STATe-of-Thoughts: Interpretable LLM Reasoning

Updated 2 July 2026

STATe-of-Thoughts is a framework that uses discrete action templates and beam search to achieve high-level, interpretable control in language model reasoning.
It replaces stochastic token sampling with structured, template-guided inference, enabling clearer attribution between reasoning strategies and output quality.
Empirical results demonstrate up to 2× more output diversity and enhanced interpretability compared to traditional chain-of-thought methods.

STATe-of-Thoughts (STATe) is a framework for interpretable, discrete-action-guided reasoning in LLMs, designed as an explicit alternative to stochastic token-level diversification methods. STATe-of-Thoughts systematizes high-level control and interpretability in LLM inference-time-compute (ITC), achieving greater output diversity and enabling direct attributions between reasoning strategies and output quality. STATe draws on and reifies developments in the conceptualization of reasoning tokens as computational state in LLMs and advances action-templated search in place of temperature sampling.

1. Conceptual Foundations: The State-over-Tokens (SoT) Paradigm

STATe-of-Thoughts is rooted in the State-over-Tokens (SoT) conceptual model, which reframes the role of intermediate “reasoning tokens” in LLMs. In the SoT view, an autoregressive Transformer LLM is considered a sequence-to-token function $\mathcal{M}: \{\text{token sequences}\} \to \text{token}$ , with each inference step updating the “state” by concatenating the generated token to the prefix: $s_{k+1} = s_k \oplus t_{k+1}$ . Critically, the entire persistent computational state, as used by the model across generation cycles, is externalized in these prefix tokens—the LLM’s internal activations (KV cache) do not persist across calls; only the emitted token sequence does (Levy et al., 14 Dec 2025).

This SoT view sharply contrasts with Chain-of-Thought (CoT) prompting, which treats intermediate tokens as a human-verifiable narrative of reasoning. In SoT, these tokens serve as a model-encoded, potentially opaque, forward-looking state enabling correct downstream reasoning steps, not as a faithful reflection of reasoning visible to a human reader. Thus, reasoning tokens, even if structured as natural language due to pretraining, are functionally a form of external working memory or "whiteboard" that the model uses to accumulate and persist partial results throughout the computation (Levy et al., 14 Dec 2025).

2. Structured Reasoning and Discrete Action Templates

STATe-of-Thoughts operationalizes state management and interpretable reasoning by introducing structured action templates. Rather than relying on token-level stochasticity (e.g., high-temperature sampling), STATe conducts explicit search over high-level, interpretable actions that encode reasoning strategies (e.g., thematic focus, discourse move). Each action template specifies a prefix (explicit textual bias for the next step) and internal reasoning guidance (rationale to include for model prompting) (Bamberger et al., 15 Feb 2026).

At each step of search, a controller policy $\pi_C$ selects up to $n$ actions $a_i^j$ from a finite action-template space $\mathcal{A}$ . Given the current state $s_{i-1}$ and action $a_i^j$ , the generator model $P_G$ produces the next reasoning step $z_i^j$ , which is concatenated for future steps. The process is guided by a process reward model (PRM) and an outcome reward model (ORM) that score intermediate and final states, respectively. This architecture enables breadth-first beam search over the combinatorial action space, with explicit branching on diverse high-level reasoning strategies (Bamberger et al., 15 Feb 2026).

The table below summarizes key contrasts:

Method	Diversification Mechanism	Role of Reasoning Steps
CoT, ToT	Stochastic (token-level temp)	Human-readable narrative
STATe	Discrete action branching	Template-guided strategy
SoT (concept)	None (re-interpretation only)	Persistent computation state

3. STATe-of-Thoughts: Algorithmic Framework

The STATe framework is formalized as search over composite states $s_{k+1} = s_k \oplus t_{k+1}$ 0, where $s_{k+1} = s_k \oplus t_{k+1}$ 1 is the initial prompt and $s_{k+1} = s_k \oplus t_{k+1}$ 2 is the sequence of reasoning steps so far. The search objective is to maximize final outcome score, approximated via breadth-first beam search.

A high-level pseudocode outline of the inference loop is:

$\pi_C$ 3 This cycle—Plan via controller $s_{k+1} = s_k \oplus t_{k+1}$ 3, Generate via generator $s_{k+1} = s_k \oplus t_{k+1}$ 4, Evaluate via $s_{k+1} = s_k \oplus t_{k+1}$ 5, Select—produces explicit, diverse, and auditable chains of reasoning choices (Bamberger et al., 15 Feb 2026).

4. Comparative Empirical Properties and Performance

STATe-of-Thoughts directly addresses several limitations of prior ITC approaches:

Response diversity: STATe achieves up to 2× more distinct semantic outputs than best-of-n or Tree-of-Thoughts approaches using high-temperature sampling on NoveltyBench. For Qwen3-30B at $s_{k+1} = s_k \oplus t_{k+1}$ 6, STATe yields a distinct output diversity $s_{k+1} = s_k \oplus t_{k+1}$ 7 versus CoT+ActionSpace $s_{k+1} = s_k \oplus t_{k+1}$ 8 and CoT $s_{k+1} = s_k \oplus t_{k+1}$ 9.
Interpretability: Explicit action sequences allow attribution of output features to templates. Sequential regression models (incorporating order and transitions between actions) explain up to 57% of outcome variance under strict synthesis, compared with $\pi_C$ 036% for presence-based models (Bamberger et al., 15 Feb 2026).
Steering generation: Targeted trajectory exploration—ranking action sequences based on estimated quality—yields arguments that outperform randomly sampled or topic-only baselines in 78–81% and 64–77% of pairwise comparisons, respectively.
Controllable branching: Beam search with action templates supports both adaptive tree depth and systematic exploration of reasoning strategies.
Ablations: Under faithful synthesis (light rephrasing, same step order), controlling for argument length raises $\pi_C$ 1 from 0.38 to 0.82, indicating how much output structure is driven by template selection and combinatorial action dynamics.

5. Integrations with Latent State Reasoning and Thinking States

STATe-of-Thoughts is complemented by parallel advances in latent reasoning mechanisms, notably the "Thinking States" approach (Amos et al., 9 Feb 2026). Thinking States maintain recurrent, compressed vector-valued representations of subthoughts generated during input processing, which are injected at specific model layers. This allows state-tracking and latent reasoning over long sequences without extending context length, outperforming CoT on extrapolation tasks (e.g., 98–100% OOD accuracy in state tracking for $\pi_C$ 2, versus 64% for CoT). The Thinking States approach demonstrates:

Fixed-size, compressible subthoughts as latent states,
Fully parallelizable, supervised training via teacher-forcing,
Recurrence-based length extrapolation,
Reduced inference latency compared to full CoT rationales.

STATe-of-Thoughts' explicit, symbolic action templates and latent-state mechanisms like Thinking States are complementary: STATe provides high-level discrete pathway control and interpretability, while Thinking States manage continuous state information for complex state-tracking and long-range dependencies (Amos et al., 9 Feb 2026, Levy et al., 14 Dec 2025).

6. Interpretability, Limitations, and Future Directions

STATe-of-Thoughts offers interpretability by making action choices explicit, facilitating rigorous hypothesis testing about step order and reasoning transitions. Analyses such as sequential feature regression (Model M2) provide quantifiable attributions between template usage and quality outcomes.

Noted limitations include:

Domain-specific action space design: Template engineering requires careful selection of reasoning dimensions (content, structure, style).
Demonstrated scope: Experiments use open-source LLMs; scaling to proprietary or specialized models remains an open question.
Nature of analysis: Most results are correlational; causal links between action templates and quality require controlled experiments or interventional designs.

Prominent research directions include: causal inference over action trajectories, integration of Monte Carlo tree search (MCTS) or reinforcement learning, inclusion of adversarial and multi-turn scenarios, and extending latent-state reasoning mechanisms into generation phases.

7. Conclusion

STATe-of-Thoughts systematizes interpretable, high-level control over LLM reasoning through discrete action templates, providing empirical and theoretical advantages over traditional stochastic ITC methods. By grounding reasoning steps in explicit, auditable action sequences, STATe enables nuanced interpretability, diversity, and targeted quality improvements while aligning with state-oriented conceptualizations from SoT and complementing latent-state methods such as Thinking States (Levy et al., 14 Dec 2025, Bamberger et al., 15 Feb 2026, Amos et al., 9 Feb 2026). These developments anchor a principled research trajectory directed at understanding, controlling, and attributing LLM reasoning processes at both symbolic and latent levels.