Chain-of-Reasoning Embedding (CoRE)

• Chain-of-Reasoning Embedding (CoRE) is a framework that maps reasoning processes into latent representations, underpinning tasks such as segmentation, retrieval, and self-evaluation.
• It utilizes various mechanisms like special-token embeddings, trajectory-based representations, variational compression, and external reward signals to integrate reasoning into model objectives.
• By coupling reasoning with downstream tasks, CoRE enhances performance, efficiency, and interpretability across multimodal and computational applications.

Chain-of-Reasoning Embedding (CoRE) denotes a family of methods that convert a reasoning process into a latent representation and then use that representation for downstream control, retrieval, compression, or self-evaluation. In recent work, the term is used in several closely related but non-identical senses: as a dedicated special-token embedding that conditions segmentation, as a reasoning-conditioned multimodal embedding for retrieval, as a trajectory of step-level hidden states for metacognitive monitoring, as a variational latent chain that replaces explicit chain-of-thought (CoT), and as an external embedding of reasoning text used as a reinforcement-learning reward signal (Xie et al., 6 Mar 2026, Wang et al., 7 Apr 2026, Li et al., 8 Jul 2025, Wang et al., 30 Jan 2026, He et al., 20 Apr 2026).

1. Conceptual scope and principal definitions

Recent formulations do not present CoRE as a single canonical object. Instead, they define a common pattern: reasoning is made explicit or latent, mapped into an embedding space, and then coupled to a task objective. This shared pattern appears across segmentation, multimodal retrieval, mathematical metacognition, latent reasoning, and audio reasoning (Xie et al., 6 Mar 2026, Wang et al., 7 Apr 2026, Li et al., 8 Jul 2025, Wang et al., 30 Jan 2026, He et al., 20 Apr 2026).

Formulation	CoRE object	Operational role
CORE-Seg	final hidden state of <seg>	prompt for MedSAM 2 decoder
MMEmb-R1	hidden state of <r_emb> after CoT	reasoning-enhanced retrieval embedding
CoRE-Eval	trajectory $\tau=(\mathbf{h}_1,\dots,\mathbf{h}_T)$	label-free early-exit signal
ReGuLaR	latent chain $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$	compressed reasoning for answer generation
Audio-DeepThinker	BGE-M3 embedding of reasoning chain	semantic similarity reward in RL

In "CORE-Seg" (Xie et al., 6 Mar 2026), the CoRE-like object is the final-layer hidden state of a learnable <seg> token appended to the answer. The paper describes this vector as “a condensed semantic anchor” that aggregates both the preceding rationale and localization cues. In "MMEmb-R1" (Wang et al., 7 Apr 2026), the corresponding object is a reasoning-conditioned embedding extracted from a special <r_emb> token after a generated rationale, contrasted with a direct embedding from <d_emb>. In "CoRE: Enhancing Metacognition with Label-free Self-evaluation in LRMs" (Li et al., 8 Jul 2025), CoRE is explicitly defined as a trajectory of step-level hidden states, one vector per reasoning step, interpreted as the internal state evolution of the model. "ReGuLaR" (Wang et al., 30 Jan 2026) instead replaces explicit reasoning tokens by a short sequence of latent states $\mathbf{Z}$, each state encoding a segment of the original CoT. "Audio-DeepThinker" (He et al., 20 Apr 2026) uses an external text embedding model to embed full reasoning chains and score generated reasoning against reference chains.

A recurring misconception is that CoRE simply means “adding CoT before the answer.” The cited work shows a stricter requirement: the reasoning representation must be coupled to a downstream objective, whether by a prompt adapter, contrastive supervision, variational regularization, geometric diagnostics, or RL reward shaping (Xie et al., 6 Mar 2026, Wang et al., 7 Apr 2026, Li et al., 8 Jul 2025, Wang et al., 30 Jan 2026, He et al., 20 Apr 2026).

2. Representational mechanisms

The most explicit special-token instantiation appears in CORE-Seg. Given image $I$ and query $Q$, the multimodal reasoning module produces

$$\hat{O}=\langle think\rangle \hat{R} \langle /think\rangle \langle answer\rangle \hat{A} \langle /answer\rangle,$$

with <seg> at the end of the answer. The final hidden state of that token is

$P \in \mathbb{R}^{B \times 1 \times D},$

and serves as the reasoning embedding. A Semantic-Guided Prompt Adapter then maps $P$ into a prompt $\hat{P}\in\mathbb{R}^{B\times1\times256}$ via stacked ResMLPs and cross-attention with a learned query, after which MedSAM 2 decodes the mask from $(Z_{\text{img}},\hat{P})$ (Xie et al., 6 Mar 2026).

MMEmb-R1 uses an analogous special-token construction, but for multimodal retrieval rather than segmentation. For any multimodal input $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$0, the direct embedding is taken from <d_emb>, while the reasoning embedding is taken from <r_emb> after a rationale $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$1: $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$2 The paper formalizes reasoning as a latent variable and approximates a reasoning distribution by generating diverse candidate rationales with multiple worker MLLMs, then selecting pair-helpful rationales through counterfactual gains $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$3 computed by an evaluator on query-target matching confidence (Wang et al., 7 Apr 2026).

CoRE-Eval defines a different representational regime. If the reasoning chain is segmented into steps $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$4, then each step embedding is the last-layer hidden state at the last token of that step,

$\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$5

and the CoRE object is the trajectory

$\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$6

This trajectory is not used as a static feature. Instead, it is analyzed dynamically through step-to-step magnitude change and cosine similarity, then compressed into a composite scalar signal

$\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$7

for periodicity detection (Li et al., 8 Jul 2025).

ReGuLaR moves from explicit reasoning embeddings to latent reasoning states. The reasoning chain $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$8 is partitioned into segments $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$9, and each segment is represented by a latent state $\mathbf{Z}$0 sampled autoregressively: $\mathbf{Z}$1 Each latent state is thus a compressed representation of a segment of CoT, and the answer is generated from $\mathbf{Z}$2 without decoding intermediate reasoning at inference (Wang et al., 30 Jan 2026).

Audio-DeepThinker uses an external embedding formulation. Generated reasoning $\mathbf{Z}$3 and reference reasoning $\mathbf{Z}$4 are embedded by BGE-M3,

$\mathbf{Z}$5

and their cosine similarity contributes directly to reward: $\mathbf{Z}$6 Here the embedding is not the policy’s own hidden state but a frozen semantic representation of the full reasoning text (He et al., 20 Apr 2026).

3. Supervision and optimization strategies

The training regimes attached to CoRE vary substantially, but all of them tie reasoning embeddings to task performance rather than treating CoT as an auxiliary output.

CORE-Seg uses a progressive strategy from supervised fine-tuning to GRPO. Stage 1 optimizes

$\mathbf{Z}$7

where $\mathbf{Z}$8 supervises both reasoning and answer generation, and $\mathbf{Z}$9 supervises the segmentation mask. Stage 2 adds GRPO with a dual-granularity reward mechanism comprising format reward $I$0, bipartite matching reward $I$1, and mask reward $I$2. The mask reward falls back to GIoU when Dice is below $I$3, and otherwise uses $I$4 with $I$5, specifically to mitigate reward sparsity (Xie et al., 6 Mar 2026).

MMEmb-R1 first aligns reasoning with pairwise retrieval by sampling selected rationales and optimizing a joint loss

$I$6

This combines reasoning-enhanced contrastive loss, CoT generation loss, and direct embedding contrastive loss. It then learns an adaptive policy over $I$7 with GRPO, using a reward

$I$8

where $I$9 balances similarity utility against rationale length. This design explicitly addresses two issues named in the paper: structural misalignment between instance-level CoT and pairwise contrastive supervision, and the inefficiency of always-on reasoning (Wang et al., 7 Apr 2026).

ReGuLaR formulates latent reasoning as a sequential VAE. Its training objective combines answer generation from latent states, reasoning reconstruction from each latent state, and KL regularization between posterior and a segment-conditioned prior: $Q$0 The prior is unusual: each reasoning segment is rendered as an image, encoded by a frozen DeepSeek-OCR encoder, and adapted into a target latent mean $Q$1. The posterior is then regularized toward $Q$2 (Wang et al., 30 Jan 2026).

Audio-DeepThinker uses RL rather than supervised reasoning fine-tuning. Stage 1 employs

$Q$3

with

$Q$4

and the similarity term is applied only when the answer is correct. Stage 2 replaces the hybrid reward with an LLM-only similarity reward,

$Q$5

to allow more diverse but still logically valid reasoning. Optimization uses GDPO with decoupled normalization of reward groups and a KL penalty with $Q$6 (He et al., 20 Apr 2026).

CoRE-Eval is the outlier: it is explicitly training-free and label-free. The framework computes the composite signal online, performs sliding-window Pearson correlation across candidate periods, and enters a Cycle state when correlation exceeds a threshold $Q$7 and the estimated period remains stable for $Q$8 steps. The reported default hyperparameters are $Q$9, $$\hat{O}=\langle think\rangle \hat{R} \langle /think\rangle \langle answer\rangle \hat{A} \langle /answer\rangle,$$0, $$\hat{O}=\langle think\rangle \hat{R} \langle /think\rangle \langle answer\rangle \hat{A} \langle /answer\rangle,$$1, and $$\hat{O}=\langle think\rangle \hat{R} \langle /think\rangle \langle answer\rangle \hat{A} \langle /answer\rangle,$$2 (Li et al., 8 Jul 2025).

4. Task instantiations and empirical results

In medical image segmentation, CORE-Seg introduces ComLesion-14K, described as the first diverse CoT benchmark for reasoning-driven complex lesion segmentation. Each sample is a 5-tuple $$\hat{O}=\langle think\rangle \hat{R} \langle /think\rangle \langle answer\rangle \hat{A} \langle /answer\rangle,$$3, and the dataset contains 13,678 samples, 8 imaging modalities, 9 anatomical regions, and 31 disease categories. The full model reports a mean Dice of $$\hat{O}=\langle think\rangle \hat{R} \langle /think\rangle \langle answer\rangle \hat{A} \langle /answer\rangle,$$4, a mean IoU of $$\hat{O}=\langle think\rangle \hat{R} \langle /think\rangle \langle answer\rangle \hat{A} \langle /answer\rangle,$$5, and a failure rate of $$\hat{O}=\langle think\rangle \hat{R} \langle /think\rangle \langle answer\rangle \hat{A} \langle /answer\rangle,$$6, compared with $$\hat{O}=\langle think\rangle \hat{R} \langle /think\rangle \langle answer\rangle \hat{A} \langle /answer\rangle,$$7 mean Dice and $$\hat{O}=\langle think\rangle \hat{R} \langle /think\rangle \langle answer\rangle \hat{A} \langle /answer\rangle,$$8 failure for the second-best baseline, LISA-3B, and $$\hat{O}=\langle think\rangle \hat{R} \langle /think\rangle \langle answer\rangle \hat{A} \langle /answer\rangle,$$9 mean Dice with $P \in \mathbb{R}^{B \times 1 \times D},$0 failure for SegZero-3B. Ablations show that removing the Semantic-Guided Prompt Adapter drops Stage-1 mean Dice from $P \in \mathbb{R}^{B \times 1 \times D},$1 to $P \in \mathbb{R}^{B \times 1 \times D},$2, and that using whole-answer hidden states as prompt yields weaker RL gains than using <seg> (Xie et al., 6 Mar 2026).

In multimodal retrieval, MMEmb-R1 reports an MMEB-V2 score of $P \in \mathbb{R}^{B \times 1 \times D},$3 with only 4B parameters. Reported baselines include Embed-RL at $P \in \mathbb{R}^{B \times 1 \times D},$4, RzenEmbed-v1 at $P \in \mathbb{R}^{B \times 1 \times D},$5, and UME-R1 at $P \in \mathbb{R}^{B \times 1 \times D},$6. An ablation on reasoning policy shows Direct-only at $P \in \mathbb{R}^{B \times 1 \times D},$7, Always Reason at $P \in \mathbb{R}^{B \times 1 \times D},$8, Always Direct at $P \in \mathbb{R}^{B \times 1 \times D},$9, Random $P$0 at $P$1, the learned adaptive policy at $P$2, and Oracle at $P$3. Latency experiments with Qwen2-VL-2B show UME-R1 at $P$4 s, MMEmb-R1 Always Reason at $P$5 s, and MMEmb-R1 Adaptive at $P$6 s on the reported subset, with the adaptive policy also achieving the highest accuracy among those three settings (Wang et al., 7 Apr 2026).

In metacognitive early exit, CoRE-Eval is evaluated on GSM8K, MATH-500, and AIME 2024 using DeepSeek-R1-Distill-Qwen 7B, 14B, and 32B. Across all models and datasets, it reduces CoT length by $P$7 to $P$8 while improving accuracy by about $P$9 on average. On AIME 2024 with the 32B model, the reported accuracy rises from $\hat{P}\in\mathbb{R}^{B\times1\times256}$0 to $\hat{P}\in\mathbb{R}^{B\times1\times256}$1, while average reasoning length falls from $\hat{P}\in\mathbb{R}^{B\times1\times256}$2 to $\hat{P}\in\mathbb{R}^{B\times1\times256}$3 tokens (Li et al., 8 Jul 2025).

In latent reasoning compression, ReGuLaR reports average accuracy $\hat{P}\in\mathbb{R}^{B\times1\times256}$4 with average reasoning length $\hat{P}\in\mathbb{R}^{B\times1\times256}$5 on the LLaMA-1B setting summarized in the paper, compared with CoLaR* at $\hat{P}\in\mathbb{R}^{B\times1\times256}$6 and $\hat{P}\in\mathbb{R}^{B\times1\times256}$7, Coconut at $\hat{P}\in\mathbb{R}^{B\times1\times256}$8 and $\hat{P}\in\mathbb{R}^{B\times1\times256}$9, and iCoT at $(Z_{\text{img}},\hat{P})$0 and $(Z_{\text{img}},\hat{P})$1. Under extreme compression with $(Z_{\text{img}},\hat{P})$2, ReGuLaR keeps $(Z_{\text{img}},\hat{P})$3 and outperforms CoLaR on GSM8K-Aug-NL, AQUA-RAT, and MATH; the MATH example reported in the details gives CoLaR at $(Z_{\text{img}},\hat{P})$4 average accuracy with $(Z_{\text{img}},\hat{P})$5 versus ReGuLaR at $(Z_{\text{img}},\hat{P})$6 with $(Z_{\text{img}},\hat{P})$7 (Wang et al., 30 Jan 2026).

In audio reasoning, Audio-DeepThinker reports $(Z_{\text{img}},\hat{P})$8 on MMAR, $(Z_{\text{img}},\hat{P})$9 on MMAU-test-mini, and $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$00 on MMSU. A Stage-1 reward ablation reports MMAR accuracy and Rubrics of $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$01 for base reward, $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$02 for base plus Audio-Thinker-style rewards, and $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$03 for base plus consistency plus hybrid similarity. The progressive Stage 1$\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$042 setup reports the best combined result in the cited summary, with MMAR $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$05 and Rubrics $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$06 (He et al., 20 Apr 2026).

5. Interpretability, overthinking, and reasoning quality

A central motivation for CoRE is that it makes reasoning operational rather than merely rhetorical. In CORE-Seg, Stage 2 reasoning is reported to become more careful and image-grounded, often using cautious language such as “possible” and “may indicate” under uncertainty, and clinician review on 200 random cases judged Stage 2 masks and reasoning to be more clinically realistic and more consistent with expert judgment. This indicates that the embedding extracted from <seg> is not only an internal control signal but part of an inspectable reasoning-segmentation pipeline (Xie et al., 6 Mar 2026).

CoRE-Eval treats the latent trajectory itself as a diagnostic object. The paper reports three geometric signatures: abrupt changes in magnitude and angle for strategy shifts, quasi-periodic oscillations for semantic redundancy and reasoning loops, and inverse correlation between magnitude and angle for latent stagnation. Early exit is triggered not by a correctness classifier but by persistent periodicity in the recent trajectory. The authors explicitly caution that CoRE geometry is not a direct measure of understanding or calibrated confidence; rather, it is a heuristic metacognitive mechanism grounded in hidden-state dynamics (Li et al., 8 Jul 2025).

MMEmb-R1 and Audio-DeepThinker both frame a related phenomenon as overthinking. MMEmb-R1 reports that always reasoning is inferior to adaptive control, and its Pareto analysis shows accuracy rising with reasoning ratio, peaking around approximately $\mathbf{Z}=\{\mathbf{z}_k\}_{k=1}^K$07 reasoning, and then declining as nearly all samples invoke reasoning. The paper identifies this decline as empirical evidence that indiscriminate reasoning can obscure salient semantic signals for simple cases (Wang et al., 7 Apr 2026).

Audio-DeepThinker contributes a mechanistic account of internal reasoning emergence. Its interpretability analysis reports that RL training reshapes upper-layer MoE gating mechanisms more than expert weights, with gating drift increasing sharply in upper layers, and that reasoning tokens “crystallize progressively in the upper transformer layers.” Logit-lens analysis further shows that the decision to enter reasoning mode is made earlier than decisions about reasoning content or final answer. This suggests that, in internal-state terms, reasoning-aware embeddings may be most informative in upper-layer representations and routing patterns (He et al., 20 Apr 2026).

ReGuLaR addresses interpretability differently. It does not preserve human-readable reasoning at inference, but it regularizes latent reasoning with rendered CoT images that preserve the exact content of the original rationale in a compact visual-semantic form. This suggests a trade-off: explicit CoRE formulations provide direct textual inspectability, whereas latent CoRE formulations emphasize compression and efficiency (Wang et al., 30 Jan 2026).

6. Limitations, unresolved issues, and broader significance

Several limitations recur across the literature. CoRE-Eval requires access to hidden states and therefore applies only to white-box or instrumentable models; it also incurs window-correlation overhead, which the paper notes may not be compensated on simpler tasks such as GSM8K (Li et al., 8 Jul 2025). CORE-Seg is currently 2D only, has slower inference due to explicit CoT generation, and relies on synthetic CoT generated by GPT-4o rather than clinician-authored reasoning (Xie et al., 6 Mar 2026). MMEmb-R1 uses a pipeline in which reasoning generation, pair-aware selection, and adaptive RL are separate stages with offline components, and its policy is binary rather than multi-level (Wang et al., 7 Apr 2026). ReGuLaR depends on CoT supervision, offline rendering and visual encoding, sentence-level or similar segmentation heuristics, and modeling assumptions such as the stated conditional independence assumption (Wang et al., 30 Jan 2026). Audio-DeepThinker explicitly notes that embedding similarity can over-constrain reasoning diversity, which is why the embedding term is removed in Stage 2; moreover, its reasoning embeddings are generic text embeddings rather than audio-specific latent states (He et al., 20 Apr 2026).

Another important correction to a common assumption is that explicit CoT by itself is not uniformly beneficial. CORE-Seg reports that “SFT with CoT (no RL)” yields a slight drop relative to an SFT baseline without explicit CoT, and that the full gains come from reward-shaped reasoning plus segmentation supervision (Xie et al., 6 Mar 2026). MMEmb-R1 similarly reports that always-on reasoning underperforms adaptive reasoning, and Audio-DeepThinker removes embedding similarity in its second stage precisely because semantic closeness to a reference chain can penalize valid alternative reasoning paths (Wang et al., 7 Apr 2026, He et al., 20 Apr 2026).

Taken together, these works define CoRE less as a single architecture than as a design space. One branch treats reasoning embeddings as special-token anchors that directly condition another module; another treats them as retrieval embeddings tied to pairwise supervision; another models them as latent trajectories for self-evaluation; another compresses them into a variational latent chain; and another uses external reasoning embeddings as process-level reward signals. This suggests that the durable core of the concept is not any specific token, encoder, or loss, but the systematic conversion of reasoning into a manipulable latent object that can be optimized, probed, regularized, or used to control computation (Xie et al., 6 Mar 2026, Wang et al., 7 Apr 2026, Li et al., 8 Jul 2025, Wang et al., 30 Jan 2026, He et al., 20 Apr 2026).