acceptodds
Withdrawn from ICLR 2027

CALR: Continuous Anchored Latent Reasoning via Render-of-Thought Compression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Visual latent reasoning compresses rendered derivations into compact intermediate states, reducing textual reasoning overhead. Existing approaches differ in how they represent these states: continuous methods avoid vocabulary constraints, whereas discrete methods improve accuracy through quantization into a finite codebook. Our analysis of representative continuous and discrete systems identifies two functional requirements: answers must rely on latent states, and those states must carry valid, problem-specific reasoning. Continuous latents influence answers despite collapsed reasoning content, whereas discrete latents retain recoverable intermediate reasoning that answer prediction largely bypasses. To address these challenges, we propose Continuous Anchored Latent Reasoning (CALR), which connects latent formation with answer use through functional anchoring. With reference latents from information-balanced compression, CALR couples latent-mediated answer supervision with derivation-level semantic anchoring: the former routes answer supervision through intermediate states, while the latter grounds their decoded content in problem-specific derivations. A parallel-to-autoregressive curriculum develops sequential reasoning by conditioning subsequent latent blocks on generated prefixes. Evaluations on five mathematical reasoning benchmarks across model families show substantial accuracy gains. Under matched budgets, CALR gains 26.0 percentage points over a comparable continuous latent reasoning method. Further analyses show that its latents support answer prediction and carry problem-specific intermediate reasoning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Render of Thought, Render Compression, Visual Latent Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GPy6OpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether visual latent reasoning states contain useful derivations and influence answer prediction. It introduces CALR, which combines information-balanced optical compression, a frozen question-blind reader, and derivation supervision from a frozen OCR decoder. A parallel producer is subsequently trained to generate four dependent latent blocks. An optional internal reader supports unified inference. Experiments cover five mathematical reasoning benchmarks and several model families. The main configuration improves GSM8K accuracy from RoT's 37.8% to 63.8% with 32 latent states. Latent replacement, decoding, and intermediate-state substitution provide evidence about the resulting representations.

Critical strengths
  1. The method addresses a concrete failure mode with a meaningful training change. Question-blind answer supervision removes the direct question-to-answer shortcut from the external supervisory route. Derivation supervision gives that route structured content. Table 1 reports a substantial 26.0-point GSM8K gain over RoT at the same latent count. Table 2 shows improvements across all twelve reported model-benchmark combinations. Table 4 also supports retaining external supervision during autoregressive refinement.

  2. The causal diagnostics are unusually helpful. Tables 11 and 15 separately test answer-side reliance and dependence between generated blocks. Table 17 is especially valuable. Donors share the original final answer but differ in an intermediate quantity. Feedback substitution produces the corresponding counterfactual answer in 75% and 80% of the two sets. Reader-only substitution does so in only 0.8%. This provides meaningful evidence that intermediate information affects subsequent generation. It strengthens the contribution beyond simply decoding readable text from latents.

Critical weaknesses
  1. The empirical tests do not fully establish the stated functional requirements. C1 in Section 2.1 concerns any change in the per-example answer distribution. Section 2.2 nevertheless declares that DLR-SFT fails C1 using aggregate accuracy, even though deletion changes accuracy from 53% to 51%. Aggregate accuracy cannot determine this distributional condition. C2 requires a valid complete derivation. Table 13 instead measures recovery of reference arithmetic steps, with a score of 0.74 for CALR. It does not report how often the full decoded derivation is valid or contains extra incorrect steps. Appendix E.5 acknowledges semantic and equation discrepancies in the transcriptions. Table 17 provides stronger causal evidence on selected cases, but does not resolve the full-dataset validity claim. The paper should distinguish these supported partial properties from satisfaction of C1 and C2 as formally defined.

  2. The comparison is not sufficiently reconciled or controlled. Unified parallel CALR scores 52.4% in Tables 1 and 10, but the full parallel ablation scores 52.2% in Table 4. CALR's answer accuracy is 63.8% in Table 1 and 63.5% in Table 13. Appendix D.1 assigns these evaluations to the full GSM8K test set. Different seeds or checkpoints could explain these values, but that distinction is not documented. The meaning of the reported uncertainty and the number of runs are also unspecified. In addition, SIM-CoT is discussed but not evaluated. The paper needs a controlled comparison to isolate the value of the proposed anchoring design. Equal latent count alone does not match training resources or inference cost across the different generation procedures.

Questions for authors
  1. Clarification: What operational threshold determines satisfaction of C1? Can you report paired prediction changes or distributional changes, rather than infer the formal condition from aggregate accuracy? For C2, what fraction of decoded trajectories are complete valid derivations, including checks for extra incorrect steps?

  2. Clarification: How were the two 120-problem sets in Table 17 constructed? Were examples selected using native correctness or successful intervention outcomes? Please specify the eligible population, donor selection procedure, and how the substituted block was aligned with the targeted intermediate quantity.

  3. Clarification: Please reconcile 52.4% versus 52.2% and 63.8% versus 63.5%. Which checkpoints, seeds, readout routes, and evaluation runs underlie these tables? What do the uncertainty values represent? The 300-example subset in Table 11 and the external-reader route in Table 14 already explain their different values; those are not the discrepancies at issue.

  4. Evidence request: How does CALR compare with SIM-CoT under the same base model, training examples, and latent budget? Can you also report end-to-end latency and training cost? These measurements would establish whether the accuracy gains justify the additional training stages and readout supervision.

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

This paper proposes CALR, a visual latent reasoning method that renders reasoning traces as images and encodes them into continuous latent states. CALR has three clean components: a resampler that compresses the rendered trace into 32 latents, a question-blind reader that produces the answer, and a low-rank producer that generates the latents from the question. On Qwen3-VL-4B-Instruct with 32 latents, CALR reaches 63.8 on GSM8K-Aug against 37.8 for RoT, and a four-benchmark average of 64.4 against 55.5. The gains carry over to Qwen3-VL-2B and LLaVA-1.6-Mistral-7B, and to Qwen3-VL-8B on MATH-500.

Critical strengths
  • The intervention battery is the part worth keeping. Separating content interventions from interface interventions, and adding question removal on top, is a better diagnostic than the accuracy tables most latent-reasoning papers stop at.
  • The gains are consistent. All 12 model-benchmark cells improve over RoT at the same K, across three backbones from two model families.
  • The recipe is portable. It needs no architecture change and no OCR model at inference time, and neither of the two main tricks is specific to vision, so it could be bolted onto text-domain latent methods.
Critical weaknesses
  • The headline +26 is on GSM8K-Aug, the training distribution, and against the weakest visual-latent baseline (RoT). Against DLR with RL, the average gain is only 1.2 with DLR using about five latents, and CALR is below it on MATH-500 at 4B. The baseline rows match DLR’s tables to the decimal, yet the paper’s own DLR-SFT reproduction disagrees (44.0 vs 48.0).
  • The efficiency claim lacks the obvious controls. There is no explicit-CoT baseline at a matched token budget (for example, CoT truncated to about 256/512... tokens), no accuracy-vs-K curve, and no latency or FLOPs.
  • The diagnostics are partly circular. C2 is measured with a decoder that appears in CALR’s training loss, so the comparison with RoT and DLR (0.34 against 0.07 and 0.09) mostly reflects which system the decoder favors. The C1 evidence is confounded as well: CALR’s reader is trained with question masking and the baselines are not, so the No-x result (61.9 against 2.0 and 1.0) is not like-for-like. A donor-following rate would say more than an accuracy drop under swap.
  • The diagnosis has been published before. “Do Latent Tokens Think?” (arXiv 2512.21711) reports the same swap finding on Coconut but is not cited. SIM-CoT is cited but never compared.
Questions for authors

N/A

Reviewer_ebY5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies visual latent reasoning via render-of-thought compression and argues that existing continuous and discrete systems fail complementary handoff requirements: latent reliance (C1) and reasoning capability (C2). Continuous latents can drive answers while carrying little problem-specific content, whereas discrete latents can be decoded into valid derivations yet are largely bypassed by the answer model. The authors propose Continuous Anchored Latent Reasoning (CALR), which couples question-blind, latent-mediated answer supervision with derivation anchoring through a frozen DeepSeek-OCR2 decoder, and trains a parallel-to-autoregressive curriculum over continuous latent blocks. On grade-school math benchmarks with Qwen3-VL-4B-Instruct, unified autoregressive CALR reaches 63.8% on GSM8K-Aug versus 37.8% for RoT under matched 32-latent budgets, and interventions indicate that the learned latents both support answers and recover problem-specific intermediate content.

Critical strengths
  1. The C1/C2 framing is concrete and useful. The preliminary interventions in Figure 2 cleanly separate the continuous failure (interface load-bearing, content collapsed) from the discrete failure (content recoverable, interface bypassed), which makes the later design choices easier to evaluate than a purely empirical comparison.
  2. The diagnostics go beyond accuracy tables. Cross-problem latent swaps drop accuracy from 64.3% to 1.4% with the question retained, decoded trajectories recover substantially more intermediate values for the current problem than for others, and the ablations in Tables 4–5 show that external supervision, reader masking, and information-balanced resampling each matter.
Critical weaknesses

1. Recoverability is trained directly, rather than reasoning in the latent space

CALR trains the latents so that the frozen OCR decoder can reconstruct the reference derivation (Eqs. 9 and 16), and then uses the same decoder to show that the latents contain reasoning (C2, Table 13). Recoverability is thus the training objective itself, not a consequence of the model reasoning in the latent space. The model may simply learn to write a recoverable textual derivation into its latents without performing the corresponding computation. The transfer results discussed in Weakness 3 are consistent with this concern.

2. The gains from OCR reconstruction and from latent reasoning are not separated

The whole method is built on the frozen DeepSeek-OCR2: it defines the reference latent space in Stage I, provides the reconstruction loss in Stage II, and serves as the evaluation decoder. However, no ablation removes the OCR term or anchors the latents without OCR; It is therefore unclear how much of the improvement comes from OCR reconstruction supervision, and how much from latent reasoning itself.

3. Large gains appear mainly on the training distribution

The model is trained only on GSM8K-Aug, where CALR improves over RoT by 26.0 points, while the gains on GSM-Hard, SVAMP, and MultiArith are only 2.7, 4.7, and 2.5 points (Table 1, Qwen3-VL-4B). GSM-Hard keeps the GSM8K test problems and only enlarges the numbers, yet CALR recovers 68% of the CoT-SFT gain over Direct-SFT on GSM8K-Aug but only 17% on GSM-Hard, with the same pattern on the other two backbones in Table 2. This suggests that the model may largely reproduce the CoT seen in training rather than acquire a transferable reasoning ability.

Questions for authors
  1. Could you explain what evidence, beyond decodability through the training decoder, shows that the latents perform the reasoning rather than store a recoverable derivation?
  2. Could you show that the improvement does not mainly come from the DeepSeek-OCR2 reconstruction supervision, but from latent reasoning itself?
  3. Could you explain why the gains are much smaller outside GSM8K-Aug, especially on GSM-Hard, and whether the decoded intermediate values remain correct there(a case study on GSM-Hard, including failure cases, would be welcome)?

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

Related papers

Discussion (0)

Sign in to comment.