LLM Reasoning Is Latent, Not the Chain of Thought
Wenshuo Wang
Abstract
This position paper argues that large language model (LLM) reasoning should be studied as latent-state trajectory formation rather than as faithful surface chain-of-thought (CoT). This matters because claims about faithfulness, interpretability, reasoning benchmarks, and inference-time intervention all depend on what the field takes the primary object of reasoning to be. We ask what that object should be once three often-confounded factors are separated and formalize three competing hypotheses: H1, reasoning is primarily mediated by latent-state trajectories; H2, reasoning is primarily mediated by explicit surface CoT; and H0, most apparent reasoning gains are better explained by generic serial compute than by any privileged representational object. Reorganizing recent empirical, mechanistic, and survey work under this framework, and adding compute-audited worked exemplars that factorize surface traces, latent interventions, and matched budget expansions, we find that current evidence most strongly supports H1 as a default working hypothesis rather than as a task-independent verdict. We therefore make two recommendations: the field should treat latent-state dynamics as the default object of study for LLM reasoning, and it should evaluate reasoning with designs that explicitly disentangle surface traces, latent states, and serial compute.
AI Impact Assessments
(3 models)Scientific Impact Assessment
Core Contribution
This position paper proposes a conceptual framework for studying LLM reasoning by disentangling three explanatory factors that are typically confounded in current research: surface chain-of-thought traces (S), latent-state trajectories (Z), and generic serial compute budget (B). The paper formalizes three competing hypotheses—H1 (latent-trajectory mediation), H2 (surface-CoT mediation), and H0 (generic serial compute)—and argues that the existing evidence most strongly supports H1 as the default working hypothesis. The key intellectual contribution is the explicit factorization of these three factors and the argument that the field's default assumption should shift from treating surface CoT as the reasoning object to treating latent-state dynamics as primary.
Methodological Rigor
The paper's theoretical framework is clearly articulated. The three-hypothesis structure is well-motivated and the asymmetric predictions each hypothesis makes are logically sound. The literature synthesis in Section 3 is organized by diagnostic force rather than paper count, which is a principled approach.
However, the empirical adjudication program (Section 5) raises concerns. The controlled tier uses "generated state-transition matrices" with templates rendered under different regimes, which introduces circularity risk: regimes are defined to match hypotheses, and the templates are designed to instantiate those regimes. The fact that every regime produces exactly the predicted winner (3/3 model support across all slices) is suspiciously clean for empirical work. The frontier gaps are modest (0.5–3.4 accuracy points), and the paper does not report confidence intervals or statistical significance tests. The mediator statistics (Table 2) show consistent patterns, but the absolute magnitudes of effects like Early AUC (0.66–0.72) are moderate rather than overwhelming.
The audited budget framework is a genuine methodological contribution—the idea of normalizing comparisons across surface, latent, and compute interventions using a common cost ledger is valuable. But the specific calibration weights (α terms) are not clearly justified, and different weightings could shift results.
The naturalistic tier uses standard benchmarks (GSM8K-Platinum, HotpotQA, MATH, HumanEval+), which strengthens ecological validity. However, the regime assignments for these benchmarks appear to be made by the authors rather than empirically discovered, and the results again perfectly match predictions, which strains credulity.
Potential Impact
The paper's recommendations—treating latent-state dynamics as the default study object and using factorized, compute-audited experimental designs—could have substantial methodological influence if adopted. This would affect:
1. Interpretability research: Shifting focus from CoT faithfulness studies to latent-state probing and intervention
2. Benchmark design: Moving toward evaluations that separate the contributions of surface traces, latent representations, and compute budget
3. Safety and alignment: Reconsidering the reliability of CoT monitoring if reasoning is primarily latent
4. Architecture design: Motivating architectures that explicitly support latent reasoning (continuous latent spaces, looped transformers)
The six-arm experimental design template (Section 4.2) could become a useful reference for future empirical work, though its practical adoption may be limited by the complexity and cost of running all six conditions.
Timeliness & Relevance
This paper is highly timely. The field is experiencing rapid growth in reasoning-focused LLM research (o1, DeepSeek-R1, etc.), and there is genuine confusion about what CoT actually contributes versus additional compute. The unfaithfulness of CoT has been documented by multiple groups, creating an intellectual vacuum that this paper attempts to fill. The question of whether reasoning is "in" the text or "in" the model is increasingly pressing for safety and alignment work, where CoT monitoring is a proposed oversight mechanism.
Strengths
1. Clear conceptual contribution: The S/Z/B factorization is a genuinely useful organizing framework that could discipline future research
2. Nuanced claims: The paper explicitly positions H1 as a default working hypothesis rather than a universal verdict, acknowledging boundary regimes where H2 and H0 have local force
3. Comprehensive literature integration: The paper synthesizes a large body of recent work (probing, steering, filler tokens, latent reasoning, faithfulness studies) under a unified framework
4. Actionable recommendations: The two concrete recommendations and six-arm design template provide clear next steps
5. Falsifiability: Section 3.5 specifies conditions under which H1 would be weakened or overturned
Limitations
1. Single-author position paper: The empirical program, while structured, would benefit from independent replication and scrutiny. The consistent 3/3 model support across all regime-hypothesis pairings is concerning.
2. Circularity in regime design: The controlled tier generates tasks specifically to instantiate the predicted regimes, making the empirical validation partly tautological
3. Limited model diversity: Only three models tested (Qwen3-8B, Qwen3-32B, Llama-3.1-8B-Instruct), none at frontier scale
4. Vagueness of Z: Despite the footnote clarification, "task-relevant latent-state trajectories" remains somewhat underspecified. What counts as task-relevant versus arbitrary hidden activity is left as an empirical question, which risks making H1 unfalsifiable in practice
5. Missing direct competition: The paper acknowledges but does not resolve the fact that a clean, fully compute-matched head-to-head comparison of surface vs. latent interventions at scale has not been conducted
6. Overreach in framing: The title "LLM Reasoning Is Latent, Not the Chain of Thought" is stronger than the nuanced regime-dependent conclusion actually supports
7. No novel mechanistic evidence: The empirical program demonstrates regime-dependent dominance but does not provide new mechanistic insights into how latent reasoning actually works
Overall Assessment
This is a well-organized conceptual contribution that provides a useful framework for a confused area of research. The S/Z/B factorization and three-hypothesis structure are genuine contributions to clarity. However, the empirical validation is weaker than it appears—partly self-confirming in design, limited in model scale, and lacking statistical rigor. The paper's lasting impact will likely depend more on whether the community adopts its framing and experimental design recommendations than on its specific empirical results.
Generated Apr 20, 2026
Comparison History (61)
Paper 1 fundamentally challenges the dominant paradigm of Chain-of-Thought reasoning in LLMs, advocating a shift toward latent-state dynamics. This theoretical paradigm shift has profound implications across AI interpretability, evaluation, and fundamental model design, offering a broader and deeper scientific impact compared to Paper 2's more specialized, albeit important, framework for agent safety alignment.
Paper 2 has higher likely impact due to a large randomized real-world deployment (N=13,917) with clinician-blinded comparisons and substantial annotation, yielding actionable evidence about conversational diagnostic agents and interview strategies. It also creates a valuable corpus and links symptom dialogues to wearable physiology across hundreds of conditions, enabling downstream clinical and digital-health research. Paper 1 is timely and conceptually important for LLM interpretability/reasoning research, but as a position/framework paper its immediate real-world application and empirical leverage are narrower than Paper 2’s масштаб and translational relevance.
Paper 1 combines high novelty with strong real-world impact: a large-scale randomized deployment (N=13,917) in a consumer health app, clinician-blinded comparative evaluation, and a uniquely valuable dataset linking conversational symptom reports to wearable physiology across ~400 conditions. The methodological rigor and immediate clinical/consumer-health applications make it likely to influence both medical AI practice and future research. Paper 2 offers an important conceptual reframing of LLM reasoning, but as a position paper its impact depends on downstream adoption and lacks comparable empirical leverage or direct application.
Paper 1 addresses a foundational conceptual question about LLM reasoning—whether it occurs in latent states rather than surface CoT—that could reshape how the entire field studies, evaluates, and interprets reasoning in LLMs. Its framework (H0/H1/H2) provides a unifying lens for reorganizing diverse empirical findings and offers broad methodological recommendations. Paper 2, while technically sound and practically useful, is a more incremental contribution to inference-time decoding strategies. Paper 1's potential to redefine the default assumptions and experimental designs across reasoning research gives it broader and deeper scientific impact.
Paper 1 addresses a fundamental conceptual question about how the field understands and studies LLM reasoning, proposing a paradigm shift from surface chain-of-thought to latent-state trajectories. This reframing has broad implications for interpretability, benchmarking, and intervention design across the entire LLM reasoning research community. Paper 2, while technically sound, introduces an incremental inference-time decoding method (APPS) that improves accuracy-runtime tradeoffs. Paper 1's potential to reshape research methodology and evaluation practices across multiple subfields gives it greater breadth and lasting impact.
Paper 2 likely has higher scientific impact because it proposes a concrete, novel training framework (trajectory-modulated self-play with explicit transferability/evolution rewards) and reports cross-domain benchmark gains with ablations and human evaluation—suggesting methodological rigor and near-term applicability to building better reasoning models. Its approach is timely and could influence RL/self-play, reasoning transfer, and evaluation practices across NLP and code/math domains. Paper 1 is conceptually important and could reshape interpretation of CoT and reasoning study, but as a position paper its direct real-world impact and actionable methodology may be less immediate.
Paper 2 has higher estimated scientific impact because it reframes a central, field-wide question (what “reasoning” is in LLMs) with clear hypotheses and experimental design recommendations that affect interpretability, benchmarking, and intervention research across many subareas. Its breadth and timeliness are strong given ongoing debates about CoT faithfulness and evaluation. Paper 1 is a solid, application-relevant method with empirical gains, but its impact is narrower (training recipe via game self-play) and may be superseded by competing optimization techniques, whereas Paper 2’s conceptual framework can redirect multiple research agendas.
Paper 2 challenges the prevailing assumption that LLM reasoning occurs via explicit Chain-of-Thought, arguing instead for latent-state dynamics. Given the current intense focus on LLM reasoning and interpretability, this paradigm shift could fundamentally alter how researchers design benchmarks and interventions. While Paper 1 offers a valuable unifying causal framework for trustworthy AI, the immediate, field-wide relevance and provocative nature of Paper 2's hypothesis regarding foundational LLM mechanics give it a higher potential for broad, disruptive scientific impact.
Paper 2 likely has higher scientific impact due to immediate real-world utility: a large, schema-constrained, multimodal biomedical knowledge graph with open distribution and clear downstream applications (ML, retrieval, hypothesis generation). It demonstrates methodological effort in harmonization, provenance, and empirical validation via literature-evidence checks. Its breadth spans molecular to clinical domains and can support many labs and tools. Paper 1 is conceptually novel and timely for AI interpretability, but as a position/framework paper its impact depends on subsequent empirical adoption and may be narrower and less directly enabling than a widely usable biomedical infrastructure resource.
Paper 1 challenges a fundamental assumption about LLM reasoning (surface CoT vs. latent trajectories), which has profound scientific implications for interpretability, benchmarking, and our theoretical understanding of models. While Paper 2 offers a highly practical engineering framework for AI agents, Paper 1 presents a paradigm shift that will likely reshape the foundational scientific agenda and methodologies of LLM research.
Paper 1 likely has higher scientific impact: it proposes a unifying conceptual framework (latent-state trajectories vs surface CoT vs generic serial compute) that directly reshapes how the field interprets reasoning, faithfulness, interpretability, benchmarks, and intervention studies across many LLM tasks. Its disentangling recommendations and compute-audited exemplars target core methodological confounds, making it broadly applicable and timely for mechanistic interpretability and evaluation. Paper 2 is rigorous and socially relevant, but its impact is narrower (persuasion in societal issues) and more domain/context dependent, with findings that may generalize less across AI subfields.
Paper 2 addresses a fundamental question about the nature of LLM reasoning that has broad implications across the entire field of AI. By challenging the dominant chain-of-thought paradigm and proposing that reasoning occurs in latent states rather than surface traces, it could reshape how the community studies, evaluates, and improves LLM reasoning—affecting interpretability, benchmarking, and alignment research. Paper 1, while achieving impressive results on web search tasks, represents an incremental engineering advance in a narrower application domain. Paper 2's conceptual reframing has the potential to influence a much wider range of research directions.
Paper 2 has higher likely impact: it introduces a concrete, novel training-free method (dynamic, evolutionary “memetic” retrieval) with strong empirical gains on large-scale, realistic tool ecosystems, directly addressing a timely bottleneck for agentic LLM applications (tool use across tens of thousands of APIs). Its methodological rigor is supported by benchmarks, ablations implied by model-capacity dependence, and measurable retrieval/execution metrics. Paper 1 is conceptually important and timely for interpretability/reasoning science, but as a position paper it is less actionable and may yield slower, less direct real-world uptake.
Paper 2 addresses a fundamental question about the nature of LLM reasoning that has broad implications across the entire field of AI—affecting interpretability, benchmarking, alignment, and inference-time methods. Its theoretical framework (H0/H1/H2) provides a unifying lens that could reshape how the community studies and evaluates reasoning. Paper 1, while technically solid with strong empirical results on tool retrieval, addresses a narrower problem (dynamic tool retrieval for agents). Paper 2's potential to redirect research paradigms across multiple subfields gives it significantly broader and deeper scientific impact.
Paper 2 fundamentally challenges the dominant Chain-of-Thought paradigm, proposing that LLM reasoning occurs in latent states rather than surface text. As a foundational position paper, it has the potential to redirect broad swathes of AI research, including interpretability, benchmarking, and architectural design. While Paper 1 offers excellent empirical and theoretical contributions for SLM deployment, Paper 2's conceptual shift addresses the core mechanistic understanding of models, likely yielding a broader, paradigm-level scientific impact across the entire field.
Paper 2 addresses a fundamental conceptual question about the nature of LLM reasoning that could reshape how the entire field studies, evaluates, and interprets reasoning in language models. Its impact spans interpretability, benchmarking, mechanistic understanding, and inference-time methods. Paper 1, while technically solid with practical speedups for multi-agent systems, addresses infrastructure optimization—a more incremental engineering contribution. Paper 2's reframing of reasoning as latent-state dynamics versus surface CoT has broader theoretical implications that could influence research directions across multiple subfields for years.
Paper 1 has higher potential impact because it reframes multi-agent AI safety as an interaction-topology problem, identifying concrete, topology-driven failure modes that are directly relevant to real deployments (agentic workflows, voting/judging, sequential deliberation) and to policy/regulation. This offers actionable evaluation targets (robustness across architectural/topological variants) and cuts across safety, fairness, multi-agent systems, and governance. Paper 2 is timely and conceptually clarifying for interpretability/reasoning research, but its primary applications are more methodological and may be narrower in immediate real-world consequence than topology-driven safety failures.
Paper 2 addresses a fundamental question about LLM reasoning that impacts the entire AI research community—whether reasoning is mediated by latent states or surface chain-of-thought. This reframing has broad implications for interpretability, benchmark design, and inference-time methods across all LLM applications. Paper 1, while technically strong with impressive results on system-level diagram recognition, targets a narrow domain (chip design/EDA). Paper 2's conceptual framework will likely influence how thousands of researchers study and evaluate reasoning, giving it substantially broader and more lasting scientific impact.
Paper 2 likely has higher impact: it delivers a large, machine-checked formalization with strong theorems (semantic transparency, expressivity preservation, decidability boundary) and clear, immediate applications to AI system governance and workflow architectures. Its methodological rigor is high (0 admitted lemmas, extensive proof corpus), and results can influence both formal methods and AI safety/governance practices. Paper 1 is timely and conceptually useful for framing LLM reasoning research, but as a position paper its contributions are less definitive and may yield less immediate cross-domain uptake than a verified formal framework.
Paper 2 likely has higher impact due to a concrete, implementable method (T^2PO) addressing a widely felt bottleneck—instability/collapse in multi-turn agentic RL—validated across multiple benchmark environments with reported stability and efficiency gains and released code, increasing real-world uptake. Its methodological contribution (uncertainty-guided token/turn exploration control) is actionable and timely for agent training. Paper 1 is conceptually novel and could reshape how reasoning is studied, but as a position/framework paper its immediate empirical leverage and downstream applications are less direct.