Get Started
Topic · 48 recaps
Reasoning
Methods that improve how language models think through problems — chain-of-thought, search, verification, and the reasoning-trained model families that have emerged since o1-style training.
Play all
...
Posts
Questions
Home
Topics
Search
Library
Questions researchers are working on
Follow a question through Rcap’s explanations and the latest papers addressing it.
Search
Can any finite formal system autonomously derive every theorem within its expressive scope?
A system may be able to express a theorem without having an autonomous procedure that produces it. The central issue is whether finite formal systems can be complete with respect to the theorems they can express.
Can chain-of-thought monitoring detect consequential computation hidden in semantically irrelevant filler tokens?
Language models may gain task performance from semantically irrelevant filler tokens without making the relevant computation interpretable in their visible reasoning. This complicates the use of chain-of-thought as evidence of what a model has computed.
Can chain-of-thought monitoring detect preferences received through tools or inferred from raw artifacts?
Chain-of-thought monitoring assumes that a model’s reasoning trace reveals the information influencing its answer. Preferences delivered through tool returns or inferred from unprocessed artifacts may affect answers without being clearly verbalized in the trace.
Can language models infer others’ mental states as social interactions evolve under unreliable information?
Socially grounded tasks require models to use interaction history, infer what participants know or intend, and distinguish reliable from unreliable information. Existing evaluations often isolate these demands, making performance in changing social environments difficult to characterize.
Can multilingual LLMs maintain mathematical reasoning when equivalent inputs use different word order or voice?
A mathematically equivalent prompt can be expressed through reordered constituents or active-passive voice. Models that rely on surface form may change their answers even when the underlying entity-quantity relations remain unchanged.
Do multilingual attention heads that retrieve context also control transitions into the target language during reasoning?
In multilingual Transformers, retrieving information from context and producing reasoning in a specified target language may involve different internal components. It is unclear whether the same attention heads perform both roles.
Do source labels bias human and LLM judgments of logical fallacies differently?
Labels about who produced a comment can influence judgments of its credibility and logical quality independently of the argument itself. It remains unclear whether humans and language models are affected by these source cues in the same way.
How can a single search agent improve multi-hop web research without sub-agents or test-time verification?
Multi-hop web research requires an agent to connect evidence across several pages while retaining the information needed for later steps. Long search trajectories can overwhelm the available context, making both training and reliable performance difficult without additional agents or verification passes.
How can agentic vision-language models acquire and use necessary external evidence without redundant tool calls?
Complex image-grounded questions may require visual details or external knowledge unavailable in the initial input. Models may pursue irrelevant evidence or fail to extract useful information from tool outputs, while final-answer supervision does not clearly teach effective evidence acquisition.
How can agents choose task-dependent world-model rollout horizons for effective multi-step planning?
Learned world models let agents imagine future environmental states without interacting with real environments. Single-step or fixed-horizon imagination may provide insufficient foresight when tasks require different amounts of planning and progress changes over time.
How can agents find people across cameras from vague witness clues under spatial-temporal and turn constraints?
Witness accounts may be partial or ambiguous, while relevant observations are distributed across camera locations and time. An agent must choose questions and searches before its interaction budget runs out.
How can AI agents adapt execution routes as runtime evidence invalidates their planned continuation?
Agents often commit to a multi-step execution route before intermediate outcomes are known. When those outcomes change what should happen next, stale actions or broad replanning can waste computation and discard useful progress.
How can AI systems translate scientific reasoning into verifiable lab workflows while respecting changing states and physical constraints?
Scientific reasoning must be connected to operations that transform physical samples and equipment over time. Without a computable account of laboratory state and constraints, planned actions may not be executable or safely verifiable.
How can alignment systems infer the multiple criteria behind human pairwise preferences?
Pairwise preference labels record which option people choose, but not the interacting considerations behind that choice. This makes it difficult to build preference models that are both faithful to judgments and interpretable.
How can audio deepfake detectors identify and localize manipulation when genuine and fake content coexist?
A whole-clip label can conceal which time intervals or overlapping sources provide evidence of manipulation. This makes mixed-authenticity audio decisions difficult to interpret and verify.
How can autoformalization preserve diverse faithful statements that improve prover search under a fixed budget?
A single formal translation can hide other faithful formulations, while syntactic differences among equivalent statements can change how a prover searches. Correctness-only, single-output evaluation therefore misses effects that matter for downstream proving.
How can automated Lean 4 proof search reuse promising partial proofs without later revisions degrading them?
In real-world Lean projects, proof attempts may contain useful partial progress even when they fail to compile. Continuing the wrong attempt can discard that progress, while exploring too many alternatives wastes inference calls.
How can autonomous robots adapt processing for every admitted event as new situations arrive?
Robots encounter events that differ in context, familiarity, and required cognitive effort, while fixed task-driven procedures attend only to selected cases. The difficulty is maintaining appropriate coverage as events arrive continuously and earlier cases may await additional evidence.
How can autonomous-driving scenario generators reliably induce collisions at a requested region of the target vehicle?
Existing autonomous-driving scenario generators can produce crashes but offer limited control over where the target vehicle is struck. This makes it difficult to construct tests for region-specific collision behavior.
How can autoregressive LLM decoding generate multiple tokens in parallel at large batch sizes without sacrificing quality?
Autoregressive LLMs repeatedly predict one next token at a time, limiting throughput and increasing serving costs as batch sizes grow. Parallelizing generation is difficult because it must retain the quality and distribution of the original autoregressive model.
Previous
1 / 10
Next