Skip to content

SHIFT: Self-reconstruction Harnesses Implicit Fine-grained Thinking for Retrieval

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/lyx0727/SHIFT
Area: LLM Reasoning
Keywords: Implicit Reasoning Retrieval, Self-Reconstruction Supervision, Dense Retrieval, Residual Projection, Bidirectional Attention Aggregation

TL;DR

This paper proposes SHIFT, an efficient reasoning-intensive retrieval framework that decouples generation noise from metric embeddings via residual projection, dynamically aggregates continuous latent thoughts through bidirectional self-attention, and supervises the latent reasoning space via fine-grained autoregressive self-reconstruction of explicit chain-of-thought trajectories with 2D relative position encodings—surpassing both cascade rewriting pipelines and prior implicit reasoning retrievers with only 3 latent thinking steps.

Background & Motivation

Dense retrieval models powered by generative Large Language Models (LLMs) have emerged as foundational pillars for modern open-domain question answering, web search, and retrieval-augmented generation (RAG) pipelines. Traditional dual-encoder paradigms typically treat causal LLMs as conventional sentence encoders, fine-tuning final token representations (such as <EOS>) using contrastive learning objectives. However, this standard formulation fails to activate the deep multi-step reasoning capabilities inherent in LLMs, rendering them vulnerable when handling complex, reasoning-intensive queries characterized by semantic ambiguity, multi-hop logical dependencies, and implicit user intent. To address this limitation, the predominant "rewrite-then-retrieve" paradigm employs an external LLM reasoner or chain-of-thought prompt to generate verbose textual rationales or expanded queries before feeding them into the retriever. While empirically effective, this cascaded workflow incurs severe latency overheads, increases token consumption, and prevents end-to-end optimization of the retrieval pipeline.

Migrating intermediate reasoning from discrete token space to continuous latent representations (as explored in recent methods like GIRCSE and LaSER) offers a promising avenue to preserve differentiability and minimize inference latency. Nevertheless, naively incorporating continuous reasoning during query encoding reveals two deep-seated structural mismatches. The first is representation mismatch: the hidden states produced by autoregressive causal LLMs are naturally optimized to summarize preceding conversational contexts for next-token generation, exhibiting severe anisotropy and collapsing into a narrow hyper-cone. Consequently, directly using or mean-pooling these generation-oriented hidden states retains extraneous contextual noise rather than discriminative retrieval signals, causing distinct latent reasoning steps to collapse into homogeneous static embeddings. The second is supervision mismatch: standard contrastive loss supervises only the final similarity between pooled query and document representations, providing zero guidance on what individual intermediate reasoning states should encode; conversely, distillation methods like LaSER enforce strict point-to-point alignment between latent states and explicit teacher hidden states, thereby copying the teacher model's severe representational redundancy where step-wise cosine similarities already approach 1.0.

To overcome these representational and supervisory bottlenecks, the authors re-examine how autoregressive generative priors can be leveraged internally without importing redundant teacher noise. A viable retriever must isolate discriminative metric features from generative noise, dynamically adapt to varying query complexities without enforcing rigid monotonic priors, and internalize fine-grained explicit reasoning trajectories via non-rigid generative supervision. Core idea: decouple generative noise via residual projection, adaptively pool latent thoughts via bidirectional self-attention, and supervise the continuous reasoning space via 2D-position-aware autoregressive self-reconstruction of explicit chain-of-thought trajectories.

Method

Overall Architecture

The overall SHIFT pipeline operates through four modular phases: (1) Continuous Soft Thinking: Given an input query \(q\), the causal LLM autoregressively generates \(K\) continuous latent soft tokens \(t_j\) and corresponding reasoning hidden states \(h_j\) by taking the expected value over the vocabulary embedding matrix under language modeling head distributions; (2) Residual Retrieval Projection: A lightweight two-layer residual projection module maps raw hidden states into a dedicated retrieval subspace, stripping away generation-oriented contextual noise while preserving logical trajectories; (3) Bidirectional Attention Aggregation: Multi-head self-attention models the all-to-all interactions among the projected latent states, dynamically reweighting reasoning steps before mean pooling into the final query vector \(v_q\); (4) Fine-Grained Self-Reconstruction: During training, step-level latent states are expanded to match explicit reasoning token lengths, modulated by 2D step-level and intra-step relative position embeddings, and tasked with reconstructing explicit reasoning traces through autoregressive next-token prediction, optimized jointly with the contrastive retrieval loss.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    A["Input Query q"] --> B["Continuous Soft Thinking<br/>Autoregressively generate K continuous soft tokens tj and hidden states hj"]
    B --> C["Residual Retrieval Projection<br/>Two-layer residual MLP filters causal generation noise"]
    C --> D["Bidirectional Attention Aggregation<br/>Multi-head self-attention interaction and mean pooling to vq"]
    D --> E["Dual-Tower Retrieval Matching<br/>Static document encoding vd without reasoning and InfoNCE loss"]
    B --> F["Fine-Grained Self-Reconstruction<br/>2D positional expansion of latent states + NTP of explicit CoT"]
    E --> G["End-to-End Joint Training<br/>L = Lcl + Lntp"]
    F --> G

Key Designs

1. Continuous Soft Thinking: Differentiable Reasoning in Continuous Latent Space

To resolve the high latency, verbosity, and non-differentiable sampling barriers of discrete chain-of-thought generation, SHIFT executes intermediate thinking directly in continuous embedding space. Given the input query token embeddings \(X = E(q)\), the causal LLM backbone computes the initial query representation \(h_0 = f_\theta(E(q))\). At each reasoning step \(j \in \{1, \dots, K\}\) (fixed to \(K=3\) in experiments), the model projects the prior hidden state \(h_{j-1}\) through the language model head \(W_{\text{lm}} \in \mathbb{R}^{m \times |V|}\) to produce vocabulary logits \(\text{logit}_j = W_{\text{lm}} h_{j-1}\), yielding a soft probability distribution \(p_j = \text{softmax}(\text{logit}_j)\). Instead of executing discrete token selection (e.g., argmax or sampling), SHIFT computes the continuous soft token \(t_j\) as the expectation of token vectors over embedding matrix \(E \in \mathbb{R}^{|V| \times m}\): $\(t_j = p_j^\top E = \text{softmax}(W_{\text{lm}} h_{j-1})^\top E\)$ This continuous token is appended to the sequence to derive the next hidden state \(h_j = f_\theta([E(q); t_1, \dots, t_j])\). Through causal attention, each latent token simultaneously attends to the query context and preceding reasoning steps, faithfully emulating the sequential dependency of chain-of-thought reasoning within fully differentiable computation that costs merely 3 forward passes.

2. Residual Retrieval Projection: Decoupling Metric Subspace from Anisotropic Generation Noise

Because the hidden states \(h_i\) of autoregressive LLMs are optimized to predict future tokens by compressing broad historical context, they suffer from well-documented representational degeneration (anisotropy), clustering within a narrow cone. Directly performing similarity matching over raw generative hidden states introduces contextual redundancy and suppresses discriminative capacity. SHIFT introduces a lightweight two-layer residual projection module \(\phi_{\text{ret}}\) over the latent reasoning sequence \(H_{\text{latent}} = [h_1, \dots, h_K]\): $\(\tilde{h}_i = h_i + \phi_{\text{ret}}(h_i)\)$ The residual architecture preserves the foundational reasoning trajectory while learning non-linear transformations that filter out generation-oriented noise. As evidenced by ablation studies, removing this residual projection induces severe performance degradation across all retrieval benchmarks, confirming the critical importance of disentangling the metric retrieval space from the generative reasoning space.

3. Bidirectional Attention Aggregation: Adaptive Multi-Step Fusion Beyond Monotonic Priors

Existing implicit reasoning retrievers commonly rely on static mean pooling or enforce monotonic inductive biases (e.g., GIRCSE's assumption that later reasoning steps are intrinsically superior; GRU or causal attention assigning higher weights to final steps). However, real-world queries vary drastically in their required reasoning depths: simple queries favor early generalized semantics, whereas intricate theorem-proving queries demand late-stage logical abstractions. SHIFT breaks free from monotonic assumptions by applying bidirectional multi-head self-attention over the projected representations \(\tilde{H} = [\tilde{h}_1, \dots, \tilde{h}_K] \in \mathbb{R}^{K \times m}\): $\(Q_i = \tilde{H} W_{Q_i}, \quad K_i = \tilde{H} W_{K_i}, \quad V_i = \tilde{H} W_{V_i}\)$ $\(o_i = \text{softmax}\left(\frac{Q_i K_i^\top}{\sqrt{m / n_h}}\right) V_i, \quad z = \text{concat}(o_1, \dots, o_{n_h}) W_O\)$ The resulting contextualized matrix \(z \in \mathbb{R}^{K \times m}\) dynamically captures multi-step interdependencies based on the query's intrinsic difficulty, and is subsequently aggregated via mean pooling: \(v_q = \frac{1}{K} \sum_{i=1}^K z_i\). This all-to-all interaction mechanism flexibly weights individual steps according to query-specific needs, avoiding the representational bottlenecks of unidirectional causal pooling.

4. Fine-Grained Self-Reconstruction: Autoregressive Supervision Without Teacher Anisotropy

Prior frameworks like LaSER enforce point-to-point alignment between student latent states and teacher hidden states extracted from explicit reasoning traces. However, empirical analysis shows that teacher hidden states at different reasoning steps already exhibit pairwise cosine similarities exceeding 0.99; enforcing alignment directly injects this severe anisotropy and redundancy into the student. SHIFT completely eliminates step-level vector alignment and replaces it with an internal autoregressive self-reconstruction objective. The explicit reasoning sequence \(R_q = \{r_1, \dots, r_L\}\) is uniformly segmented into \(K\) chunks of length \(l_k\). Each continuous soft token \(t_k\) is repeated \(l_k\) times to form an expanded sequence \(T_{\text{exp}}\). To inject structured intra-step geometry, a 2D position embedding \(p_{k,i} = [\alpha_k; \beta_i]\) is constructed, where \(\alpha_k = \frac{k}{K} - 1\) denotes normalized global step position and \(\beta_i = \frac{i}{l_k} - 1\) denotes normalized intra-step token position. An MLP \(\phi_{\text{pos}}\) generates a position-dependent residual: $\(\tilde{t}_{\sum_{j=1}^{k-1} l_j + i} = t_k + \phi_{\text{pos}}([t_k; p_{k,i}])\)$ The LLM is then tasked with predicting the explicit reasoning tokens via next-token prediction: $\(\mathcal{L}_{\text{ntp}} = - \sum_{i=1}^L \log P(r_i \mid I, q, \tilde{t}_{<i}; \theta)\)$ This fine-grained self-reconstruction utilizes the LLM's intrinsic generation capability to anchor abstract latent tokens to concrete semantic reasoning steps, maintaining healthy inter-step distinction (pairwise cosine similarities remain balanced between 0.77 and 0.84) while imparting rich reasoning structure.

Loss & Training

SHIFT is trained end-to-end under a joint objective. For candidate documents, since static documents contain no complex instructional needs and require scalable offline indexing, they are encoded without reasoning directly from the final layer <EOS> token: \(v_d = f_\theta([E(d); \text{<EOS>}])\). Contrastive learning is conducted using the InfoNCE objective over query embedding \(v_q\) and document embedding \(v_d\): $\(\mathcal{L}_{\text{cl}} = -\log \frac{\exp(\text{sim}(v_q, v_{d^+})/\tau)}{\sum_{d \in B} \exp(\text{sim}(v_q, v_d)/\tau)}\)$ where \(\tau = 0.02\) is the temperature hyper-parameter, and batch \(B\) includes in-batch negatives alongside hard negatives mined from the training corpus. The total loss is the unweighted sum of contrastive and self-reconstruction objectives: $\(\mathcal{L} = \mathcal{L}_{\text{cl}} + \mathcal{L}_{\text{ntp}}\)$ Training is conducted on 4 A100 GPUs for 1 epoch using LoRA (\(r=64, \alpha=32\)), optimized by AdamW with a learning rate of \(1 \times 10^{-4}\), a warmup ratio of 0.1, and a global batch size of 64. Sequence lengths are capped at 1024 tokens during training and extended to 8192 during evaluation.

Key Experimental Results

Main Results

Evaluations are conducted on the reasoning-intensive in-domain benchmark Bright (12 subsets spanning StackExchange, LeetCode, and theorem proving, evaluated by nDCG@10), as well as out-of-domain benchmarks FollowIR (instruction following, evaluated by MAP@5, nDCG@5, and pairwise ranking p-MRR) and BrowseComp-Plus (deep research retrieval, evaluated by Recall@5, 100, 1000).

Table 1: Main Retrieval Performance on the Bright Benchmark (nDCG@10)

Model Family Model Size StackEx. (Bio/Earth/Rob) Coding (Leet/Pony) Theorem (TheoQ/TheoT) Overall Avg.
Standard Dense BGE-M3 0.6B 7.2 / 13.3 / 12.6 15.6 / 30.7 5.6 / 5.0 11.4
Standard Dense Qwen3-Embedding 0.6B 12.7 / 26.3 / 12.5 14.3 / 0.7 17.2 / 26.5 14.4
Standard Dense Qwen3-Embedding 8B 14.7 / 17.9 / 9.1 17.4 / 0.8 16.8 / 24.5 14.0
Basic Contrastive Fair Baseline (Qwen3) 0.6B 28.8 / 31.6 / 15.1 8.8 / 3.5 14.1 / 15.4 18.3
Basic Contrastive Fair Baseline (Qwen3) 8B 49.7 / 51.2 / 23.4 3.7 / 3.2 16.8 / 31.8 25.7
Basic Contrastive Fair Baseline (LLaMA3.1) 8B 57.7 / 40.9 / 20.1 4.3 / 5.7 12.0 / 17.7 22.5
Rewrite-then-Retrieve Pipeline (Qwen3) † 0.6B 57.8 / 51.6 / 15.0 0.6 / 14.6 14.9 / 15.5 22.4
Rewrite-then-Retrieve Pipeline (Qwen3) † 8B 53.1 / 54.3 / 20.5 3.2 / 15.2 17.4 / 38.8 28.1
Rewrite-then-Retrieve Pipeline (LLaMA3.1) † 8B 63.1 / 54.5 / 17.1 1.6 / 11.4 16.1 / 30.3 26.9
Explicit Reasoning Search-R3 (Qwen2.5) 1.5B 13.8 / 8.3 / 4.5 16.7 / 3.0 8.5 / 11.7 7.7
Implicit Reasoning GIRCSE (Qwen3) 0.6B 29.0 / 32.8 / 13.6 11.1 / 1.1 13.5 / 20.6 19.1
Implicit Reasoning GIRCSE (Qwen3) 8B 59.0 / 56.5 / 19.0 3.2 / 3.6 14.0 / 27.2 26.0
Implicit Reasoning LaSER (Qwen3) 0.6B 50.0 / 45.9 / 18.4 9.1 / 2.7 16.0 / 22.4 23.1
Implicit Reasoning LaSER (Qwen3) 8B 58.0 / 51.8 / 26.0 9.2 / 11.7 18.7 / 34.0 29.3
Implicit Reasoning LaSER (LLaMA3.1) 8B 58.4 / 48.1 / 17.0 1.7 / 5.9 14.6 / 19.2 24.4
Ours SHIFT (Qwen3-0.6B) 0.6B 49.8 / 48.7 / 20.5 12.3 / 5.3 21.8 / 23.6 25.0
Ours SHIFT (Qwen3-8B) 8B 54.8 / 49.4 / 26.7 14.2 / 14.1 26.6 / 35.6 31.0
Ours SHIFT (LLaMA3.1-8B) 8B 54.4 / 46.4 / 24.0 5.6 / 7.1 16.8 / 21.2 26.3

Table 2: Out-of-Distribution Benchmark Performance (FollowIR & BrowseComp-Plus)

Model Family & Variant Size FollowIR: Core17 (MAP@5 / p-MRR) FollowIR: News21 (nDCG@5 / p-MRR) FollowIR: Avg (Score / p-MRR) BrowseComp-Plus (R@5 / R@100 / R@1000)
Qwen3-Embedding-0.6B 0.6B 2.8 / 8.9 27.3 / 3.6 13.8 / 5.1 8.0 / 29.1 / 64.8
Qwen3-Embedding-8B 8B 3.1 / 6.2 25.5 / 8.8 12.9 / 7.2 7.7 / 31.6 / 61.3
Fair Baseline (Qwen3-0.6B) 0.6B 2.1 / 3.5 13.4 / -0.6 7.4 / -0.2 3.5 / 21.3 / 46.9
Fair Baseline (Qwen3-8B) 8B 2.8 / 4.4 18.9 / 2.0 11.0 / 1.7 11.3 / 37.4 / 63.2
GIRCSE (Qwen3-0.6B) 0.6B 3.0 / 3.9 12.8 / -0.9 7.2 / -1.6 6.9 / 25.2 / 52.8
GIRCSE (Qwen3-8B) 8B 3.0 / 4.2 22.6 / 0.7 11.4 / 2.0 13.0 / 40.8 / 68.1
LaSER (Qwen3-0.6B) 0.6B 2.0 / 3.4 25.0 / 1.0 11.5 / 0.0 6.8 / 26.8 / 54.9
LaSER (Qwen3-8B) 8B 4.1 / 5.8 21.8 / 0.2 12.5 / 2.4 11.7 / 38.4 / 66.9
SHIFT (Qwen3-0.6B) 0.6B 2.0 / 6.7 23.1 / 0.2 11.3 / 1.6 6.9 / 27.4 / 56.1
SHIFT (Qwen3-8B) 8B 4.3 / 6.4 24.6 / 0.4 13.9 / 2.5 13.2 / 40.6 / 66.8
SHIFT (LLaMA3.1-8B) 8B 4.1 / 2.5 23.0 / 0.6 12.6 / 0.5 6.8 / 26.9 / 52.7

Ablation Study

An extensive suite of ablations on Qwen3-0.6B assesses the utility of each component, aggregation strategy, reconstruction architecture, and document-side reasoning.

Table 3: Ablation Analysis of SHIFT (Qwen3-0.6B)

Variant / Configuration Bright (nDCG@10) FollowIR (Score) FollowIR (p-MRR) BC-Plus (R@5) BC-Plus (R@1000) Observation
SHIFT (Full Model) 25.0 11.3 1.6 6.9 56.1 Bidirectional MHA + Residual Proj. + NTP Self-Reconstruction
w/o Reconstruction 22.1 9.8 -6.7 5.5 48.6 Removing NTP loss causes severe representation collapse
w/o Residual Projection 22.8 10.1 -5.4 6.0 52.8 Retaining causal generative noise degrades retrieval metrics
w/o 2D Position 24.4 11.3 1.6 6.8 56.1 Lack of intra-step relative position hampers trajectory expansion
Retrieval with <EOS> 20.3 10.6 0.0 4.1 47.2 Static single-token retrieval wastes multi-step reasoning signals
Mean Pooling 24.3 11.1 4.5 6.8 52.9 Uniform weights cannot adaptively reweight task-critical steps
Last Pooling 20.9 9.7 0.8 4.7 51.0 Monotonic assumption that later steps are better fails severely
Causal MHA Pooling 21.8 9.6 0.8 5.1 48.9 Causal masking prevents backward cross-step refinement
GRU Pooling 15.6 5.8 -5.5 3.3 46.6 Recurrent gating introduces an information bottleneck and bias
BERT Reconstruction 21.7 9.8 -5.1 4.2 47.1 External 1-layer BERT decoder fails to bridge generative reasoning
Prefix Reconstruction 23.4 11.6 0.2 6.4 52.8 Prefix-tuning style slot fitting is less efficient than 2D expansion
w/ Document Reasoning 24.9 11.7 2.0 6.8 56.2 3-step implicit thinking on documents adds cost with no gain
Basic Contrastive Learning 18.3 7.4 -0.2 3.5 46.9 Standard contrastive fine-tuning without implicit reasoning

Key Findings

  1. Self-Reconstruction Is the Foundation of Effective Latent Reasoning: Omitting the autoregressive reconstruction loss drops Bright nDCG@10 from 25.0 to 22.1 and plunges FollowIR p-MRR from +1.6 to -6.7. Contrastive learning alone operates only on final pooled outputs, leaving intermediate latent steps semantically ungrounded. Autoregressive self-reconstruction actively anchors latent tokens to fine-grained logical trajectories.
  2. Prior Implicit Retrievers Suffer from Teacher-Induced Representation Collapse: Step-wise analysis reveals that in LaSER, the pairwise cosine similarities among its 3 latent steps are nearly 1.0, meaning each step performs identically in isolation. This stems from distilling teacher states that are already hyper-anisotropic (average similarity >0.95). In contrast, SHIFT maintains healthy inter-step cosine similarities (0.77–0.84), allowing each step to contribute distinct semantic facets and monotonically enhance retrieval across diverse tasks.
  3. Bidirectional Attention Disproves Monotonic Late-Step Priors: Methods assuming that the final step is uniformly superior (e.g., Last Pooling at 20.9 or Causal MHA at 21.8) are substantially outperformed by Bidirectional MHA (25.0). Cosine similarity distributions between pooled query embeddings and step representations (Figures 5 & 6) prove that optimal reasoning depth is task-dependent (e.g., Econ and TheoQ rely heavily on earlier vs. intermediate steps), which bidirectional self-attention captures dynamically.
  4. Document-Side Reasoning Is Wholly Superfluous: Executing 3 latent reasoning steps on documents yields virtually identical performance (24.9 vs. 25.0 on Bright) while tripling indexing latency. Queries possess goal-directed intent requiring reasoning disambiguation, whereas documents are static semantic targets best captured by lightweight <EOS> encoding.

Highlights & Insights

  • Harnessing Intrinsic Generation to Supervise Abstract Representations: Rather than designing auxiliary decoders or enforcing restrictive contrastive pulls, SHIFT ingeniously utilizes the causal LLM's natural next-token prediction capability to decode explicit reasoning paths from continuous latent embeddings via 2D relative position expansion.
  • Exposing Geometric Anisotropy in Explicit CoT Teachers: The work offers profound theoretical insight into why distilling explicit reasoning into latent retrievers previously stagnated: autoregressive teacher representations naturally suffer from severe anisotropy, meaning hard feature alignment simply clones contextual noise rather than reasoning logic.
  • Task-Adaptive Reasoning Synthesis via Bidirectional Pooling: Demonstrating that reasoning utility across intermediate steps is non-monotonic and domain-specific, SHIFT establishes bidirectional self-attention as a versatile mechanism for aggregating latent thoughts according to query complexity.

Limitations & Future Work

  • Dependence on Synthetic Explicit CoT Trajectories: The training pipeline relies on explicit reasoning paths generated by proprietary LLMs (e.g., GPT-4o-mini). Domain hallucinations or superficial reasoning styles present in the source dataset (ReasonEmbed) can transfer biases into the latent reasoning space.
  • Static Latent Budget (\(K=3\)): Fixing the latent reasoning budget to exactly 3 steps simplifies batch training but imposes an unnatural constraint across varying query difficulties. Future investigations should pursue adaptive early exiting or dynamic token allocation based on real-time query complexity.
  • Black-Box Interpretability of Latent States: Because continuous soft tokens operate entirely within latent space, inspecting and diagnosing intermediate reasoning errors remains difficult compared to explicit text rationales, necessitating the development of dedicated continuous probing diagnostics.
  • vs LaSER (Jin et al., 2026): LaSER applies dual-view distillation to align dense retriever latent states directly with the step-wise hidden states of an explicit CoT teacher, which inadvertently transfers the teacher's severe anisotropy and causes all latent tokens to collapse into near-identical vectors. SHIFT replaces cross-model vector alignment with fine-grained autoregressive self-reconstruction, preserving step diversity while integrating residual projection and bidirectional pooling.
  • vs GIRCSE (Tsai et al., 2026): GIRCSE iteratively refines soft embeddings via autoregressive contrastive decoding, but operates under a strict causal assumption that later tokens are always superior and lacks explicit reasoning supervision. SHIFT proves that bidirectional multi-head attention superiorly aggregates multi-step semantics, guided by fine-grained CoT reconstruction.
  • vs Rewrite-then-Retrieve Cascades (Query2doc, BGE-Reasoner): Cascaded pipelines require generating dozens or hundreds of discrete reasoning tokens during inference, incurring immense computational latency. SHIFT matches or exceeds 8B rewrite-then-retrieve pipelines on Bright (31.0 vs. 28.1) using only 3 latent forward steps, providing orders-of-magnitude faster inference.

Rating

  • Novelty: ⭐⭐⭐⭐☆ [Identifies representation and supervision mismatches in latent retrieval; introduces 2D position expansion and generative self-reconstruction]
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Evaluates on Bright (12 subsets), FollowIR, and BrowseComp-Plus, accompanied by step-wise geometric similarity and ablation analyses]
  • Writing Quality: ⭐⭐⭐⭐⭐ [Clear motivation, compelling narrative flow, well-crafted architectural figures, and rigorous empirical validation]
  • Value: ⭐⭐⭐⭐☆ [Offers an elegant, computationally efficient framework for operationalizing deep LLM reasoning in industrial-scale retrieval and RAG]