acceptodds
Withdrawn from ICLR 2027

Phase Geometry of Reasoning: Predicting and Controlling Evidence Integration in Large Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models often produce different probabilistic judgments when equivalent evidence is reordered, reframed, or combined, yet the latent thought dynamics responsible for these changes remain poorly understood. We operationalize a thought as an atomic, belief-changing state transition induced by processing one evidence item, and represent reasoning as a trajectory through a latent state space. We hypothesize that these trajectories follow complex-valued, noncommutative dynamics in which phase-dependent interference determines how successive thoughts combine. To test this hypothesis, we introduce Complex Probabilistic Interventional Reasoning (CPIR), a controlled framework that traces model judgments across three-step reasoning sequences with known priors and likelihood ratios. CPIR intervenes on individual thought updates through evidence masking, polarity reversal, strength modification, order permutation, paraphrasing, and compositionally held-out combinations of these operations. We compare a complex-valued state-space model with normative Bayesian inference, logistic interaction models, nonlinear real-valued recurrent models, and parameter-matched controls. On five frozen synthetic environments with planted complex-coherent dynamics, CPIR identifies the correct compositional family and predicts held-out domains, paraphrases, and longer sequences, reducing joint-holdout logit RMSE from 0.072 to 0.023 relative to the strongest real-valued baseline. When transferred to Qwen3-14B and DeepSeek-V3.2 judgments, the same prespecified procedure discovers target-specific real and heterogeneous predictive regimes, demonstrating selective model assignment across targets. Extending CPIR to sequential LLM evaluation, repeatability and competence controls reveal that the observed cross-order reversal rate (11.1%) is matched by same-order variation (10.0%) for the competence-qualified judge, motivating checklist elicitation that reduces reversals to 0.7%. At equal inference cost, averaging opposite criterion orders improves Brier score from 0.156 to 0.146 over averaging two same-order evaluations, with the gain concentrated in probabilistic calibration. Together, these results show that CPIR can recover structured thought composition, discriminate among compositional regimes, and improve the stability and calibration of sequential LLM evaluation.

Primary area:
interpretability and explainable AI
Keywords:
Large language models, complex-valued representations, probabilistic reasoning, mechanistic interpretability, causal intervention, activation steering, evidence integration, contextual interference, order effects, compositional generalization, reasoning faithfulness, quantum cognition

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_pX9EOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Complex Probabilistic Interventional Reasoning, an intervention-based framework for selecting among real, complex, and heterogeneous dynamics of LLM belief updates, and applies it to LLM-as-a-judge evaluation.

Critical strengths

The manuscript is fluently written and adopts the vocabulary of rigorous methodology throughout (frozen splits, pre-specified gates, matched controls).

Critical weaknesses
  1. The title centers on phase geometry, but the complex model is not selected on either real LLM (Table 4); the only support comes from synthetic environments where complex dynamics were planted.
  2. The paraphrase and domain held-out results for both LLMs are described in Sec. 5 but never reported.
  3. The target qcq_c is defined as a frequency of binary samples (Eq. 11) but described as a reported probability (Sec. 5); samples per condition and decoding settings are not given.
  4. The headline 11.1% → 0.7% result is based on 144 candidate pairs (24 scenarios × 6 pairs), i.e., roughly 16 → 1 pair.
  5. The main text contains one two-row table; all substantive results are in the appendix.
  6. Several defined metrics (IEE/OEE, NLL, the geometry-consistency residual) and listed baselines (normative Bayes, logistic, scalar real SSM, nonlinear real RNN) never appear in any result, and the gate counts (7/13/6) are never enumerated.
  7. Multiple cross-references are incorrect: Appendix A does not contain the “complete parameter grid”, Appendix F does not report gate outcomes, Appendix G does not contain the estimators, and the “†” in Table 6 is undefined.
  8. Several figures have overlapping labels or legends (Figs. 2c, 3a, 3b, 4a, 4b), and Figs. 2a–c and 4a duplicate Tables 2–5.
  9. On real LLMs, all fitted families have logit RMSE of 1.17–2.43, and no prior-only or normative Bayes baseline is reported, so it is unclear whether any family is better than learning nothing.
Questions for authors

Given the number of incorrect cross-references, unreported metrics, and missing results, could the authors clarify how the manuscript was checked before submission?

Reviewer_ZQfcOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces CPIR (Complex Probabilistic Interventional Reasoning), a framework for studying how large language models (LLMs) integrate evidence probabilistically. A "thought" is operationally defined as a belief-changing state transition induced by processing one controlled evidence item, and reasoning is modeled as a trajectory through a latent state space. The central hypothesis is that these trajectories may follow complex-valued, noncommutative dynamics: each thought type is a 2×2 complex operator, and phase-dependent interference determines how successive thoughts combine, producing order effects. CPIR collects LLM probability judgments over controlled evidence sequences with known priors and likelihood ratios and applies interventions (evidence masking, polarity reversal, strength modification, order permutation, paraphrasing, and compositional combinations). A frozen, prespecified selection procedure compares the complex operator model against normative Bayes, logistic interaction models, real state-space models (2D and 3D), a generic GRU, a phase ablation (B ≡ 0), and a diagonal/no-coherence ablation, under compositional holdouts (unseen lengths, unseen wording, unseen domains, and their joint combination). Validation has three parts. (1) Synthetic environments with planted diagonal/real/complex-coherent dynamics: CPIR recovers the correct family; on five frozen complex-coherent environments the complex model attains joint-holdout logit RMSE 0.0231 vs. 0.0716 for the strongest real baseline and 0.0796 for the GRU, and it is sample-efficient (crossing the 0.05 threshold with 40 sequences). (2) Application to Qwen3-14B and DeepSeek-V3.2 judgments with the same frozen procedure: the procedure selects phase-ablated real dynamics for DeepSeek on both holdouts, and a heterogeneous regime for Qwen (GRU best on length extrapolation, Real-3D best on the joint shift); no complex-coherent dynamics are found in either real model. (3) An LLM-evaluation application: on a 24-scenario, 4-domain benchmark with three judges, cross-order pairwise reversals (11.1%) are comparable to same-order rerun instability (10.0%) for the competence-qualified judge (DeepSeek-V3.2); checklist elicitation reduces reversals to 0.7%, and, at equal inference cost, averaging opposite criterion orders improves Brier score from 0.1563 to 0.1457 (difference 0.0106, CI [0.0045, 0.0171]), with the gain localized to probabilistic calibration.

Critical strengths
  1. The paper pre-registers selection gates, evaluates on compositional holdouts (unseen combinations of length, wording, domain), matches input information across all families, reports scenario-clustered bootstrap intervals in the evaluation section, and explicitly reports its negative results (real dynamics for DeepSeek, heterogeneous for Qwen; the order-marginalization analysis is honestly labeled hypothesis-generating). This is exemplary practice for interpretability research, makes the framework reusable, and I verified that the reported numbers are internally consistent (e.g., 67.8% = 1 − 0.0231/0.0716; factor 14.4 = 0.3323/0.0231; 93.7% = 1 − 0.007/0.111).
  2. A practical, statistically careful evaluation result. Same-order rerun controls cleanly establish that cross-order reversals (11.1%) sit within rerun noise (10.0%; difference 1.1pp, CI [−1.8, 4.1]) for the qualified judge. Checklist elicitation reduces reversals to 0.7% while lowering criterion error, and equal-budget opposite-order averaging improves Brier score with a bootstrap interval excluding zero, with the benefit correctly localized to calibration (regret CI includes zero). This is a credible, useful contribution to LLM-as-judge methodology.
Critical weaknesses
  1. The central hypothesis is not supported on any real LLM, yet the title and abstract overpromise. The headline claim—complex-valued, noncommutative "phase geometry of reasoning"—is validated only on synthetic data with planted complex-coherent dynamics (Section 4). On real models (Section 5), DeepSeek-V3.2 selects phase-ablated real dynamics and Qwen3-14B a heterogeneous regime; neither exhibits complex dynamics, and the best joint-holdout logit RMSEs are ≈ 2.2 and ≈ 1.9 (Table 4)—all families fit the LLM judgments very poorly in absolute terms, so "family selection" among poor fits is itself in need of justification. The abstract's framing ("the latent thought dynamics responsible for these changes remain poorly understood" … "we hypothesize that these trajectories may follow complex-valued, noncommutative dynamics") and the title "Phase Geometry of Reasoning" exceed what the experiments establish. As the paper's own data falsify the headline hypothesis on real models, the delivered contribution is (a) a model-selection framework and (b) an evaluation-stabilization protocol—both worthwhile, but the paper should be reframed accordingly. Note also that Section 6 does not depend on complex geometry at all; the two halves of the paper are only loosely coupled.
  2. The synthetic complex-vs-real comparison is confounded by state dimension. The complex model has a 2D complex state (4 real components; dynamics on S³ with 8 real parameters), and its "exact realification" (Eq. 10) is a 4×4 real block. Yet the strongest real control is 3D (S²) and the "parameter-matched" real model is 2D (S¹). Under the normalized recurrence, lower-dimensional real states cannot represent the same projective dynamics regardless of phase, so no tested baseline matches the complex model in state capacity. No 4D real baseline (either unconstrained, or with the same 8-parameter near-identity structure) is ever fitted. It is therefore unclear whether the large gains on complex truth (0.0056 vs. 0.6127 / 0.3515 in Table 2; 0.0231 vs. 0.3368 / 0.0716 in Table 3) reflect complex phase structure or simply larger state capacity. Tellingly, in the one comparison that holds state dimension fixed on real data (phase ablation vs. complex, both S³ with the same near-identity operator family), the complex model loses for DeepSeek. The authors must clarify what "parameter-matched" means (parameter count, state dimension, initialization, near-identity constraint) and add a 4D real baseline to make the phase claim identifiable.
Questions for authors

None

Reviewer_fjUBOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes CPIR, a framework for modeling probabilistic evidence integration through latent state transitions. It compares complex-valued operators with real-valued state-space models, recurrent models, and ablations under compositional holdouts. Synthetic experiments show recovery of planted diagonal, real, and complex dynamics, with the complex model achieving substantially lower extrapolation error when the generator is complex-coherent. On Qwen3-14B and DeepSeek-V3.2 judgments, however, complex-specific support is absent; the paper instead assigns heterogeneous and real-valued regimes. A separate LLM-judge study compares cross-order instability with same-order variability, demonstrates reduced reversals through checklist elicitation, and reports a secondary calibration improvement from opposite-order averaging.

Critical strengths

Useful repeatability and competence controls. Comparing cross-order reversals with same-order reruns prevents ordinary sampling variability from being automatically interpreted as structured order dependence. For the qualified DeepSeek judge, the difference is only 1.1 percentage points with an interval spanning zero (Appendix G, Table 6). This is a meaningful diagnostic contribution.

Explicit acknowledgment of real-coordinate equivalence and negative evidence. The paper correctly recognizes that complex operators have an exact real representation and reports that neither real LLM supports the complex-specific gates. These qualifications prevent an unwarranted claim of an intrinsically complex or quantum substrate.

Critical weaknesses

The empirical evidence does not substantiate the central discovery claim on LLMs. The strong complex-model results come from environments explicitly generated by complex-coherent operators (§4). This validates recovery under correct specification, not the relevance of phase geometry to naturally trained models. On real LLMs, the best joint-holdout logit RMSE remains approximately 2, and no noise-floor analysis establishes how adequately any family explains the judgments (§5; Appendix F, Table 4). Moreover, the real-family assignment relies on relatively small point-estimate differences without reported uncertainty for the decisive family contrasts. A best-performing model in a restricted, poorly fitting candidate set is not sufficient evidence of a discovered dynamical regime. This gap undermines the paper’s main scientific contribution.

The practical improvements are not demonstrated consequences of the proposed geometry. Checklist elicitation and opposite-order averaging do not use the learned state, phase, operators, or predicted intervention effects (§6). No comparison establishes that CPIR’s fitted dynamics select a better intervention than a simpler protocol based on repeatability checks and rubric decomposition. For the qualified judge, global cross-order reversals are already statistically indistinguishable from same-order variation, so reducing reversals does not itself demonstrate control of a structured order effect. The calibration improvement is useful but explicitly secondary and lacks an independent confirmatory replication. Consequently, these experiments do not establish the claimed connection between identifying thought geometry and controlling evaluation behavior.

Questions for authors

Clarification: What absolute adequacy and uncertainty criteria justify assigning DeepSeek a real dynamical regime rather than reporting that none of the tested families adequately explains its behavior? Please provide paired uncertainty for the decisive model comparisons and an estimate of the elicitation noise floor.

Clarification: Which checklist or averaging decision depends on the fitted CPIR dynamics? Is there existing evidence that the geometric modeling contributes beyond competence checks, same-order reruns, and direct prompt comparisons?

Clarification: Can the authors provide the complete operator-generator architectures, trainable parameter counts, optimization and regularization settings, probability-collection protocol, and fitted unconstrained four-dimensional real control? Exact realification verifies algebraic equivalence, but does not replace a competitive independently fitted real baseline for assessing the structured inductive bias.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.