work better than expert ones? Can large language models classify argument schemes reliably? Why does argument scheme classification stumble where other NLP tasks succeed?
Three orthogonal axes exhaustively organize argument scheme space First-order second-order distinction unifies classical and modern argument divisions LLM paraphrases outperform formal definitions for scheme classification prompts Large LLMs classify argument schemes only with few-shot examples and descriptions Argument scheme classification carries higher cognitive load than other argument NLP tasks
What are the three orthogonal axes that structure the argument scheme periodic table? How does the first-order and second-order distinction unify classical and modern argument theory? Can smaller scheme inventories or critical questions replace direct scheme classification? What are the nine possible proposition-type combinations in arguments? How do internal and external topoi differ in classical rhetoric? Can formal argumentation structure replace ad-hoc fallacy classifications?
Can retrieval-augmented language models forecast like human experts?
Can language models augmented with search and reasoning match or exceed the forecasting accuracy of competitive human crowd forecasters on events beyond their training data? This tests whether AI can handle genuine predictive judgment.
Can live benchmarks prevent contamination in prediction tasks? Can language models beat human venture capital experts? Can LLMs actually forecast time series better than we think? Can decomposing forecasting into stages unlock numerical and contextual reasoning? How do reasoning models actually break under pressure?
Live contamination-free benchmarks reveal forecasting requires agentic reasoning LLMs surpass human venture capital experts where the expert bar is modest Workflow architecture dominates raw model capability in LLM forecasting tasks Multi-stage decomposition enables forecasting by integrating numerical and contextual reasoning Reasoning models excel at formal tasks but fail at social cognition.
Do newer language model generations improve forecasting ability without additional training? Why are post-cutoff test sets essential for evaluating genuine forecasting ability? What role does retrieval mechanism design play in forecast accuracy? How do AI researcher forecasts compare across different timeline question phrasings? How do search and reasoning workflows improve forecasting performance over base models? Can language models match competitive crowd forecasters on real future events? How much does domain expertise actually improve human forecasting under uncertainty? What privacy-preserving evaluation methods best capture real-world forecasting ability? How should forecasting methods adapt to a post-AGI regime?
How do transformers perform analogical reasoning across domains?
Exploring whether transformers solve analogy problems through a distinct mechanism separate from composition, and whether this involves abstract relational structure rather than memorized computation patterns.
Do transformers actually learn systematic compositional reasoning? How do transformers learn to reason across multiple steps? Do foundation models learn world models or task-specific shortcuts? Does transformer reasoning leave a geometric signature in representation space? Do language models fail at reasoning due to complexity or novelty?
Transformers reduce compositional reasoning to linearized subgraph matching Multi-hop reasoning emerges through three stages with cosine clustering signature Foundation models develop task-specific heuristics, not generalizable world models. Reasoning tasks produce 2-3x higher mean curvature than lexical variation tasks Reasoning failures are driven by instance-level unfamiliarity, not task complexity.
Can three axes organize all possible argument schemes?
Can a small set of orthogonal distinctions—subject vs. predicate, order level, and proposition types—capture the full space of valid argument structures? This matters because it could replace ad-hoc scheme lists with a systematic framework.
Can argument schemes be organized by formal principles instead of lists? Do first-order and second-order arguments unify classical and modern divisions? Can large language models classify argument schemes reliably? Why does argument scheme classification stumble where other NLP tasks succeed? Why do LLMs fail at understanding what remains unsaid?
Formal ordering principles replace ad-hoc argument scheme lists with periodic structure First-order second-order distinction unifies classical and modern argument divisions Large LLMs classify argument schemes only with few-shot examples and descriptions Argument scheme classification carries higher cognitive load than other argument NLP tasks LLMs replicate statistical patterns but lack communicative optimization principles.
How does evaluative stance differ from structural argument analysis? What distinguishes contrasting aspects from related aspects in question structure? Can the eight-dimension rubric predict which question types need decomposition? What are the three orthogonal axes that structure the argument scheme periodic table? How does the first-order and second-order distinction unify classical and modern argument theory? How do first-order and second-order arguments differ in formal structure? What are the nine possible proposition-type combinations in arguments? Can unfilled cells in the periodic table represent undiscovered argument schemes? How do internal and external topoi differ in classical rhetoric?
What three separate factors drive chain-of-thought performance?
Can we isolate and measure the distinct contributions of output probability, memorization, and genuine reasoning to CoT success? Understanding their relative weights matters for knowing when CoT actually reasons versus when it relies on shortcuts.
Does chain-of-thought reasoning reveal genuine inference or pattern matching? Do language models actually use their reasoning steps? Does logical validity actually drive chain-of-thought gains? Does more thinking time always improve reasoning accuracy? Do reasoning traces need to be semantically correct?
Chain-of-thought mimics reasoning form through learned schemata, not genuine inference Language models generate faithful reasoning chains far less often than assumed Invalid CoT reasoning performs nearly as well as valid reasoning Reasoning accuracy peaks then declines with excessive thinking tokens Corrupted reasoning traces teach as well as correct ones
What detection methods can catch each distinct CoT bypass strategy? Does chain-of-thought monitoring fundamentally degrade under optimization pressure? What structural features force users to evaluate the epistemic status of outputs? What explains the 87 percent to 12 percent cliff in plan executability? Why do top performers produce shorter chains of thought in their strongest domains? What happens to chain-of-thought performance across distribution shifts? How does the three-component definition apply to test-time scaling laws? What determines the optimal thinking token threshold for a given task? What interference occurs when planning and synthesis happen in the same component? What saliency patterns distinguish successful from failed chain-of-thought reasoning? How can judges evaluate thinking without seeing the actual thoughts? What mechanism makes keyword probability the strongest predictor of priming?See all 62 inquiring lines on this note →
Can LLMs reason creatively beyond conventional problem-solving?
Explores whether large language models can engage in truly creative reasoning that expands or redefines solution spaces, rather than just decomposing known problems. This matters because existing reasoning methods may miss creative capabilities entirely.
LLM research ideation collapses into narrow clusters despite high novelty LLMs generate novel ideas through unconstrained combination but cannot evaluate them Reasoning topologies classify formally as path graphs, trees, and directed graphs LLMs excel at evidence organization but fail at speculative entity connections Language compositionality enables agents to imagine out-of-distribution goals
What makes colorless green ideas fail where Jabberwocky succeeds? Why do LLMs generate ideas that sound novel but fail during execution? How do constrained versus unconstrained domains flip LLM novelty patterns? Why does embedding evaluation criteria in prompts reduce creative scope? Can few-shot examples narrow generative diversity in creative tasks? Why does LLM research ideation collapse into low diversity despite high novelty? How can LLMs evaluate their own creative outputs for utility and novelty? What graph structures would enable transformational creative reasoning in LLMs? Can prompting for specific creative paradigms improve ideation diversity? When should an LLM engage extended reasoning versus responding directly? Why do research ideation systems suffer from diversity collapse despite high novelty metrics? Can LLMs improve at metaphor if they handle decoupled semantics better?See all 40 inquiring lines on this note →
How do transformers learn to reason across multiple steps?
Does multi-hop reasoning in transformers emerge through distinct learning phases, and what geometric patterns in hidden representations explain when reasoning succeeds or fails?
Do language models actually use their encoded knowledge? Do base models already contain hidden reasoning ability? Does the choice of RL algorithm actually matter for reasoning? Do reasoning cycles in hidden states reveal aha moments? Can neural networks learn compositional skills without symbolic mechanisms?
Encoded knowledge in LMs often fails to causally influence generation Base models contain latent reasoning capability unlocked by minimal training RL algorithm choice is interchangeable because the pretrained prior bounds exploration Reasoning cycles in hidden states correspond to aha moments Compositional generalization emerges from scaling when training covers the task space
What is selective resonance and why do transformers not perform it? How do transformers perform multi-hop reasoning across distant training documents? Do transformers learn generalizable algorithms or instance-based patterns? What graph structures better support multi-hop reasoning than pairwise edges? Do grokking phases correspond to transitions between nesting levels? How does error propagation limit transformer performance on complex tasks? Can symbolic mechanisms improve transformer compositional abilities? How do humans and LMs differ on multi-hop reasoning? Can we detect and measure circuit formation before generalization emerges? What hidden computations happen inside transformer layers during reasoning? Can we decode what individual circuits inside transformers are doing? Why are pairwise relations insufficient for representing higher-order multi-hop reasoning?See all 25 inquiring lines on this note →
Do large language models use one reasoning style or many?
Explores whether LLMs share a universal strategic reasoning approach or develop distinct styles tailored to specific game types. Understanding this matters for predicting model behavior in competitive versus cooperative scenarios.
Why do correct reasoning traces contain fewer tokens? When does explicit reasoning actually help model performance? Why do LLM persona prompts produce inconsistent outputs across runs? Why do reasoning models fail under manipulative prompts? Do personality types shape how AI agents make strategic choices?
Correct reasoning traces in o1-like models are shorter than incorrect ones Task structure determines whether explicit reasoning helps or harms LLM persona prompts fail because model uncertainty overwhelms persona-specific knowledge Manipulative prompts reduce reasoning model accuracy by 25 to 29 percent Personality-primed agents adopt strategically divergent behavior aligned with psychology
Does Habermas's strategic action framework explain LLM dialogue behavior? Can step-level deliberation flags guide other reasoning systems? Why must procedural skills consolidate before strategic reasoning can develop? Why do language models fail at planning despite understanding strategies? Why do LLM agents fail where game-theoretic bots succeed? Should LLM reasoning be studied as latent state trajectories rather than surface text? Do agents inform neighbors when adopting strategies in their reasoning? What role does sequence model in-context learning play in multi-agent cooperation? Can extended reasoning training capture individual strategic thinking styles? How should reasoning prompts adapt based on question complexity and type? Can multi-agent LLM systems overcome diversity collapse through structured disagreement? Which game type reveals minimax reasoning in language models?See all 35 inquiring lines on this note →
Do large language models reason symbolically or semantically?
Can LLMs follow explicit logical rules when those rules contradict their training knowledge? Testing whether reasoning operates independently of semantic associations reveals what computational mechanisms actually drive LLM multi-step inference.
Can large language models translate natural language to logic faithfully? Do foundation models learn world models or task-specific shortcuts? Why do language models ignore information in their context? Does semantic grounding in language models come in degrees? Why do neural networks fail at compositional generalization?
LLMs produce syntactically valid logic but fail at semantic fidelity consistently. Foundation models develop task-specific heuristics, not generalizable world models. Language models fail to integrate context when prior associations are strong LLM grounding is multi-dimensional, not binary or uniform The binding problem explains neural networks' systematic generalization failure
Can LLMs infer situational context the way humans do pragmatically? Why do LLMs fall for and deploy logical fallacies with equal confidence? Why do LLMs fail inter-annotator agreement tests on argument evaluation? How does surface salience compete with background knowledge in model inference? Can prompt-based debiasing overcome entrenched LLM model priors? How do different LLM integration paradigms affect inheritance of pretraining biases? How does LLM-PKG compare to mining product relations directly from interaction data? Does epistemic drift operate the same way across all languages? Can evidence density alone shift an LLM from generation to reasoning? How do transformers perform multi-hop reasoning across distant training documents? How much of LLM reasoning failure stems from missing knowledge versus signal weighting? Do modern architectures in NLP and vision rely on dot products intentionally?See all 258 inquiring lines on this note →
Do language models make rational strategic decisions in games?
Explores whether LLMs consistently apply game-theoretic reasoning to reach optimal strategies, and whether their performance holds as games become more complex. Understanding this matters for deploying LLMs in negotiation and competitive settings.
Different LLMs exhibit distinct strategic reasoning styles tied to game type Negotiation agreement tracking demands bilateral monitoring unlike single-user dialogue systems Workflow architecture dominates raw model capability in LLM forecasting tasks Models fail to track individualized reasoning styles over time Reasoning models fail systematically across exploration, mode selection, and social understanding.
Does logical validity actually drive chain-of-thought gains?
What if invalid reasoning in CoT exemplars still improves performance? Testing whether logical correctness or structural format is the real driver of CoT's effectiveness.
Does chain-of-thought reasoning reveal genuine inference or pattern matching? Do language models actually use their reasoning steps? Why do chain-of-thought examples fail across different conditions? Do large language models reason symbolically or semantically? Do reasoning traces need to be semantically correct?
Chain-of-thought mimics reasoning form through learned schemata, not genuine inference Language models generate faithful reasoning chains far less often than assumed Chain-of-thought exemplars are brittle across order complexity diversity and style Large language models reason through semantic associations not symbolic logic. Corrupted reasoning traces teach as well as correct ones
What makes counterfeiting social warrant different from counterfeiting factual claims? How does validation skill replace production skill in AI systems? What makes accountability and validity-orientation non-behavioral properties? What makes emotional alignment more effective than logic when reasoning errors are exposed? What makes colorless green ideas fail where Jabberwocky succeeds? What makes quasi-beliefs real enough to explain AI behavior? Does good simulation eventually count as genuine realization? Why does item discrimination matter more than surface-level question plausibility? Can proxy evaluation of ideas accurately predict their quality without implementation? What makes Beck's diagram effective for constraining simulated patient behavior? How much does faithfulness vary naturally in reasoning without evaluation pressure? Do recency-focused prompts and in-context examples work equally well for order recovery?See all 197 inquiring lines on this note →
Why does partial formalization outperform full symbolic logic?
Explores whether injecting some symbolic structure into natural language reasoning works better than completely formalizing problems. Matters because it could reveal the optimal balance between structure and semantics for LLM reasoning.
Can large language models translate natural language to logic faithfully? Do large language models reason symbolically or semantically? Can symbolic solvers fix how LLMs reason about logic? Can structured argument prompts make LLM reasoning more rigorous? Do formal language prototypes improve reasoning across different domains?
LLMs produce syntactically valid logic but fail at semantic fidelity consistently. Large language models reason through semantic associations not symbolic logic. Symbolic solver integration enables reliable logical reasoning through structured feedback Structured critical questions force LLMs to check warrants and improve reasoning Formal language prototypes boost cross-domain reasoning and planning performance
What structural constraints matter more than model depth for CF? How does the frame problem differ between symbolic and statistical reasoning systems? How should emotional states integrate into symbolic reasoning systems? What are the five structure types and which tasks does each one suit best? Can symbolic solvers rescue language models from logical reasoning failures? Would hybrid systems combining LLMs with symbolic solvers overcome the retraction limitation? Can we distinguish between semantic and symbolic reasoning in language models? What graph structures would enable transformational creative reasoning in LLMs? Do reasoning languages like Prolog follow the same two-constraint transfer pattern? What makes symbolic operations different from general knowledge questions? Can LLMs improve at metaphor if they handle decoupled semantics better? Why does GraphRAG prioritize corpus completeness while LogicRAG prioritizes query adaptivity?See all 65 inquiring lines on this note →
How much does the order of premises actually matter for reasoning?
When you rearrange the order of logical premises in a deduction task, does it change how well language models can solve it? This tests whether LLMs reason abstractly or process input sequentially.
Why do chain-of-thought examples fail across different conditions? Does training data format shape reasoning strategy more than domain? Do large language models reason symbolically or semantically? How much does demo position alone affect in-context learning accuracy? How do language models perform syllogistic reasoning internally?
Chain-of-thought exemplars are brittle across order complexity diversity and style Training format shapes reasoning strategy 7.5 times more than domain Large language models reason through semantic associations not symbolic logic. Demo position creates spatial bias shifting accuracy up to twenty percent Syllogistic reasoning uses a three-stage circuit contaminated by world knowledge
Does reasoning ability actually degrade with longer inputs?
Explores whether modern language models can maintain reasoning performance when processing long contexts, and whether technical capacity translates to practical reasoning capability over extended text.
Language models degrade at temporal reasoning as task complexity increases Iterative expansion and reflection loops increase RAG knowledge density significantly More thinking tokens hurt LLM accuracy beyond a critical threshold Long-context LLMs resolve retriever-reader imbalance by shifting burden to reader Vector embeddings measure semantic association, not task relevance
How can we measure whether assistance preserved the user's reasoning state? How do readers selectively hold frame-related words in mind? Can input augmentation and rephrasing compensate for smaller model limitations? How does the knowing-doing gap widen as tasks become more complex? How do transformers perform multi-hop reasoning across distant training documents? Does sentence-level granularity capture enough structure for complex reasoning tasks? Does irrelevant content degrade reasoning even when it fits the context window? How does SONAR embedding quality affect downstream reasoning accuracy? What distinguishes genuine reasoning activation from memorization-assisted answer recall? Can meaning-level metrics like Semantic Entropy avoid length bias? Can latent reasoning architectures work as retrofits to existing models? Can context compression preserve what matters without introducing bias?See all 155 inquiring lines on this note →
Can models identify what information they actually need?
When a reasoning task is missing a key piece of information, can language models recognize what's absent and ask the right clarifying question? QuestBench tests this capability directly.
Why do reasoning models overthink ill-posed questions? Can language models recognize when text is deliberately ambiguous? Does reasoning fine-tuning make models worse at declining to answer? Why do LLMs struggle to connect unrelated entities speculatively? Can models learn to ask clarifying questions instead of guessing?
Reasoning models lack critical thinking to reject ill-posed questions Language models fail systematically at recognizing and disambiguating multiple interpretations Reasoning fine-tuning degrades model abstention capacity by approximately 24 percent LLMs excel at evidence organization but fail at speculative entity connections Models can be trained to proactively identify missing information and request clarification
Can AI systems identify important unanswered questions that emerge during reasoning? Can models identify what information they are missing in underspecified problems? How do humans decide which level of clarification to request? Can models identify information gaps without just guessing or refusing to answer? What makes some clarifying questions more useful than others? What specific information must be exported from the language system? How does ambiguity detection connect to models' ability to ask clarifying questions? Can models learn to identify what information is missing from questions? Do reasoning models overthink ill-posed questions instead of recognizing incompleteness? Can models identify what information they are missing in underspecified tasks? How does proactive critical thinking enable models to identify missing information? Can language models ask clarifying questions when sentences are ambiguous?See all 21 inquiring lines on this note →
Do first-order and second-order arguments unify classical and modern divisions?
Does the formal distinction between first-order and second-order arguments map onto both the classical internal-external topoi divide and the modern reasonable-fallacious distinction? If so, it would reveal a single structural axis underlying two separate critical traditions.
Can three axes organize all possible argument schemes? Can argument schemes be organized by formal principles instead of lists? Can formal argumentation make AI decisions truly contestable? Can structured argument prompts make LLM reasoning more rigorous? What happens to social order when AI removes ritual constraints?
Three orthogonal axes exhaustively organize argument scheme space Formal ordering principles replace ad-hoc argument scheme lists with periodic structure Formal argumentation frameworks make AI decisions contestable in ways unstructured outputs cannot Structured critical questions force LLMs to check warrants and improve reasoning AI conversation lacks the ritual machinery that sustains human social order.
What are the three orthogonal axes that structure the argument scheme periodic table? How does the first-order and second-order distinction unify classical and modern argument theory? How do first-order and second-order arguments differ in formal structure? What makes an argument fallacious according to formal linguistic criteria? How do internal and external topoi differ in classical rhetoric?