How does active reasoning through interaction differ from passive single-turn problem solving?
This explores the contrast between reasoning that unfolds through back-and-forth exchange — turn-taking, asking, branching — versus a model trying to solve everything inside one silent pass, and what the corpus says each mode is good and bad at.
This explores the contrast between reasoning that unfolds through back-and-forth exchange — turn-taking, asking, branching — versus a model trying to solve everything inside one silent monologue. The corpus suggests the difference isn't mainly about having more time to think; it's about structure. A single-turn solver runs as an uninterrupted internal monologue, and several notes show that monologue has a characteristic failure shape: it wanders. Reasoning models explore "like tourists, not scientists," abandoning promising paths prematurely and drifting into invalid branches, with success probability collapsing as problems get deeper Why do reasoning models abandon promising solution paths? Why do reasoning LLMs fail at deeper problem solving?. And more thinking doesn't rescue it — accuracy actually peaks and then declines as a model pours more tokens into one pass, overthinking easy problems and underthinking hard ones Does more thinking time always improve reasoning accuracy?.
The interesting move in the collection is that several papers recover the *benefits of interaction without needing a second party* — they make a single model reason against itself. DialogueReason restructures one model's internal chain as a conversation between distinct agents, and that dialogue format beats plain monologue precisely on tasks needing multiple approaches, because it breaks the fixed-strategy, fragmented-attention rut of solving in one voice Can dialogue format help models reason more diversely?. In the same spirit, separating a "decomposer" from a "solver" prevents planning and execution from interfering with each other Does separating planning from execution improve reasoning accuracy?, and modular cognitive tools — reasoning steps run as isolated tool calls — lifted GPT-4.1 on competition math without any retraining Can modular cognitive tools unlock reasoning without training?. The throughline: interaction, even simulated, imposes turn boundaries that a free-running monologue lacks, and those boundaries are where the gains come from.
There's a second, more literal sense of "reasoning through interaction" — reasoning with a *user* rather than at a problem. Here the corpus surfaces something you might not expect: today's models are passively built. Optimizing for next-turn reward structurally strips out initiative, so agents wait to be asked instead of clarifying, probing, or volunteering Why do AI agents fail to take initiative?. Yet that passivity is trainable away — proactive behavior rose from near-zero to ~74% with RL — and proactivity pays off concretely, cutting conversation turns by up to 60% by offering relevant information before it's requested Could proactive dialogue make conversations dramatically more efficient?. So the single-turn solver isn't just a reasoning style; it's a behavioral default the training objective quietly enforces.
What's worth taking away is that "interaction" turns out to be a way of *organizing exploration*, not just a UI choice. Abstractions that force breadth-first search beat piling on depth Can abstractions guide exploration better than depth alone?, and the quality of any extended thinking depends on whether training taught the model to use those steps for gap analysis rather than self-doubt Does extended thinking help or hurt model reasoning?. Passive single-turn solving fails not because it thinks too little but because nothing in it forces the model to branch, check, hand off, or ask — and the whole cluster of dialogue, modular, and proactive methods is really a set of ways to manufacture those interruptions.
Sources 10 notes
Reasoning LLMs exhibit two reinforcing failures: wandering (invalid exploration) and underthinking (premature path-switching). Decoding-level interventions like thought-switching penalties improve accuracy without fine-tuning, suggesting viable solutions exist but are abandoned prematurely.
Current reasoning models lack the three properties of systematic exploration: validity, effectiveness, and necessity. This causes success probability to drop exponentially with problem depth, making medium problems solvable but deep problems catastrophically harder.
Increasing thinking tokens from ~1,100 to ~16K reduced benchmark accuracy from 87.3% to 70.3%, revealing a non-monotonic relationship where models overthink easy problems and underthink hard ones.
DialogueReason, which structures a single model's internal reasoning as dialogue between distinct agents in separate scenes, overcomes monologue reasoning's fixed-strategy and fragmented-attention weaknesses, especially on tasks requiring multiple problem-solving approaches.
Modular architectures with separate decomposer and solver models outperform monolithic LLMs, with decomposition ability transferring across domains while solving ability does not. The separation prevents planning-execution interference and produces more generalizable skills.
Four cognitive tools implemented as sandboxed LLM calls improved GPT-4.1 on AIME2024 from 26.7% to 43.3% without any RL training. Modularity enforces operation isolation that pure prompting cannot guarantee, eliciting pre-existing reasoning capability.
Research shows next-turn reward optimization structurally removes initiative from models, but proactive behaviors like critical thinking and clarification-seeking are trainable (0.15% to 73.98% with RL). The core challenge is balancing proactivity with civility to avoid intrusion.
Simulations show proactivity—providing relevant information without being asked—cuts dialogue turns by 60% in medium-complexity domains. This behavior mirrors human conversation and Grice's maxims but is almost entirely absent from AI datasets and research benchmarks.
RLAD jointly trains abstraction and solution generators, showing that allocating test-time compute to diverse abstractions outperforms parallel solution sampling at large budgets. Abstractions create structured breadth-first exploration that prevents the underthinking failure mode of depth-only reasoning chains.
Vanilla models use thinking mode counterproductively, inducing self-doubt that degrades performance. RL training reverses this, transforming the same mechanism into beneficial gap analysis. Training mediates reasoning quality, not just quantity.