acceptodds
Under review as a conference paper at ICLR 2027

Parallelizable Probabilistic Latent Reasoning for Diverse Reasoning Trajectories

Abstract

Language models relying on verbalized Chain-of-Thought (CoT) are constrained to token-by-token generation in the language space, where many tokens serve textual coherence rather than core reasoning. This entangles genuine differences in logical trajectories with superficial variations in verbalization. Moving reasoning into a compact, continuous latent space offers a promising alternative for efficiently representing and exploring diverse reasoning trajectories. However, current latent models remain largely deterministic, capturing only a single reasoning trajectory rather than modeling a distribution over multiple valid alternatives. While stochastic approaches parameterize distributions over latent reasoning steps, they frequently suffer from unstable optimization and limited empirical performance. Furthermore, their reliance on autoregressive latent prediction during training precludes effective parallelization, leading to prohibitively long training times. To overcome these bottlenecks, we propose Probabilistic Latent Reasoning Models (PLRM), a parallelizable variational framework for probabilistic latent reasoning. PLRM adopts a JEPA-style objective in which the posterior infers latent distributions in parallel from observed reasoning trajectories, while the prior learns to predict the subsequent latent distribution via teacher forcing. To model heterogeneous reasoning processes, we introduce a discrete strategy variable together with a mixture-based prior, allowing the model to preserve the Gaussian-mixture structure induced by multiple distinct reasoning trajectories rather than collapsing them into a single Gaussian. Across diverse benchmarks, PLRM achieves performance competitive with deterministic latent reasoning models on single-trajectory datasets while enabling efficient training and consistently outperforming competing stochastic baselines on multi-trajectory datasets.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Autoregressive Language Model, Language Model, Continuous Chain-of-Thought, Joint-Embedding Predictive Architecture

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_y9cWOpenReview ↗
Rating
2(−1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Probabilistic Latent Reasoning Models (PLRM), a parallelizable variational framework designed to enable probabilistic latent reasoning. PLRM employs a Joint-Embedding Predictive Architecture (JEPA) to infer target latent distributions from CoT traces in parallel. PLRM also employs a discrete categorical strategy variable together with a mixture prior to capture heterogeneous strategies to enable the modeling of multi-trajectory distributions. Computer simulations show improvements in terms of mathematical and common sense reasoning tasks relative to stochastic latent reasoning models for GSM8K-Aug-NL and CommonsenseQA dataset.

Critical strengths

The major contributions of this work are:

  1. introduction of PLRM that supports variational latent reasoning and allows the transfer of reasoning from the language space to the latent space. Three important features of conventional supervised fine-tuning are preserved by PLRM in these regards: modeling of distributions in the latent space; dense supervision; and teacher-forced next-latent distribution prediction.
  2. PLRM forces the latent representation to preserve the essential reasoning content while filtering out incidental details of how that reasoning is expressed verbally.
  3. PRLM models heterogeneous strategies through a mixture-based prior.
Critical weaknesses

The performance improvement is not clear on purely equation-dominant applications where the math equations dominate the reasoning--on on Llama-3.2-1B, PLRM presents 50.00% accuracy on GSM8K, less than the deterministic methods: SIM-CoT (55.19%) and CODI (54.44%) (Table 2). Not clear if PLRM features exhibited on very few datasets can be generalized.

Training of the posterior target encoder may be challenging since detailed knowledge of CoT traces is needed to estimate the target latent distributions. Not clear in what realistic circumstances consistent estimation of the target distributions is possible.

It is not clear how the posterior token allocation (number of CoT tokens for each latent step) and number of mixture models can be set up automatically in a consistent manner without making use of an adhoc and computationally demanding procedure.

To overcome the joint posterior-prior training from collapsing into trivial solutions, PLRM employs many loss regularizers. The unaddressed problem is how to adjust properly the contribution of each regularizer?

It remains an open question whether the model can scale effectively to larger architectures or handle open-ended, real-world tasks with multiple possible solution paths, such as creative writing.

Questions for authors

The writing style of the paper could be significantly improved to enable even readers unfamiliar with the topic to understand the basic contributions of this paper. The paper is written in a style friendly only to the specialists not to the novice readers. In the Abstract of the paper the abbreviation "JEPA-styel" is used without spelling it out. Please spell out all abbreviations before using them and cite the necessary references to inform the readers about the prior results used in paper. On page 1, 5 lines from bottom of the page, I recommend correcting this typo "This is not only introduces ..." Also, on page 6, 6 lines above equation (9), please correct this typo "the prioir constructs".

On page 4, 2 lines from the top, it is not clear "the model can continue decoding directly from the current KV cache without recomputing the context representations". Please explain.

On page 5, in the paragraph that starts with "Preventing Representation Collapse", please can you justify why this methodology is sufficient and helps in this case "we apply VISReg (Wu et al., 2026) to the posterior means μq to stabilize joint posterior–prior training. We regularize μq rather than sampled latent z, since Gaussian sampling noise could satisfy the regularization even when the posterior means collapse."

It would be useful to provide rigorous analytical justifications for the claims made and to explain the achieved performance improvements formally.

2 comments

Thank you for reading our paper closely and sharing your feedback. We particularly appreciate your recognition that PLRM can retain essential reasoning content while filtering out incidental details of its verbal expression.

W1. Performance on equation-heavy tasks

On equation-dominant GSM8K with Llama-3.2-1B, PLRM scores 50.00%, compared with 54.44% for CODI and 55.19% for SIM-CoT. We separately evaluate transfer from GSM8K-Aug-NL to SVAMP, GSM-Hard, and MultiArith. Those results characterize generalization across the evaluated arithmetic benchmarks.

W2. Estimating posterior targets

In our supervised setting, ground-truth CoT traces are available during training. The posterior encodes information from these traces into target latent distributions, which supervise the prior to predict corresponding latent distributions from the input question alone. For real-world settings where ground-truth CoTs are unavailable, we believe PLRM could potentially be combined with self-bootstrapping approaches such as STaR (Self-Taught Reasoner), which generate and refine reasoning trajectories for training. We acknowledge that extending PLRM to such settings remains an open direction and will clarify this practical limitation and potential extension in the revised manuscript.

W3. Choosing token allocation and component count

We determine posterior token allocation (PTL) using a simple, consistent rule: dividing the 95th-percentile reasoning-trace length of each dataset by the fixed number of latent steps, K=6K=6. This requires only dataset-level token-length statistics, without computationally expensive hyperparameter searches or additional model training. For standard benchmarks with one annotated reasoning trace per example, we use M=1M=1. For HiddenBranch, Table 5 evaluates M=1M=1 through M=4M=4, with M=2M=2 achieving the highest average TV Score.

W4. Choosing regularizer weights

To mitigate posterior representation collapse, PLRM applies VISReg to the posterior means. Importantly, we use identical VISReg hyperparameters across all six dataset–backbone configurations in Table 6, without dataset-specific tuning:

HyperparameterShared Value
VISReg weight (λreg\lambda_{\mathrm{reg}})0.9
VISReg projections512

These shared settings show that our reported results do not rely on benchmark-specific tuning of VISReg hyperparameters.

W5. Larger models and open-ended tasks

Our current evaluation covers seven benchmarks for mathematical and commonsense reasoning, including out-of-distribution transfer and a controlled test of heterogeneous reasoning strategies on HiddenBranch. Open-ended generation tasks such as creative writing remain to be evaluated.

Training and evaluating PLRM with larger backbones and on more diverse benchmarks within the rebuttal window is challenging given the time and GPU resources available. We will continue these experiments through the end of the rebuttal period and add any results we obtain by the deadline.

Q1. Accessibility, acronyms, and typos

We have added an overview of PLRM as Figure 1 and clarified the explanation of the method in the main text. We have also spelled out abbreviations at first use, including Joint-Embedding Predictive Architecture (JEPA) and corrected the grammatical and typographical errors you identified.

Q2. KV cache reuse

During inference, PLRM first generates the latent trajectory autoregressively from the input question xx. As each projected latent embedding is processed by the model, its attention keys and values are stored in the KV-cache alongside those of the input question. Thus, when generating the answer conditioned on x,zx,z, the required context is already represented in the existing KV-cache. Answer decoding simply continues from this cache, without requiring a separate forward pass to reprocess the question and latent trajectory.

Q3. Why regularize the posterior mean?

In PLRM, the posterior provides latent targets derived from observed reasoning traces, which the prior learns to predict. If the posterior means collapse to nearly identical representations across examples, these targets may no longer preserve meaningful differences between reasoning trajectories, weakening the supervision provided to the prior.

Importantly, since z=μq+σqϵz=\mu_q+\sigma_q\epsilon, sampled latents can still exhibit substantial variance due to Gaussian noise even when their means have collapsed. However, such noise-induced variation does not necessarily encode meaningful differences between reasoning traces. Regularizing the sampled latents alone could therefore satisfy a variance constraint without addressing the underlying collapse.

We apply VISReg directly to the posterior means to encourage non-degenerate variation across examples, while the answer-prediction and CoT-reconstruction objectives encourage this variation to retain task-relevant information. Although this does not mathematically guarantee the absence of collapse, it directly addresses the failure mode described above. We will clarify this motivation in the revised manuscript.

Q4. Analytical justification

Proposition 1 shows that KL alignment during joint posterior–prior training penalizes information in zz about the particular verbalized CoT that cannot be predicted from the question xx. Together with the answer prediction objective, this penalty encourages zz to preserve the reasoning needed to solve the problem while discarding incidental verbal details.

On GSM8K-NL with GPT-2, PLRM reaches 36.30% accuracy with 11.28 tokens on average, compared with 30.93% and 67.00 tokens for CoT-SFT. PLRM also outperforms SIM-CoT, which reconstructs CoT tokens at each latent step, on both GPT-2 (36.30% vs. 17.20%) and Llama-3.2-1B (42.08% vs. 31.69%).

To test what the latents retain, we trained separate decoders on the same Llama-3.2-1B PLRM posterior latents, encoded from the ground-truth GSM8K-NL CoT. The table reports cross-entropy (CE), number recall, and exact match with the ground-truth CoT in that order.

Natural languageEquations
0.495 / 84.5% / 21.0%0.318 / 79.5% / 48.7%

Although the posterior encoded the ground-truth natural-language CoT from GSM8K-NL, the decoder recovered the full sequence of corresponding GSM8K equations exactly on 48.7% of problems, compared with 21.0% for the original CoT. This suggests that the posterior latents retain the mathematical steps of the reasoning more readily than their exact wording.

Reviewer_kiD3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PLRM, a probabilistic latent reasoning framework that jointly learns a CoT-conditioned variational posterior and a problem-conditioned autoregressive prior. The posterior supplies latent target distributions, enabling parallel teacher-forced training of the prior. Training combines CoT reconstruction, answer prediction, distribution alignment, and representation regularization. A categorical strategy variable further enables a mixture prior to model heterogeneous reasoning trajectories. The authors evaluate PLRM using GPT-2 and Llama-3.2-1B on arithmetic and commonsense benchmarks, alongside HiddenBranch, a ProofWriter-derived benchmark with multiple valid reasoning–answer pairs. Results show improvements over the evaluated stochastic baselines, mixed performance relative to deterministic methods, favorable training-time scaling compared with sequential-rollout approaches, and improved valid-branch coverage from the mixture prior on HiddenBranch.

Critical strengths
  1. The paper presents a well-motivated framework for jointly learning latent reasoning targets and a problem-conditioned predictive prior. The posterior provides CoT-conditioned target distributions, while the prior learns to predict these distributions through parallel teacher-forced training. This gives a concrete implementation of the paper’s JEPA-style motivation: latent targets can adapt jointly to answer prediction and predictability from the problem. I consider this the paper’s main strength because it connects representation learning and training efficiency within a coherent probabilistic framework.
Critical weaknesses
  1. The experiments do not convincingly establish a favorable accuracy–efficiency trade-off. On Llama-3.2-1B, PLRM achieves 50.00% on GSM8K, compared with 54.44% for CODI and 58.91% for CoT-SFT. Although PLRM performs well on CSQA, its improvements are not consistent across the evaluated tasks. Figure 1 demonstrates an advantage in training-time scaling over methods requiring sequential latent rollout, but reports time per iteration rather than time to reach comparable accuracy; CoLaR also supports parallel training and is faster in this comparison. Together with the evaluation being limited to GPT-2 and a 1B model on arithmetic and commonsense benchmarks, these results leave unclear when PLRM offers a compelling practical advantage without scaling.
  2. The claimed result of stochastic latent reasoning is not directly validated by the evaluation. The paper does not report pass@k or success rates under matched inference budgets, which would help establish whether sampling alternative latent trajectories improves problem-solving performance. All standard benchmarks use M=1, so the benefit of the strategy-conditioned mixture is evaluated only on HiddenBranch. This controlled benchmark usefully measures coverage of multiple valid answers, but does not establish the benefit of exploring different reasoning paths toward the same answer. Moreover, increasing M from 1 to 2 improves TV Score from 0.6252 to 0.7097 while reducing Hit Rate from 93.70% to 92.05%. This demonstrates a coverage–accuracy trade-off, but leaves the practical value of the additional diversity unresolved.
Questions for authors
  1. Clarification — baseline reproduction: Table 2 reports 67.32% CSQA accuracy for CODI, whereas Table 8 lists a previously reported result of 74.0%, close to PLRM’s 74.28%. What differences in training, checkpoint selection, or evaluation explain this discrepancy? This clarification would materially affect my assessment of the CSQA improvement.

  2. Clarification — efficiency measurement: Table 6 provides PLRM’s training hyperparameters, and Figure 1 specifies the token-to-step convention for CoT-SFT. Could the authors clarify the hardware, actual batch sizes, and sequence lengths used specifically for the timing comparison, and whether the Table 6 settings apply? Does the reported iteration time include all forward and backward computations, including posterior computation, CoT reconstruction, and answer prediction for PLRM? Please clarify how comparable training workloads were ensured across methods.

  3. Clarification — stochastic evaluation: Appendix B.2 defines TV Score using R samples per input and describes Hit Rate as deterministic/greedy decoding. Could the authors specify the value of R, the latent and text decoding settings used for TV Score, and how the strategy variable and continuous latents are handled when computing Hit Rate? In particular, are strategies selected by argmax and latents set to their prior means for the deterministic evaluation? These details are important for interpreting the coverage results and their comparison across methods.

  4. Additional experimental evidence — benefit of sampling: Can the authors provide pass@k or success vs inference-cost results comparing PLRM with CoT-SFT and the stochastic latent baselines under matched inference budgets? If prior-mean rollouts are not already used for the reported deterministic evaluation, please also compare them with sampled latent rollouts to help quantify the benefit of stochastic sampling at inference.

2 comments

Thank you for your detailed review and careful assessment of our method. We especially appreciate your observation that jointly learning latent targets and the problem-conditioned prior connects representation learning with efficient parallel training.

W1. Accuracy and efficiency trade-off

We thank the reviewer for this important observation. We acknowledge that PLRM does not consistently outperform deterministic methods, particularly on equation-dominant GSM8K. However, PLRM achieves competitive accuracy while consistently outperforming the evaluated stochastic latent reasoning baselines. Importantly, PLRM aims to balance effective stochastic latent modeling with parallelizable training. Deterministic latent methods requiring sequential latent rollout incur increasing training costs as reasoning steps grow, whereas PLRM supports parallel teacher-forced training. Although CoLaR also supports parallel training and has lower per-iteration cost, PLRM achieves substantially higher accuracy across the evaluated stochastic baselines.

Thus, our intended contribution is not uniformly superior accuracy or the lowest training cost, but a practical balance between stochastic reasoning capability, predictive performance, and training efficiency. We will clarify this positioning and the scope of our efficiency claims in the revised manuscript.

W2, Q4. Benefits of stochastic reasoning under matched inference budgets

On GPT-2 GSM8K-NL, PLRM achieves higher pass@kk than CoT-SFT and CoLaR at all four matched token budgets. We set the budgets using 10, 30, 60, and 100 PLRM rollouts. For each baseline, we use the largest number of rollouts kk that fits each budget.

Method (tokens/rollout)PLRM pass@10 (112 tokens)PLRM pass@30 (338 tokens)PLRM pass@60 (675 tokens)PLRM pass@100 (1,125 tokens)
PLRM (11.3)50.1% (10)57.3% (30)61.5% (60)64.2% (100)
CoT-SFT (68.6)31.1%43.9%53.9%61.6%
CoLaR (34.2)25.3%29.9%33.4%36.1%

PLRM’s pass@kk advantage over CoT-SFT is largest at the smallest token budget (19.0 percentage points at 112 tokens).

On HiddenBranch, increasing MM from 1 to 2 raises pass@3 from 96.72% to 97.20% and pass@5 from 97.84% to 98.51%. Pass@1 falls from 90.66% to 89.97%, and greedy Hit Rate falls from 93.70% to 92.05%. The gains at higher kk measure how often a valid assumption is recovered at a fixed sample count. We used prior-mean rollouts for all PLRM results in Tables 1 and 2 and for the deterministic Hit Rate on HiddenBranch.

Q1. CODI performance on CSQA

To accommodate GPU utilization, we adjusted the effective batch sizes for Llama-1B and applied corresponding linear learning-rate scaling, while keeping the GPT-2 settings unchanged. To further investigate the CSQA discrepancy, we additionally reproduced CODI using its original settings during the rebuttal period:

Backbone / DatasetEffective Batch (Original → Ours)Learning Rate (Original → Ours)Accuracy (%)
GPT-2UnchangedUnchanged—
Llama-1B / GSM8K128 → 1128e-4 → 7e-454.44
Llama-1B / CSQA64 → 568e-4 → 7e-467.32
Llama-1B / CSQA (original-setting rerun)64 → 648e-4 → 8e-469.78

Restoring the original settings improved CSQA accuracy from 67.32% to 69.78%, although it remained below the published 74.0%, which is comparable to PLRM's 74.28%.

For further details on our baseline reproduction procedure and comparisons with published results, please refer to our response to Reviewer MG2v (W2).

Q2. Timing setup and comparable workloads

All methods were timed on a single NVIDIA B200 GPU using a controlled single-batch training protocol. For each backbone and reasoning budget K∈{2,6,10,14,18}K\in\{2,6,10,14,18\}, we constructed a batch of 64 repetitions of the same fixed question–answer pair, with synthetic reasoning inputs adapted to each method's format. The question–answer pair, batch size, and reasoning budget were held consistent across methods. Both the physical and effective batch sizes were 64, with no gradient accumulation. This protocol was separate from the main benchmark training settings in Table 6. For each configuration, we performed six training updates on the same batch, discarded the first as warm-up, and averaged the remaining five. CUDA was synchronized at both timing boundaries. Iteration time includes all required forward and backward passes and the optimizer update, including PLRM's posterior computation, CoT reconstruction, and answer prediction.

Q3. TV Score and Hit Rate decoding settings

We thank the reviewer for pointing out the missing evaluation details. We computed TV Score using 10 stochastic samples per input across 6,128 HiddenBranch test inputs.

For stochastic evaluation, CoT-SFT uses token sampling at temperature 1.0, while CoLaR and ReGuLaR additionally sample Gaussian latents and latent termination decisions at the same token-sampling temperature. PLRM samples strategies and continuous latents from its learned prior without additional temperature scaling, while greedily decoding answer tokens. The sampled strategy remains fixed throughout each latent trajectory.

For deterministic Hit Rate, PLRM selects the most probable strategy from its learned input-conditioned strategy distribution (argmax) and keeps it fixed throughout the trajectory. Each intermediate latent state is set to the corresponding conditional prior mean, rather than being sampled, and answer tokens are decoded greedily.

We have added these evaluation settings to Appendix B.2 of the revised manuscript.

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

This paper proposes Probabilistic Latent Reasoning Models (PLRM), a variational framework that aims to carry three important properties of CoT supervised fine-tuning (stochasticity, dense supervision, and parallel training) to continuous latent reasoning. A posterior qϕ(z∣x,τ)q_\phi(z \mid x, \tau) infers a sequence of Gaussian latents from the verbalized CoT in a single forward pass. A problem-conditioned prior pθ(z∣x)p_\theta(z \mid x) learns to predict them via teacher forcing. The authors frame this setup as a JEPA-style objective. To handle problems with multiple valid reasoning paths, the paper further introduces a discrete strategy variable. Experiments show improvements over stochastic latent baselines.

Critical strengths
  • The problem is clearly formulated. The paper explicitly identifies the three properties that are lost when moving from CoT reasoning to latent reasoning, and it addresses all.

  • The teacher-forced prior gives a practical training-efficiency advantage over baselines that require sequential latent rollout (Fig. 1).

  • The strategy-conditioned mixture prior is a reasonable way to address the multimodal problem.

Critical weaknesses
  • The JEPA discussion in Section 3.2 overlooks the reconstruction term. Based on DKL(qϕ(z∣x,τ) ∥ pθ(z∣x))D_{\mathrm{KL}}(q_\phi(z \mid x,\tau)\,\|\,p_\theta(z \mid x)), the paper claims that the latent representation retains core reasoning information at an abstract level while discarding verbal details. However, term 1 in line 196, the reconstruct term pθ(τ∣z)p_\theta(\tau \mid z), trains the model to reconstruct the entire reasoning trace from the latent zz. This reconstruction-based loss encourages the latent to maintain reasoning details, which contradicts the above claim and is fundamentally different from the core objective of JEPA, i.e., “predictive representations over reconstructing every detail of the input.” Therefore, the JEPA argument does not account for the full objective and rests on an incorrect justification.

  • In line 226, VISReg is applied based on this JEPA framing, but there is no ablation study with and without VISReg.

  • The baseline results in Tables 1 and 2 differ from the numbers reported in the original papers. Table 8 in the appendix provides the reported results, but the main paper does not mention whether the main-table values were reproduced, which may cause confusion. For example, when compared against the reported values, PLRM underperforms the deterministic baseline CODI in many settings.

  • HiddenBranch appears to be an overly favorable setting for the strategy-conditioned approach. As shown in Appendix B.2, each CoT explicitly states the strategy at the beginning (e.g., “B1: Assume Anne is big.”). Since the method explicitly models this structure in the data, this seems sufficient to give it an advantage over the baselines, and the comparison may appear method-favored and unfair.

  • Regarding the strategy-conditioned prior, the main table includes math problems, which often admit diverse solutions. One could therefore naturally expect the CoT reasoning traces to be multimodal, yet there are no experiments increasing MM to 2–4 on these benchmarks.

  • The paper lacks a figure illustrating the overall method, which makes it difficult to follow how the posterior, prior, strategy variable, and losses interact, and how the model works.

Questions for authors
  • From JEPA perspective, how does performance change when the reconstruction term pθ(τ∣z)p_\theta(\tau \mid z) is removed or the KL weight is increased? Is there direct evidence that zz discards verbalization-specific information?

  • What happens without VISReg? Does the posterior actually collapse?

  • What are the results with M>1M>1 on Table 1 or Table 2? And if its possible, can you report sampling-based metrics such as pass@k to show the benefit of stochastic latent reasoning?

  • Were the baselines in Tables 1 and 2 re-trained? If so, what settings were used, and why are the numbers lower than originally reported?

4 comments

We appreciate the care you put into reviewing our work. Thank you in particular for highlighting how PLRM preserves stochasticity, dense supervision, and parallel training, and for noting the efficiency of the teacher-forced prior.

W1, Q1. JEPA framing, reconstruction, and verbalization

We tested whether PLRM’s latents preserve the mathematical content of a reasoning trace more readily than its exact wording, even with the CoT reconstruction loss.

Both Llama-3.2-1B checkpoints were trained on GSM8K-NL. Each posterior encoded the ground-truth natural-language CoT into latents. We trained decoders from scratch on the checkpoints’ latents. The decoding targets were the ground-truth CoTs from GSM8K-NL (natural language) and GSM8K (equations). The table reports cross-entropy (CE), number recall, and exact match with the ground-truth CoT in that order.

Latent inputNatural languageEquations
With CoT reconstruction loss0.495 / 84.5% / 21.0%0.318 / 79.5% / 48.7%

Even with the CoT reconstruction loss, the mathematical steps are more readily recoverable from zz than the verbal details. Decoding reproduced the full sequence of reference equations on 48.7% of problems and the exact natural-language CoT on 21.0%. This supports the JEPA motivation for learning predictive representations that preserve reasoning content with less reliance on verbatim reconstruction.

Removing the reconstruction loss improves GSM8K-NL accuracy from 36.30% to 38.1% on GPT-2 and from 42.08% to 48.2% on Llama-3.2-1B. Thus, the CoT reconstruction loss controls the trade-off between answer accuracy and interpretability.

W2, Q2. VISReg ablation and posterior collapse

Without VISReg, the posterior collapses at the default CoT reconstruction weight. We ran a new GPT-2 ablation on GSM8K-NL. We disabled VISReg and multiplied the CoT reconstruction loss LexpL_{\mathrm{exp}} by 1, 3, or 10. The default reconstruction weight in the paper is 1. All other training settings follow the main paper.

MetricPaper setting (VISReg on)No VISReg · ×1No VISReg · ×3No VISReg · ×10
Posterior–prior KL (bits/example)2450.629.274
Answers unchanged after swapping latents across examples6%93%73%55%
Test accuracy34.7%24.6%20.6%14.4%

At the default reconstruction weight (×1), the posterior collapses to the prior, with the KL dropping from 245 to just 0.62 bits per example. Swapping latents across examples leaves 93% of answers unchanged, compared with 6% with VISReg. Test accuracy drops from 34.7% to 24.6%.

Increasing the CoT reconstruction weight only partially alleviates collapse. Even at ×10, swapping latents leaves 55% of answers unchanged, compared with 6% achieved when VISReg was used. Test accuracy declines as the weight increases, from 24.6% to 20.6% to 14.4%.

Thus, VISReg was essential for preventing posterior collapse in this ablation. Increasing the CoT reconstruction weight only partially mitigated collapse and reduced answer accuracy.

W3, Q4. Baseline retraining and reporting

We retrained the baselines in Tables 1 and 2 using their official implementations. Table 8 lists the available published scores. For further details on our baseline reproduction procedure, hyperparameter adjustments, and comparisons with published results, please refer to our response to Reviewer MG2v (W2).

W4. Strategy cues in HiddenBranch

PLRM already outperforms all baselines on HiddenBranch even without using a strategy variable (i.e., setting M=1M=1). We built HiddenBranch to test whether PLRM can learn the diversity of reasoning trajectories when distinct trajectories lead to different valid answers for the same input. However, as described in Section 3.3, a single Gaussian prior can struggle when the same input has multiple reasoning trajectories. Figure 3 shows that the prior blurs distinct posterior modes at the first latent step, and the posterior modes collapse toward a common region over later steps. We introduced the strategy variable to address this limitation. With M=2M=2, the mixture keeps the posterior modes separated across latent steps (Figure 3) and improves valid-answer coverage, raising Avg. TV Score from 0.6252 to 0.7097 (Table 5).

W5, Q3. Multiple components and sampling metrics

We use M=1M=1 in Tables 1 and 2 because the standard benchmark training data provide one reasoning trace per input, as described in Appendix A.3 (lines 793–795). With one observed trace τ\tau per input xx, the aggregated posterior qˉϕ(z∣x)\bar q_\phi(z\mid x) simply reduces to the trace-conditioned posterior qϕ(z∣x,τ)q_\phi(z\mid x,\tau) (Section 3.3, lines 232–234). The target is therefore unimodal at each latent step. A single strategy suffices. With M>1M>1, multiple Gaussian priors are trained to fit the latent distribution of the same CoT.

For pass@kk evaluation, please see our response to reviewer kiD3 (W2, Q4), where we compare PLRM with CoT-SFT and CoLaR under matched token budgets on GSM8K-NL.

To examine whether sampling diverse latent trajectories can support reinforcement learning, we applied GRPO on SVAMP to PLRM and CoT-SFT, both trained on GSM8K-NL with Llama-3.2-1B-Instruct.

MethodBefore GRPOAfter GRPO
PLRM49.40%66.3%
CoT-SFT58.6%72.7%

PLRM improved by 16.9 percentage points, compared with 14.1 for CoT-SFT, narrowing the accuracy gap from 9.2 to 6.4 points. These results show that GRPO can adapt PLRM to a new task without ground-truth CoTs for that task and suggest that other reinforcement learning methods used for LLMs could also be applied to latent reasoning.

W6. Method overview figure

We acknowledge that the existing training schematic in Appendix Figure 4 may not sufficiently convey the overall framework, particularly the interactions among the posterior, strategy-conditioned prior, and training objectives. To improve clarity, we have prepared a revised overview figure (Figure 1) for the main text, illustrating (a) parallel posterior target construction, teacher-forced prior training, and the associated losses, (b) the motivation for the mixture prior in preserving distinct reasoning modes, and (c) inference through strategy sampling, autoregressive latent rollout, and answer generation. You can find the figure in the revised PDF.

Thank you for the rebuttal and the additional experiments. However, my main concerns are not fully resolved.

W1/Q1, W2/Q2 (JEPA framing, reconstruction, and VISReg).

The JEPA framing is still not justified, because the objective itself includes reconstruction term. My concern is not an empirical one about how the trained model behaves, but a logical one about what the objective contains. The reconstruction term is part of the objective, so the objective explicitly asks the latent to reconstruct the CoT, which runs against the JEPA principle. My question was how an objective that contains a reconstruction term, which by itself breaks the core assumption of JEPA, can be called a JEPA. How well the trained model reconstructs does not answer this question. The finding that removing the reconstruction loss improves accuracy is interesting. However, the resulting objective is no longer the ELBO of log⁡pθ(y,τ∣x)\log p_\theta(y, \tau \mid x) derived in the paper, and it does not use a joint-embedding predictive loss either. If the JEPA claim rests on the KL term between the posterior and the prior, any VAE with a learned conditional prior could be called a JEPA. The VISReg result is also not convincing. Since the JEPA framing itself is not justified, it is not logically clear why VISReg improves performance. It is rather interesting that VISReg helps a variational model built with a reconstruction term, which I think comes from its regularization effect. A JEPA claim needs a logical justification, and this is still missing.

W3/Q4 (baseline numbers).

This is partially resolved, but the paper should clearly state that the baseline numbers are reproduced. The reproduced numbers differ considerably from the published ones. However, this is understandable, since such differences can arise from factors such as library versions. Please also point to Table 8 for the published numbers.

W5/Q3 (multiple reasoning strategies on math benchmarks).

The justification for M=1M = 1 is not convincing and is inconsistent with the paper's own claim. The paper and the response to W1 argue that the latent abstracts away verbal details and captures the underlying reasoning. If so, the model should be able to learn distinct reasoning strategies across different problems, even when each problem has a single CoT. Arguing that multimodality cannot arise because each problem has only one CoT contradicts this claim. Moreover, a math problem can usually be solved in several ways, and modern reasoning models are expected to combine solution patterns learned during training to solve new problems at test time. Therefore, not showing this ability on math benchmarks makes the evidence for the paper's claim weak, and the benefit of the mixture remains limited to the constructed HiddenBranch dataset.

Thank you for your follow-up questions and for clarifying your remaining concerns.

W1/Q1, W2/Q2 (JEPA framing, reconstruction, and VISReg).

A joint-embedding predictive architecture can include a reconstruction term. Var-JEPA [1] provides a direct example. Its ELBO combines observation reconstruction with a conditional prior that predicts a learned target latent distribution from context.

LLM-JEPA [2] likewise jointly optimizes a next-token prediction loss and a JEPA embedding-prediction loss in its standard training objective.

PLRM follows the same predictive principle. The prior uses the question and preceding latents to predict the next latent distribution, with the CoT-conditioned posterior providing the target. This prediction of a learned target in latent space is why we describe PLRM’s predictive objective as JEPA-style.

We removed CoT reconstruction in response to the reviewer’s question about its effect on accuracy. This ablation changes the ELBO for the joint likelihood of the answer and the CoT. It retains the prior–posterior KL, so the prior still learns to predict the posterior latent targets.

We use VISReg to discourage collapse of the posterior targets during joint training. If the posterior produces the same Gaussian for every example, the prior can match it and reduce the alignment KL to zero. CoT reconstruction can counteract this by encouraging the latents to retain trace information, but increasing its weight only partially alleviated collapse in our ablation (W2/Q2). VISReg helps keep the targets distinct even when the CoT is reconstructed imperfectly, while answer prediction encourages the latents to retain the reasoning information needed to solve the problem.

W3/Q4 (baseline numbers).

We have clarified that the baseline scores in Tables 1 and 2 were reproduced using the official implementations and added a reference to Table 8 for the published scores.

W5/Q3 (multiple reasoning strategies on math benchmarks).

With one annotated CoT per training problem, the posterior target in PLRM is necessarily a single Gaussian. The empirical aggregated posterior qˉϕ(z∣x)\bar q_\phi(z\mid x) contains exactly one term, qϕ(z∣x,τ)q_\phi(z\mid x,\tau). We introduce multiple strategies (M>1M>1) to preserve distinct modes when the aggregated posterior is multimodal. For the datasets in Tables 1 and 2, each training problem has a single Gaussian target, so we use M=1M=1.

At inference, PLRM’s autoregressive prior can represent a multimodal joint distribution over the full latent trajectory zz for the same problem, even with M=1M=1. Each step is conditionally Gaussian, with a mean and variance that can depend on the preceding sampled latents. Different early samples can therefore lead to different distributions at later steps and produce distinct reasoning trajectories.

Our rebuttal experiments show the practical value of sampling trajectories with M=1M=1. PLRM achieves higher pass@kk than CoT-SFT and CoLaR under matched inference budgets on GSM8K-NL, and GRPO on SVAMP improves PLRM accuracy from 49.40% to 66.3%. HiddenBranch provides a controlled evaluation of the additional benefit of using M>1M>1.

Thank you again for your thoughtful feedback. We would be happy to answer any further questions.

[1] Moritz Gögl and Christopher Yau. [Var-JEPA: A Variational Formulation of the Joint-Embedding Predictive Architecture – Bridging Predictive and Generative Self-Supervised Learning]. ICML 2026.

[2] Hai Huang, Yann LeCun, and Randall Balestriero. [LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures]. ICLR 2026.

Reviewer_3iJXOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Probabilistic Latent Reasoning Models (PLRM), a parallelizable variational framework for probabilistic latent reasoning. During training, a posterior constructs latent target distributions from observed CoT traces, while a problem-conditioned prior learns to predict the latent sequence through teacher forcing, enabling dense supervision and parallel optimization across latent steps. To capture multiple valid reasoning processes for the same problem, PLRM further introduces a discrete strategy variable and a mixture prior. Experiments with GPT-2 and Llama-3.2-1B-Instruct show that PLRM outperforms existing stochastic latent-reasoning baselines while remaining competitive with deterministic methods.

Critical strengths
  • The paper studies an important problem in efficient LLM reasoning: whether lengthy verbalized CoT can be replaced with compact continuous states without losing answer-relevant information. This direction has clear potential to reduce redundant token generation and inference cost.
  • The stochastic formulation is well motivated. Unlike deterministic latent reasoning, which represents only a single trajectory, probabilistic latent states can express uncertainty and support the exploration of multiple valid reasoning paths.
Critical weaknesses
  • The methodological novelty of PLRM appears limited. It merely combines CoLaR’s probabilistic parallel latent prediction, ReGuLaR’s variational CoT guidance, and standard mixture-VAE techniques. The principal technical departure is the replacement of externally constructed latent targets with a jointly learned posterior, whose incremental benefit is also not clearly isolated through controlled comparisons.
  • The theoretical argument for preserving heterogeneous reasoning modes is incomplete. Equation 8 analyzes only the divergence between the aggregated posterior and the mixture prior, whereas the expected alignment objective also penalizes information distinguishing alternative trajectories for the same input. Therefore, matching the aggregated posterior alone does not eliminate mode-collapse pressure.
  • The empirical evaluation is insufficient to substantiate the claimed effectiveness and generality of PLRM. Experiments are restricted to two relatively small backbones, while the ablations do not isolate the contributions of key components. More importantly, PLRM does not demonstrate a consistent or substantial advantage over existing methods.
Questions for authors
  • Could the authors report total training-hours and end-to-end inference latency for PLRM and baselines?
  • PLRM outperforms CoT-SFT on several benchmarks despite compressing the same CoT supervision into substantially fewer latent steps. What mechanism accounts for this improvement, and can the authors provide controlled evidence that it arises from the proposed formulation?
  • Equation 20 determines the strategy responsibility solely from the first latent step and keeps this assignment fixed throughout the trajectory. In this case, how does the proposed method handle trajectories that share an initial reasoning prefix but diverge only at later steps?
3 comments

Thank you for your thoughtful review and for engaging with the motivation behind our work. We particularly appreciate your recognition that compact latent representations could reduce token generation while preserving the information needed to answer the question.

W1. Novelty and the learned posterior

PLRM’s main distinction is that it jointly trains the CoT-conditioned posterior and the question-conditioned prior. The answer prediction loss encourages the posterior to retain the reasoning needed to solve the problem, and KL alignment encourages it to produce latent targets that the prior can predict from the question alone.

In contrast, CoLaR constructs its latent targets from scaled sums of token embeddings, and ReGuLaR regularizes its posterior with visual features extracted from rendered CoT. PLRM’s advantage is particularly large on GSM8K-NL, where the CoTs contain extensive verbal detail. With GPT-2, PLRM achieves 36.30% accuracy, compared with 20.32% for CoLaR and 19.32% for ReGuLaR (Tables 1–2).

Table 3 isolates the benefit of joint training on GPT-2 CommonsenseQA. PLRM achieves 43.40% accuracy, while training the posterior first and freezing it before training the prior yields 18.35–19.49%, close to CoLaR’s 21.62%. This comparison shows the benefit of letting the latent targets adapt as the prior learns to predict them.

W2. Preserving heterogeneous reasoning modes

The expected KL in Eq. (7) includes both Iq(C;τ∣X)I_q(C;\tau\mid X) and Iq(Z;τ∣X,C)I_q(Z;\tau\mid X,C). Of these two terms, only the second can penalize trajectory-specific information during joint posterior–prior training.

We compute the responsibilities ρ\rho in the E-step of our EM-style update in Eq. (20). In our official implementation, we stop gradients through ρ\rho to hold the assignments fixed during the gradient-based M-step. Thus, Iq(C;τ∣X)I_q(C;\tau\mid X) is constant and contributes zero gradient during the M-step, so trajectories assigned to different strategy components retain their separate assignments in that update. This lets distinct first-step modes align with separate Gaussian priors, reducing mode-collapse pressure.

W3. Empirical support and generality

Our empirical evaluation spans seven benchmarks. GSM8K, GSM8K-NL, and CommonsenseQA cover standard mathematical and commonsense reasoning tasks from prior latent reasoning work. SVAMP, GSM-Hard, and MultiArith test out-of-distribution transfer. HiddenBranch, which we constructed for this study, tests coverage of multiple valid assumptions for the same input, each associated with a different proof trajectory.

Training and evaluating PLRM with larger backbones and on more diverse benchmarks within the rebuttal window is challenging given the time and GPU resources available. We will continue these experiments through the end of the rebuttal period and add any results we obtain by the deadline.

Q1. Total training time and inference latency

On GPT-2 GSM8K, we use a single NVIDIA B200 GPU and match batch sizes across methods for each comparison. Total training takes 7.4 GPU-hours for PLRM, 9.8 for CODI, 4.6 for CoLaR, 6.4 for CoT-SFT, and 111.9 for Coconut. Across all 1,319 GSM8K test questions at batch size 1, the mean end-to-end inference latency per question from GPU input to the final answer is 33.2 ms for PLRM, 31.9 ms for CODI, 54.0 ms for CoLaR, 73.1 ms for CoT-SFT, and 74.6 ms for Coconut.

Q2. Gains over CoT-SFT

PLRM’s advantage over CoT-SFT on GPT-2 GSM8K-NL (36.30% vs. 30.93%) comes from reducing the burden of reproducing the rationale’s wording. CoT-SFT learns both the computations and their verbal expression, which can be challenging for capacity in a small model. PLRM predicts the answer directly from a compact latent trajectory. CODI also predicts the answer directly from latent steps without generating a verbal explanation and reaches 35.03%. Adding SIM-CoT’s step-level token reconstruction loss to CODI lowers accuracy to 17.20% in the same GPT-2 setting, supporting our conjecture why verbalizing CoT can be challenging.

To test what PLRM retains in its latents, we used the Llama-3.2-1B model trained on GSM8K-NL with CoT reconstruction. The posterior encoded the ground-truth natural-language CoT, and we trained separate decoders on the same latents to recover either that CoT or the corresponding GSM8K equations. Each cell reports cross-entropy, number recall, and exact match, in that order.

Latent inputNatural languageEquations
With CoT reconstruction loss0.495 / 84.5% / 21.0%0.318 / 79.5% / 48.7%

The full equation sequence was recovered exactly on 48.7% of problems, compared with 21.0% for the original natural-language CoT. This supports the interpretation that PLRM retains core reasoning information while abstracting away some of the original wording.

Q3. Trajectories that diverge after a shared prefix

Trajectories with a shared verbal prefix can receive different strategy assignments by allowing the posterior to see where they diverge. Equation (20) computes an EM-style soft assignment by comparing the first-step posterior qϕ(z1∣x,τ)q_\phi(z_1\mid x,\tau) with each strategy-specific prior component pθ(z1∣x,c=m)p_\theta(z_1\mid x,c=m). When the two traces produce different first-step posteriors, the prior can learn a separate first-step Gaussian for each verbal trace under a different strategy.

Thanks for your response and I have raised my score (1->2) accordingly. However, since the current evaluation remains insufficient to substantiate the claimed effectiveness and generality of PLRM (e.g. more and larger backbones, the consistent advantage over existing methods), I still lean toward rejection.

Reviewer_MG2vOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a variational framework, Probabilistic Latent Reasoning Models (PLRM), which represents reasoning as stochastic trajectories in continuous latent space rather than AR generated CoT text. It uses a posterior to infer latent targets from observed reasoning traces and a problem-conditioned prior learns to predict these targets with teacher forcing. To handle problems with several valid reasoning paths, authors also introduce a mixture prior.

Evaluations are done on math and commonsense reasoning benchmarks. HiddenBranch derived from ProofWriter is introduced to study whether models can produce multiple valid answers and reasoning trajectories for the same input.

Critical strengths

The proposed method addresses an existing limitation of the deterministic latent reasoning. In particular, while replacing CoT with latent steps can reduce the cost of generating intermediate reasoning, deterministic representations fails to directly model alternative reasoning paths. PLRM tackles this issue by introducing stochastic latent trajectories and a mixture prior to represent multiple possible reasoning paths.

Critical weaknesses
  1. The experiment scope is too narrow to support the broader claims about probabilistic latent reasoning. The experiments use GPT-2 and Llama-3.2-1B-Instruct, which is small and old. In addition, most main results are on GSM8K-Aug variants and CommonsenseQA. The only evidence that the model captures distributions over multiple trajectories comes from HiddenBranch, a dataset that authors built themselves. It is unclear how the proposed method performs with stronger models and on more challenging reasoning tasks.

  2. Table 8 has lots of incomplete results, which is inadequate for comparison. In addition, the filled numbers also weaken the main comparison.

  • Where prior work does report numbers, they are consistently higher than the authors' own reruns in Tables 1 and 2.
  • Against the reported numbers, PLRM's lead on CSQA disappears.
  • On GSM8K and GSM8K-NL it sits below CODI and Explicit CoT for both backbones. The claim of "particularly strong performance on GSM8K-Aug-NL and CommonsenseQA" is untenable.
  1. The number of samples RR used for the TV Score and the sampling temperature for each method are not given.

  2. The HiddenBranch evaluation does not directly establish diversity in reasoning trajectories. The main metrics measure answer validity and how closely that the sampled answer distribution matches a uniform distribution over valid answers. These metrics demonstrate coverage of alternative answers, but do not show whether the model learns distinct reasoning strategies.

  3. The evidence that mixture components capture distinct reasoning strategies is limited.

Questions for authors

Why does PLRMs significantly underperform CoT-SFT on GSM8K-Aug-NL and GSM8K with Llama-3.2-1B-Instruct? What causes these performance drops, and how sensitive are the results to the number of latent steps and latent dimension?

2 comments

Thank you for your careful and constructive review. We especially appreciate your observation that modeling alternative reasoning paths remains an open challenge for efficient deterministic latent reasoning, which is what motivated PLRM’s stochastic trajectories and mixture prior.

W1. Experimental scope

Training and evaluating PLRM with larger backbones and on more diverse benchmarks within the rebuttal window is challenging given the time and GPU resources available. We will continue these experiments through the end of the rebuttal period and add any results we obtain by the deadline.

For additional evidence on how PLRM models distributions over multiple reasoning trajectories, please see our responses to Reviewer yooq (W5, Q3) and kiD3 (W2, Q4).

W2. Published versus reproduced baselines

Table 8 reports the originally published baseline results alongside our reproductions in Tables 1 and 2 to ensure transparent comparisons. Missing entries indicate settings not reported in the original studies. Importantly, reproduction discrepancies are not uniformly negative. On Llama-1B GSM8K, our reproduced ReGuLaR and CoLaR accuracies (35.38% and 44.35%) exceed their published results (34.9% and 40.1%), respectively. On GPT-2 GSM8K-NL, PLRM (36.30%) exceeds the published CODI (35.3%) and Explicit CoT (34.8%) results.

We made our best effort to reproduce the originally reported baseline performance. We followed the official implementations, keeping GPT-2 configurations unchanged. For some Llama-1B settings, we adjusted effective batch sizes for GPU utilization and applied linear learning-rate scaling. We also conducted additional learning-rate searches for CoLaR and ReGuLaR. During rebuttal, we reran CODI on CSQA using its original settings, obtaining 69.78%, a relatively modest improvement over our previously reported 67.32%. Both results remain below the published CODI accuracy (74.0%), which is comparable to PLRM's 74.28%.

We acknowledge that deterministic baselines achieve higher accuracy in some settings. However, our primary contribution is not uniformly superior accuracy over deterministic methods, but effective probabilistic joint representation learning with parallelizable teacher-forced training, achieving substantially stronger accuracy than the stochastic latent baselines.

W3. TV Score sampling protocol

We computed TV Score using 10 stochastic samples per input across 6,128 HiddenBranch test inputs. CoT-SFT uses text sampling at temperature 1.0. CoLaR and ReGuLaR use default Gaussian sampling and a token temperature of 1.0. PLRM samples a strategy and continuous latents directly from its learned prior without temperature scaling and keeps the strategy fixed throughout each rollout.

W4, W5. HiddenBranch reasoning diversity

HiddenBranch is fundamentally designed so that each valid answer corresponds to a distinct reasoning strategy for the same input. Each answer specifies a different missing assumption and has a corresponding proof of the same target proposition. Hit Rate measures whether the model generates a valid assumption, and TV Score quantifies coverage of alternative valid assumptions.

In addition to reporting these quantitative results, we decode latent trajectories generated by PLRM into verbalized CoTs. Figures 8–10 in Appendix C.2 show how PLRM derives different valid answers for the same input through distinct reasoning strategies when conditioned on different strategy modes.

Q1. GSM8K performance and latent sensitivity

Switching from GPT-2 to Llama-3.2-1B-Instruct widens the gap to CoT-SFT for all latent reasoning methods on GSM8K-NL and most on GSM8K (Tables 1–2). Llama already achieves 48.7% zero-shot accuracy on GSM8K, versus 1.4% for GPT-2. CoT-SFT can build directly on Llama’s existing text-based reasoning, while latent methods must learn to represent that reasoning in latent space.

On GSM8K-NL in particular, the CoT reconstruction loss contributes to PLRM’s accuracy gap with CoT-SFT. The reconstruction loss in Eq. (5) is important for explainability because it allows PLRM to decode the verbalized CoT from the latent trajectory. However, strong reconstruction supervision forces PLRM to devote part of its limited latent capacity to verbal details that are unnecessary for reasoning.

Under the default hyperparameters, GPT-2 strikes a balance between accuracy and explainability. On Llama-3.2-1B, the default reconstruction weight shifts the balance toward reconstruction and reduces accuracy. Setting the reconstruction weight to zero while keeping all other hyperparameters fixed raises GSM8K-NL accuracy from 42.08% to 48.2% on Llama-3.2-1B and from 36.30% to 38.1% on GPT-2. This illustrates the trade-off between explainability and accuracy.

Latent sensitivity

To assess latent sensitivity, we ran additional ablations on GPT-2 GSM8K, varying the number of latent steps and the latent dimension separately. The default setting uses six latent steps and a latent dimension of 768.

Varying latent steps (768 dimensions)

Latent steps13610
Test accuracy24.9%32.4%38.40%35.6%

Varying latent dimension (6 steps)

Latent dimension128256512768
Test accuracy36.5%37.7%37.4%38.40%

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair for overseeing the review process and all reviewers for their careful evaluation and constructive feedback. During the rebuttal, we provided additional experiments and clarifications addressing the reviewers’ major questions. We would like to take this opportunity to briefly summarize what we have done and what makes our work meaningful.


What makes our work unique?

  • Jointly learned reasoning representations. PLRM jointly trains a CoT-conditioned posterior and a problem-conditioned prior, allowing target latent representations to be learned through answer prediction and predictability from the problem rather than being constructed heuristically.
  • Probabilistic reasoning with parallel training. The posterior constructs latent target distributions in parallel, and the prior learns through parallel teacher forcing. This enables more efficient training than approaches that require sequential latent rollouts.
  • Strong performance with diverse reasoning trajectories. Together, joint representation learning and parallel training allow PLRM to substantially outperform the evaluated stochastic latent reasoning baselines while retaining the ability to generate diverse reasoning trajectories.

What we addressed during the rebuttal

What is the benefit of joint training?

In our response to Reviewer 3iJX (Q2), the reconstruction experiment shows that PLRM’s latent representations preserve core reasoning information more accurately than verbal details. Table 3 in the main paper further demonstrates the importance of joint training through a controlled comparison between fixed targets and jointly learned target representations.

Does stochastic sampling improve problem solving?

On GPT-2 GSM8K-NL, PLRM outperforms CoT-SFT and CoLaR under matched inference budgets corresponding to PLRM’s pass@10, pass@30, pass@60, and pass@100. To further examine the practical benefit of generating diverse reasoning trajectories, we applied GRPO on SVAMP, improving PLRM accuracy from 49.40% to 66.3% without using ground-truth CoTs.

How sensitive is PLRM to its latent settings?

We added ablations varying latent steps and dimensions. As shown in Table 6 in the main paper, PLRM achieves robust performance with largely shared hyperparameters and minimal dataset-specific tuning. For dataset-dependent settings, we explained explicit selection rules.

What are the training and inference costs?

We reported the total training time and mean end-to-end inference latency, together with the hardware, batch sizes, and timing scope. On GPT-2 GSM8K with a B200, PLRM takes 7.4 GPU-hours, compared with an estimated 111.9 GPU-hours for Coconut, a 15.1× training speedup. Under each method’s default configuration, mean inference latency is 33.2 ms per question for PLRM versus 73.1 ms for CoT-SFT, a 2.20× inference speedup.


We respectfully ask that PLRM be evaluated in light of these contributions and the additional evidence provided during the rebuttal. We would be happy to clarify any remaining questions.

Thank you for your time and consideration.

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.