TraceRetriever: Trace-Centric Retrieval Approach

• TraceRetriever is a trace-centric retrieval design pattern that uses explicit intermediate structures like token influence, agent trajectories, and reasoning chains to enhance query processing.
• It applies techniques such as gradient-based token attribution and trajectory supervision to diagnose and secure retrieval outcomes in diverse applications including RAG and legal precedent search.
• The approach improves performance metrics and explains retrieval decisions, offering actionable diagnostics for robust retriever learning and evidence tracing.

TraceRetriever denotes retrieval systems that do not treat retrieval as a single similarity lookup, but instead operate over a trace: token influence attributions, agent interaction trajectories, explicit reasoning chains, or rhetorically salient document segments. In recent work the term is used directly for legal precedent retrieval under partial case knowledge, and is also invoked to describe systems that trace answer provenance in retrieval-augmented generation (RAG), learn retrievers from agent trajectories, and compress Chain-of-Thought into retrieval embeddings (Nigam et al., 1 Aug 2025, Chen et al., 24 Jun 2026, Zhou et al., 30 Mar 2026, Hao et al., 3 Mar 2026). This suggests that TraceRetriever is best understood as a trace-centric retrieval design pattern rather than a single fixed architecture.

1. Conceptual scope

Across the recent literature, “trace” refers to different intermediate structures that mediate retrieval. In poisoning analysis, the trace is causal influence from retrieved tokens to an answer string. In agentic search, it is the sequence of search, browse, and reasoning actions. In multimodal retrieval, it is an explicit Chain-of-Thought that is later compressed into an embedding. In legal retrieval, it is the rhetorically salient segmentation of a judgment into facts, issues, arguments, and reasoning (Chen et al., 24 Jun 2026, Zhou et al., 30 Mar 2026, Hao et al., 3 Mar 2026, Nigam et al., 1 Aug 2025).

Paradigm	Trace object	Representative use
Provenance tracing in RAG	High-influence retrieved tokens and phrases	Poisoning detection and target-answer recovery (Chen et al., 24 Jun 2026)
Agentic retrieval learning	Search–browse–reason trajectories	Retriever supervision and query refinement (Zhou et al., 30 Mar 2026)
Reason-then-embed retrieval	Generated CoT followed by `<	emb
Role-aware legal retrieval	Rhetorical segments of judgments	Precedent retrieval from partial case knowledge (Nigam et al., 1 Aug 2025)

The common denominator is that retrieval quality is improved, explained, or diagnosed by making an intermediate structure explicit. Rather than assuming that the query alone is the correct retrieval interface, TraceRetriever systems expose what the model attended to, what the agent did, or what evidence chain it constructed.

2. Token-level provenance and corpus-poisoning detection

A security-oriented interpretation of TraceRetriever appears in the token-influence framework TRACE for poisoned RAG corpora. The setting is standard RAG: a retriever such as Contriever retrieves top-$k$ documents, and an LLM $M_\theta$ answers conditioned on the question and those passages. The threat model is corpus poisoning, exemplified by PoisonedRAG, where an attacker injects multiple passages that consistently assert a target answer. TRACE turns the retrieved set into a traceback object by scoring each token embedding $x_i$ with

$$Influence(x_i)=\left\|\nabla_{x_i}\log P_\theta(y\mid x_{1:n})\right\|_2,$$

using the affirmation string $y=\text{"Yes, the answer is:"}$ in Phase 1 and candidate phrases $p$ in Phase 2 (Chen et al., 24 Jun 2026).

The pipeline has two stages. Phase 1 computes per-document token influence, removes punctuation, frequent function words, and tokens already in the question, then keeps the top-$k_1$ influential tokens, merges contiguous indices into phrases, and retains phrases recurring in at least $a_1$ retrieved documents. Phase 2 re-runs attribution with each candidate phrase $p$ as the prediction target, keeps top-$k_2$ influential tokens that belong to $M_\theta$0, and retains tokens recurring across at least $M_\theta$1 documents. The final target token set $M_\theta$2 serves both as an explanation of what drives the prediction and as a poisoning flag: $M_\theta$3 implies detection.

This design is explicitly lightweight: it uses gradients of the target LLM itself, introduces no auxiliary classifier, and requires no external verifier. In experiments on Natural Questions, HotpotQA, and MS-MARCO, across Llama‑3.1‑8B‑Instruct, Gemma‑3‑4b‑it, Qwen‑3.5‑4B, Vicuna‑7B‑v1.5, Mistral‑7B‑Instruct‑v0.3, and Phi‑4‑mini‑instruct, TRACE achieves true positive rates consistently above 90% on almost all LLMs and datasets under high-sensitivity settings, with Phi‑4‑mini reaching an average TPR of 98.87% across the three datasets. Under stricter thresholds, false positive rate can be kept below 10%, and under balanced settings TPR remains above 80% overall. For target-answer recovery, detection accuracy is about 90% on NQ and MS‑MARCO and above 70% on HotpotQA; runtime is about 1–9 seconds per query versus about 6–17 seconds for RAGForensics (Chen et al., 24 Jun 2026).

The significance of this line of work is that TraceRetriever is not merely retrieving documents; it is tracing which retrieved tokens causally support an answer string and using that trace for both security and interpretability. The same machinery also exposes a limitation: the method assumes white-box gradient access and depends on recurrence across multiple poisoned documents.

3. Learning retrieval from agent trajectories

A second major meaning of TraceRetriever is retrieval learned from agent traces rather than from human click logs. In LRAT, a deep-research agent generates trajectories

$M_\theta$4

where $M_\theta$5 is the reasoning state, $M_\theta$6 is an action such as [Search] or [Browse], and $M_\theta$7 is the observation. LRAT identifies three supervision signals: browsing actions, unbrowsed rejections, and post-browse reasoning traces. Browsed documents become naive positives, unbrowsed documents in the same candidate set become negatives, and post-browse reasoning is used both to filter positives with a Qwen3-30B-A3B-Thinking-2507 judge and to assign relevance intensity weights through weighted InfoNCE. The filter retains 97.2% of ground-truth evidence documents and 74.8% of browsed non-evidence documents. Empirically, WebExplore‑8B with Qwen3‑Embedding‑0.6B improves evidence recall from 47.7 to 55.9, Tongyi‑DeepResearch improves in-domain success rate from 52.7 to 68.0, and average steps drop from 26.7 to 20.7 (Zhou et al., 30 Mar 2026).

Agentic-R formulates the same problem more explicitly as multi-turn utility estimation. A trajectory is

$M_\theta$8

with reasoning traces $M_\theta$9, intermediate queries $x_i$0, retrieved passages $x_i$1, and final answer $x_i$2. Passage utility is defined by combining local relevance scores $x_i$3, assigned by Qwen2.5‑72B‑Instruct, with global answer correctness $x_i$4, obtained by continuing the agent’s trajectory with each candidate passage and checking exact match against the gold answer. The retriever is then trained on positives satisfying $x_i$5 and $x_i$6, with query input $x_i$7, and improved iteratively together with the agent. Across seven QA benchmarks, Agentic‑R reaches average EM 45.00 with its own search agent, compared with 41.78 for REPLUG (Liu et al., 17 Jan 2026).

Critic-R closes the loop further by using the agent’s introspective reasoning trace after retrieval as a first-class supervision source. A critic model reads the global question, current sub-query, retrieved documents, and the agent’s introspective trace, then decides whether the documents are sufficient. Critic-R-Zero performs inference-time query refinement by producing a new instruction and query when the critic returns “no”; Critic-Embed converts successful and failed refinement trajectories into retriever supervision, without manual relevance annotation, using an InfoNCE objective over positives and intra-trajectory hard negatives. On the multi-hop average at top-1, Stella‑400M scores 0.3472/0.4470 EM/F1, Agentic‑R scores 0.3670/0.4564, Critic‑Embed scores 0.3794/0.4806, and the full Critic‑R system reaches 0.3957/0.4959. Removing introspective feedback lowers Critic‑Embed to 0.3614/0.4521, indicating that the reasoning trace is a materially useful retrieval signal (Alam et al., 30 May 2026).

These systems make explicit that TraceRetriever in agentic search is not only a retriever but also a supervision extractor. Browsing actions, failed search branches, post-browse reasoning, and introspective critiques become retrieval labels. A plausible implication is that trace collection can serve as a scalable alternative to human relevance annotation, provided that the agent and any LLM-as-judge are sufficiently reliable.

4. Reasoning traces as retrieval representations and evidence chains

A third interpretation of TraceRetriever uses an explicit reasoning trace as the retrieval representation itself. In the multimodal TRACE framework, the query is first processed autoregressively to generate either an empty trace or a Chain-of-Thought $x_i$8, followed by a special token: $x_i$9 The hidden state immediately before \<|emb|\> becomes the query embedding. Candidate embeddings are produced without CoT, which makes the design deliberately asymmetric. The training objective combines generative cross-entropy over the trace with retrieval contrastive loss, and the routing mechanism is implicit: if \<|emb|\> is the most probable first output token, the model skips reasoning; otherwise it generates CoT first. On the M‑BEIR benchmark, TRACE improves average performance from 23.0% for base Qwen2.5‑VL‑7B used as a generic encoder to 58.8%, exceeds LamRA‑Ret by 2.2 R@K points overall, gains 4.2 R@5 on CIRR, and shows the intended speed–accuracy trade-off: on MSCOCO it reaches 8.25 QPS with R@5 89.10, while on CIRR it reaches 6.48 QPS with R@5 57.03 (Hao et al., 3 Mar 2026).

A text-only evidence-construction version of the same idea appears in TRACE for multi-hop QA. Given noisy retrieved documents $$Influence(x_i)=\left\|\nabla_{x_i}\log P_\theta(y\mid x_{1:n})\right\|_2,$$0, the method constructs a knowledge graph $$Influence(x_i)=\left\|\nabla_{x_i}\log P_\theta(y\mid x_{1:n})\right\|_2,$$1 of triples extracted from each document title and text, then autoregressively selects reasoning chains $$Influence(x_i)=\left\|\nabla_{x_i}\log P_\theta(y\mid x_{1:n})\right\|_2,$$2 of triples according to

$$Influence(x_i)=\left\|\nabla_{x_i}\log P_\theta(y\mid x_{1:n})\right\|_2,$$3

The chain constructor uses a triple ranker based on E5‑Mistral and a triple selector based on LLaMA3‑8B‑Instruct, with beam search and adaptive termination. TRACE‑Triple feeds only the selected reasoning chains to the answer generator; TRACE‑Doc uses the chains to vote for source documents and then feeds the selected documents. On HotpotQA, 2WikiMultiHopQA, and MuSiQue, the method improves EM by up to 14.03 percentage points relative to using all retrieved documents, and the reasoning-chain context is often sufficient despite being much shorter than the full retrieved context (Fang et al., 2024).

Both systems instantiate a “reason first, retrieve or answer later” logic. In one case the trace is compressed into an embedding; in the other it is preserved as an explicit chain of knowledge triples. The technical commonality is that retrieval is conditioned on a structured intermediate representation rather than directly on the raw query.

5. Role-aware legal precedent retrieval

The legal system named TraceRetriever addresses a different but closely related problem: realistic precedent retrieval under partial case knowledge. The motivating observation is that lawyers rarely begin from a complete judgment; they typically know the facts, issues, and perhaps some argumentation, but not the final reasoning or decision. TraceRetriever therefore segments judgments into rhetorical roles and retrieves using only selected roles rather than the full document (Nigam et al., 1 Aug 2025).

The segmentation stage uses a Hierarchical BiLSTM–CRF rhetorical role classifier trained on Indian judgments. Sentences are segmented with spaCy, encoded using a domain-specific sent2vec model trained on Indian Supreme Court judgments, contextualized with a BiLSTM, and labeled with a linear-chain CRF. Query construction then concatenates only chosen roles such as Facts, Issue, Arguments, Reasoning, or Decision. Retrieval is hybrid. A dense index is built in Milvus using Snowflake/snowflake-arctic-embed-m-v2.0 with 768-dimensional vectors and an IVF-FLAT index with nlist = 2048; BM25 scores are computed over the dense candidate pool; rankings are combined with Reciprocal Rank Fusion; and the fused candidates are re-ranked with the cross-encoder BAAI/bge-reranker-v2-m3. Long documents are chunked for re-ranking, and chunk scores are aggregated by weighted average.

On IL-PCR, the strongest role configuration is Facts+Issue+Reasoning. With Vector DB retrieval it yields P@k 0.1843 at $$Influence(x_i)=\left\|\nabla_{x_i}\log P_\theta(y\mid x_{1:n})\right\|_2,$$4, R@k 0.2088, F1@k 0.1636, MAP 0.3783, and MRR 0.3924. Facts-only queries are substantially weaker, and Facts+Issue+Decision is worse than Facts+Issue+Reasoning, indicating that reasoning is more discriminative than the decision section. On COLIEE 2025, performance is lower overall—BM25 MAP 0.1410, Vector DB MAP 0.1695, Cross-encoder MAP 0.1574—which the paper attributes to cross-jurisdiction transfer and weaker rhetorical segmentation outside the Indian domain (Nigam et al., 1 Aug 2025).

This legal instantiation makes the TraceRetriever idea especially concrete. The trace is not a model-internal attribution or an agent trajectory; it is a rhetorically structured abstraction of the case itself. Retrieval quality improves because the query is aligned with how precedent search is actually performed.

6. Diagnostics, evaluation, and limitations

TraceRetriever-style systems require evaluation tools that expose retrieval behavior rather than only end-task accuracy. RAGTrace provides such infrastructure for RAG workflows. It logs query, retrieved chunks, generated answer, and evaluation metadata; computes composite metrics such as BLEU, ROUGE, GPTScore, Question Correctness, and Topic Relevance; and adds granular diagnostics including Retrieval Failure Value, Prompt Fragility Value, Generation Anomaly Value, and Standard Anomaly Value. Its Chunk‑Relink Graph compares chunks across original, optimized, and “standard” retrieval strategies, while its answer view links answer spans to supporting chunks. The Retrieval Failure Value combines hit rate on gold chunks with entropy over similarity distributions, and the Prompt Fragility Value measures semantic consistency of retrieved chunks across paraphrased prompts (Cheng et al., 8 Aug 2025).

For agentic search, AgenticRAGTracer supplies hop-aware supervision missing from conventional QA benchmarks. It contains 1,305 data points, distinguishes inference and comparison topologies, and exposes explicit atomic questions, answers, and evidence for each hop. On the hardest portion, GPT‑5 attains only 22.6% EM. The benchmark’s main diagnostic finding is that failures are often distorted reasoning chains that either collapse prematurely or wander into over-extension, which makes it directly relevant to TraceRetriever systems that try to trace or repair multi-step retrieval reasoning (You et al., 22 Feb 2026).

The limitations reported across the literature are consistent. Token-influence tracing in poisoned RAG requires white-box gradient access, depends on recurrence across multiple poisoned documents, and is currently validated only on generative QA (Chen et al., 24 Jun 2026). Trajectory-based retriever learning depends on the quality of the initial agent, uses LLM-as-judge filtering, and approximates relevance intensity by reasoning length, which is only a coarse proxy (Zhou et al., 30 Mar 2026). Multimodal reasoning-trace retrieval introduces autoregressive overhead and depends on the quality of teacher-generated CoT traces (Hao et al., 3 Mar 2026). Legal TraceRetriever depends on a rhetorical role classifier trained on Indian judgments, which transfers imperfectly to COLIEE (Nigam et al., 1 Aug 2025).

Taken together, these works indicate that TraceRetriever is becoming a general retrieval principle: retrieve with explicit traces, retrieve from traces, or retrieve the traces themselves. A plausible implication is that future systems will combine several of these forms at once—for example, agent-trajectory supervision, reasoning-trace embeddings, and provenance diagnostics within the same retrieval stack—so that retrieval, explanation, and robustness are optimized jointly rather than treated as separate subsystems.