UIUC's TimeClaw Beats 235B Models at Time Series Analysis Using Fewer Tokens
TimeClaw gives LLM agents a time-series-native runtime, evolving toolbox, and episodic memory — beating models up to 235B parameters while using fewer tokens
Takeaways
- TimeClaw is an agentic harness from UIUC that wraps frozen LLMs with time-series-native tools, episodic memory, and a self-evolving toolbox for contextualized temporal reasoning.
- It addresses two fundamental flaws in current LLM-on-time-series approaches: tokenization distorts numerical proximity, and agentic workflows weren't designed for structured temporal signals.
- On the TSRBench benchmark, TimeClaw achieves 49.8% accuracy — outperforming Qwen3-235B and all specialized time-series LLMs — while using the fewest tokens of any method tested.
- A capability evolution loop autonomously synthesizes reusable tools (e.g.,
portfolio_sharpe,capm_regression) from successful training trajectories, yielding a 38.9% improvement on the TSAIA finance benchmark. - Multimodal memory retrieval uses both a text embedding of task context and a 20-dim numerical fingerprint of the series, enabling two-stage retrieval that matches tasks by both meaning and signal shape.
- Code is open-source (Apache 2.0) and supports OpenAI, Google Gemini, and Anthropic Claude backends via LangChain.
Most LLM-based time series tools make the same mistake: they dump raw numbers into the model's context window and hope for the best. A team from the University of Illinois Urbana-Champaign argues this approach is fundamentally broken, and their new framework TimeClaw is built to fix it.
The core problem: two misalignments
When you serialize a time series into text tokens, two things go wrong simultaneously. The first is datatype misalignment: numerical temporal signals encode statistical properties such as trends and periodicity that can be distorted when serialized into tokens and processed through language-oriented attention mechanisms. The second is subtler but equally damaging.
The second is agentic process misalignment: existing agentic workflows are primarily designed for text-centric information processing. Long numerical sequences may further burden the context window and distract the agent from reliable temporal reasoning. The paper even proves this formally: numerically close values like 99.999999 and 100.0 can be farther apart in token space than values that are numerically distant, meaning the model's internal notion of "similarity" is simply wrong for numbers.

There's also a context window problem. Long serialized time series increase the number of irrelevant or weakly relevant tokens competing for the same unit attention mass. If some tokens act as attention sinks and absorb disproportionate mass, the remaining mass available to temporally relevant positions is further reduced. In other words, the model literally stops paying attention to the parts of the series that matter.
The harness approach
TimeClaw's answer is an agent harness , a structured runtime that wraps a frozen LLM and reshapes how it interacts with temporal data. Instead of forcing numbers into the token stream, the harness keeps time series objects in a structured workspace and exposes them only through typed, executable interfaces. The agent analyzes them through auditable temporal operations, while intermediate analytical states are maintained in a structured workspace.