FILD: Relational State Fields for Evidence-Gated Revision in Lookahead Decoding

While greedy decoding provides a strong default, bounded lookahead can reveal better continuations beyond the locally preferred token. The challenge is to use these partial futures to decide when to revise the model's native trajectory. We introduce FILD, a training-free decoder that builds evidence for revision from the model's own lookahead computation. Using the same frozen model, FILD uses self-verification to assess continuation quality and combines these assessments with native token likelihoods to identify a challenger. This proposal reflects a fixed-horizon assessment; the unfolding candidate futures reveal how support for it develops relative to the native continuation. FILD organizes their hidden states into a relational state field, assessing whether the challenger is closer than the native continuation to the same reference branches across model depths and continuation horizons. For unfinished candidate futures, revision additionally requires support that agrees across depths and persists across adjacent horizons. If all branches terminate within the lookahead horizon, the proposal suffices for revision. This procedure reuses states already computed during exploration, requiring no auxiliary trained probe or router. Across four language models and six generation benchmarks, FILD achieves the strongest aggregate performance among the evaluated decoders and the highest six-task average for each model. Component analysis shows that self-verified proposals improve factual question answering, with relationally guided revision providing further gains on reasoning tasks.