DATUM

Latent state tokens let a 14B language model match GPT-5.1 on chess collaboration tasks

single source · 1 articles · research · confidence: medium · first seen 2026-09-01 20:00 UTC

An arXiv paper describes a way to connect a language model to a non-language agent (a chess engine) without converting the engine's state into text. Instead, the engine's continuous internal state is projected into the model's token stream as learned latent state tokens, re-encoded as the position advances. On LLAMIA-Bench, six chess collaboration tasks covering imitation, assessment and explanation that neither solves alone, the authors report a 14B model trained this way, LLAMIA, matches or exceeds task specialists and frontier models including GPT-5.1 with tool access, and generalises out-of-distribution where task-specific finetunes collapse. Verbalising the state consistently underperforms, and the gap widens from 4B to 14B parameters.

What this means for you

This is a research result, not a product. If you build systems that orchestrate non-language agents, the method suggests feeding learned state tokens may avoid the performance loss of text summaries, but it has not been demonstrated beyond this chess benchmark.

Key facts

  • ·LLAMIA-Bench comprises six collaborative chess tasks covering behavioural imitation, state assessment and natural-language explanation. source
  • ·The method, latent state internalization, projects a subagent’s continuous representations into the LLM’s token stream as learned state tokens. source
  • ·LLAMIA is a 14B-parameter model trained with latent state internalization. source
  • ·The authors report LLAMIA matches or exceeds task specialists and GPT-5.1 with tool access on all six benchmark tasks. source
  • ·Verbalised integration underperforms, and the performance gap widens as the language model scales from 4B to 14B parameters. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire