Do tools actually expand what language models can reason about?
Explores whether tool access fundamentally breaks through reasoning limits in pure-text models, or merely optimizes existing capabilities. Understanding this distinction clarifies whether tools are luxury features or necessity for genuine capability growth.
Tool-Integrated Reasoning (TIR) — letting a model call a Python interpreter or other external tool mid-reasoning — reliably outperforms pure-text reasoning, but the field has demonstrated this empirically without a principled account of why and when it helps. This paper proves it: TIR enables a strict expansion of both the model's empirical and feasible support, breaking the "invisible leash" that constrains pure-text models. Tools make complex algorithmic strategies practically achievable within finite token budgets — strategies that are otherwise impossible or intractably verbose to express in text alone. Crucially, the advantage is not confined to compute-heavy arithmetic; it extends to problems requiring abstract insight.
On the training side, the paper identifies that reward shaping for TIR is unstable and proposes Advantage Shaping Policy Optimization (ASPO), which directly modifies the advantage function rather than the reward to guide behavior without destabilizing training.
This is the reasoning-side companion to Can models store unlimited facts without growing larger?: one proof concerns factual capacity, this one concerns reasoning reach. Together they give a formal foundation for why agentic harnesses beat bigger models — and they sharpen Does the reasoning cliff depend on how we test models?, which observed empirically that tool access dissolves apparent reasoning ceilings.
Inquiring lines that read this note 11
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
Do language models learn genuine linguistic structure or just surface patterns? Do harness improvements transfer across model scales or memorize shortcuts? How effectively do deterministic tools improve language model reasoning on formal tasks? Why does finetuning cause catastrophic forgetting of model capabilities? Why do benchmark improvements fail to reflect actual reasoning quality? Which computational strategies best support reasoning in language models? Why does verification consistently lag behind AI generation? What memory architectures best support persistent reasoning across extended interactions? How does objective evolution guide discovery better than fixed planning? How can AI agents autonomously learn and transfer skills across tasks?Related concepts in this collection 3
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Can models store unlimited facts without growing larger?
Does external tool use let language models recall facts without being constrained by parameter count? This matters because it could reshape how we scale knowledge capacity beyond architectural limits.
companion proof on the factual-capacity axis
-
Does the reasoning cliff depend on how we test models?
If language models hit a capability wall in text-only reasoning tasks, does that wall disappear when they can use tools? What does this reveal about what we're actually measuring?
the empirical observation this proof explains
-
Can modular cognitive tools unlock reasoning without training?
Can reasoning capabilities be elicited by structuring LLM calls as isolated cognitive operations—understanding, recalling, examining, and backtracking—rather than through reinforcement learning?
a concrete instantiation of tool-augmented reasoning expanding what the base model can do
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Understanding Tool-Integrated Reasoning
- Efficient Tool Use with Chain-of-Abstraction Reasoning
- Position: LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks
- A Comment On "The Illusion of Thinking": Reframing the Reasoning Cliff as an Agentic Gap
- Insert-expansions For Tool-enabled Conversational Agents
- On the Reasoning Capacity of AI Models and How to Quantify It
- Provable Benefits of In-Tool Learning for Large Language Models
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
Original note title
tool-integrated reasoning provably expands an LLM capability frontier — tools unlock strategies impossible or intractably verbose in pure text