Existing approaches to training and evaluating search agents focus heavily on final-answer quality, largely neglecting how generated answers map back to retrieved evidence. Recent methods track citations to enforce grounding, but citation support alone does not establish source independence or whether an answer appropriately handles conflicting evidence. We argue that credibility in open-web retrieval depends on principled cross-source evidence grounding that accounts for source independence and systematically attributes and adjudicates conflicting evidence, while hierarchical judgments of these grounding decisions provide a unified signal for training and evaluation. To this end, we propose an evaluation framework, Cross-Source Evidence Grounding (CSEG), that employs a fine-grained, six-layer decision tree over the complete search trajectory. We introduce CSEG-Bench and trainable reward models to provide targeted supervision following this framework. Our approach helps eliminating the severe data skew of natural trajectories by providing targeted coverage over complex, deep-layer decision paths, such as cross-source conflict resolution and nuanced attribution, that standard rollouts fail to capture. Specifically, (1) we define the hierarchy at label, node, and path levels; (2) we construct 1,160 verified trajectories over nine canonical paths by rewriting answers on real retrieved document pools, with independent frontier judges reproducing the entire gold path and human experts deciding inclusion; (3) we train CSEG-RM, which reaches 83.9% exact accuracy on our benchmark while lying on the performance-token efficiency Pareto frontier. Experiments that use this reward alone to supervise a frozen-retrieval answer policy raise the independently judged mean label from 2.83 to 4.19, and show that learning the hierarchy generalizes across layers at a fraction of frontier-judge cost.
Figures & tables
Figure 1: Same answer, different evidence quality. All four trajectories answer 1945 and receive outcome reward 1, yet their evidence differs: (A) two citations from independent evidence roots; (B) three citations from the same evidence root; (C) conflicting evidence is retrieved but ignored; and (D) conflicting evidence is incorrectly cited as support. An outcome-only reward cannot distinguish these trajectories.
Representative works
Claim support
Retrieved pool
Root indep.
Conflict
Authority
Traj.
RM resource
GAIA, BrowseComp
✗
✗
✗
✗
✗
✗
✗
Search-R1, R1-Searcher, ReSearch
✗
▲ !
✗
✗
✗
▲ !
✗
ALCE, AttrScore, SelfCite
✓
✗
✗
✗
✗
▲ !
▲ !
RAGAS, ARES, RAGChecker
✓
▲ !
✗
▲ !
✗
▲ !
▲ !
ClashEval, WikiContradict, RAMDocs, et al.
▲ !
✓
✗
✓
▲ !
▲ !
✗
DRAGged into Conflicts
✓
✓
✗
✓
▲ !
✓
✗
Table 1: Construct-level comparison of grounding evaluation and reward resources. ✓ : explicit; ▲ ! : partial; ✗ : not targeted. Symbols indicate task coverage, not latent capability. Appendix A defines the columns and discusses each family.
Figure 2: CSEG hierarchy and benchmark construction. (a) Rectangles denote layers and red capsules denote terminal outcomes. Red arrows indicate failures (labels 1–6); either all-pass outcome yields label 7. (b) Evidence scenario construction, answer rewriting on frozen pools, and full-path validation yield 1,160 trajectories. Roots and paths are schematic.
Layer
Aspect
Criterion
L1
Citation presence
A core claim cites an existing, topically relevant document.
L2
Multiple sources
At least two distinct cited sources support the core claim.
L3
Source independence
Cited support traces to at least two independent evidence roots.
Aggregation has appropriate scope and avoids over- or under-claiming.
L5-X
Conflict attribution
The acknowledged disagreement is attributed at source level.
Table 2: The CSEG hierarchy. Labels encode the depth of the first failed decision.
Label
L1_ fail
L2_ fail
L3_ fail
L4_conflict _missed
L5_consensus _fail
L5_conflict _fail
L6_fail_ conflict
all_pass_ conflict
all_pass_ consensus
n
78
79
184
140
105
151
181
128
114
Table 3: Composition of CSEG-Bench ( n=1,160 ). n counts trajectories for each terminal path.
Model
Total
Think
Exact
Macro R.
MAE ↓
QWK ↑
Path ED ↓
Tokens
Qwen3-4B
4.0B
high
0.3828
0.3934
1.5560
0.3590
1.7612
3,725
Qwen3-8B
8.2B
high
0.4233
0.4543
1.3793
0.4809
1.5767
4,578
GPT-OSS-20B
20.9B
low
0.4940
0.5454
1.1845
0.5035
1.5078
1,156
GPT-OSS-20B
20.9B
medium
0.5293
0.5919
1.2517
0.4656
1.5690
3,763
GPT-OSS-20B
20.9B
high
0.4164
0.4998
1.9655
0.2376
2.1121
23,860
Muse-Glimmer-30B
30B
high
0.4595
0.5187
1.8129
0.3035
2.2440
5,175
Table 4: Benchmark results on CSEG-Bench ( n=1,160 ). “Total” is the overall parameter count; “Tokens” is mean completion tokens per judgment. Bold marks the best value per block. † is also the rejection-sampling SFT teacher.
Figure 3: Thinking effort, training, and accuracy–cost trade-offs on CSEG-Bench ( n=1,160 ). (a) Selected zero-shot models improve (top) or regress (bottom) at higher thinking effort; the two plots use different vertical scales. (b) Qwen3-4B/8B accuracy across training routes. (c) Accuracy versus mean completion tokens; the dashed staircase is the Pareto frontier and gray points are other zero-shot configurations. Dark/light blue mark 4B/8B CSEG-RM; squares, triangles, and stars denote SFT, GRPO, and SFT+GRPO.
Quantity
Step 0 (Qwen3-8B base)
Step 200 (SFT+GRPO-400 RM)
Mean label
2.834
4.193
Median label
3
4
Label histogram, 1–7
186/42/123/63/53/20/25
20/4/172/110/123/33/50
P(label≥4)
0.314
0.617
P(label≥5)
0.191
0.402
P(label=7)
0.049
0.098
Table 5: Answer policy before and after CSEG-supervised optimization, on 512 frozen-retrieval questions. Scores come from an independent frontier judge running the same six-layer pipeline. The reward model is Qwen3-4B SFT+GRPO-400 (grader exact 0.8388).
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Terminal type
n
Share
Depth
L1 fail (no valid citation)
23
9.0%
shallow
L2 fail (single source)
4
1.6%
shallow
L3 fail (multiple URLs, one evidence root)
85
33.2%
shallow
L4 conflict missed
32
12.5%
middle
L5-X fail (conflict attribution)
16
6.3%
deep
L5-C fail (consensus aggregation)
9
3.5%
deep
Appendix
Table 6: Gold distribution over 256 production trajectories, every one of which passes the service’s closed-QA outcome verifier. These are the original audit counts, before consensus paths bypassed L6.
Policy
Mean label
Fail by L3
Production agent (Table 6 )
4.48
43.8%
DeepSeek-V4-Flash (low)
4.11
52.4%
Qwen3-8B (thinking)
2.89
70.1%
GPT-OSS-120B (low)
2.37
84.2%
Appendix
Table 7: Open-weight policies rolled out on the same 256 frozen (q,D) prefixes as Table 6 , rewriting only the final answer. The production row is the outcome-passing gold audit (mean label computed from that distribution); the other rows are not outcome-filtered and do not replace the 43.8% / 33.2% claims above.
Figure 4: A concrete CSEG judgment: inputs (left), the six layer decisions with case-specific rationales (center), and the resulting path, node verdicts, terminal label, and path metrics (right).
Concept
Related expressions
Usage and distinction
Reasoning configuration
Thinking effort; reasoning effort; deliberation
Effort denotes a model-specific setting. Thinking mode indicates whether thinking is enabled; deliberation describes the generated reasoning process. Matching setting names do not establish equal reasoning budgets across models.
Label prediction
Exact accuracy; terminal accuracy; accuracy
In benchmark label results, these refer to exact label accuracy: agreement between predicted and gold terminal labels. Exact whole-path match instead requires the entire decision path to agree.
Evaluation role
Judge; scorer; grader; reward model (RM)
A judge makes evaluation decisions; the scorer composes layer decisions through routing. An RM supplies a training reward. Grader is contextual shorthand for the scoring model, not a separate model family; an independent evaluation judge is distinct from the RM used for policy training.
Efficiency measure
Completion tokens; token efficiency; cost
Completion-token count is the reported cost proxy. Accuracy–token comparisons do not by themselves measure latency, monetary cost, or compute across model sizes.
Hierarchy structure
Layer; node
There are six layers and seven node classifiers because layer L5 has alternative L5-C and L5-X nodes. The node prompts use the same layer terminology. Terminal label 7 means that every visited layer passed: five on the consensus branch or six on the conflict branch. No seventh layer is evaluated.
Evidence provenance
Document; source; evidence root
A document is a retrieved item; cited source refers to a document cited as evidence. Its evidence root is the origin of its support for the core claim. Different documents or websites need not provide independent evidence roots.
Appendix
Table 8: Terminology guide. Related expressions are not interchangeable unless explicitly identified as shorthand; model-specific settings retain their original names.
Type
Label
Required terminal behavior
Retrieved-pool requirement
L1_fail
1
L1 fail
—
L2_fail
2
L1 pass; L2 fail
independent
L3_fail
3
L1–L2 pass; L3 fail
redundant
L4_conflict_missed
4
conflict exists but is missed
independent, conflict
L5_consensus_fail
5
consistent pool; aggregation fails
independent, no conflict
L5_conflict_fail
5
conflict identified; attribution fails
independent, conflict
Appendix
Table 9: The nine canonical paths. “Independent,” “redundant,” and “conflict” are defined relative to the core claim.
train balanced to 488 per (layer, verdict) cell; dev natural
Data-source ablation arm
7,320 / 911
Qwen3-8B rollouts on the same seeds, truncated to the same rows
CSEG-Bench
1,160 trajectories
nine paths; 692 two-judge plus 468 single-judge; 5,166 layer decisions; final reporting only
Production audit (App. B )
256 trajectories
natural prevalence; motivates synthesis (RQ1)
Frozen-prefix policy rollouts
256 prefixes
same (q,D) as App. B ; answer only; Table 7
Appendix
Table 10: Resources and split discipline. Benchmark, reward-model training, and answer-policy training rest on pairwise disjoint base pools. Layer examples are per-layer flattenings of a trajectory; a trajectory stopping at layer N contributes N of them.
Model
L1
L2
L3
L4
L5-C
L5-X
L6
(1,160)
(1,082)
(1,003)
(819)
(219)
(460)
(423)
Qwen3-8B SFT
0.520
0.935
0.823
0.753
0.466
0.410
0.751
Qwen3-4B SFT+GRPO
0.863
0.965
0.888
0.847
0.709
0.982
0.924
Qwen3-8B SFT+GRPO
0.583
0.973
0.897
0.856
0.766
0.937
0.890
Appendix
Table 11: Conditional macro F1 per layer on CSEG-Bench, with gold reach n in the header. A node is scored only on examples whose gold path reaches it, and a prediction that stopped earlier counts as “not reached.”
Training data
Model
Exact
Macro R.
MAE ↓
QWK ↑
Path ED ↓
Constructed, balanced
Qwen3-4B
0.7353
0.7615
0.6672
0.7052
0.8466
Qwen3-8B natural rollouts
Qwen3-4B
0.6095
0.6410
0.8284
0.6641
1.0578
Constructed, balanced
Qwen3-8B
0.7672
0.7957
0.5776
0.7358
0.7026
Qwen3-8B natural rollouts
Qwen3-8B
0.5655
0.6260
0.9181
0.6386
1.1543
Appendix
Table 12: Training-data source, GRPO from the base model for 400 steps with every other setting byte-identical. Both arms hold 7,320 training and 911 development layer examples built on the same seed questions.
Figure 5: Step-200 high band (labels 5–7) on the same 512 prefixes. A weaker GRPO-only grader also moves the policy, but yields a proportion of trajectories with labels 5–7 about 10 percentage relatively lower than the official SFT+GRPO-400 reward.
Figure 6: Training dynamics of the official Qwen3-4B SFT+GRPO-400 reward model. (a) Supervised distillation, one epoch. (b–f) GRPO against the verifiable layer verdict. The validation markers are the in-loop split, not CSEG-Bench; reported RM quality remains Table 4 .
Figure 7: Training reward assigned by the frozen Qwen3-4B SFT+GRPO-400 CSEG-RM while optimizing the Qwen3-8B answer policy. This curve shows the policy fitting the reward model. Predictive validity is the independent-judge comparison in Table 5 , not this figure.
Node
Layer
Verdict
Next
L1
Citation presence
pass
L2
L1
Citation presence
fail
STOP (label 1)
L2
Multiple sources
pass
L3
L2
Multiple sources
fail
STOP (label 2)
L3
Source independence
pass
L4
L3
Source independence
fail
STOP (label 3)
Appendix
Table 13: Routing of each node verdict.
Slot
Meaning
Use
{user_question}
The user original question/query
Filled from the trajectory at inference.
{retrieved_sources}
Formatted list of retrieved web documents, including source URL, site, authority level, publish time, and content excerpt
Filled from the trajectory at inference.
{final_answer}
The agent final answer text (excluding tool call sections)
Filled from the trajectory at inference.
{agent_search_queries}
The search queries issued by the agent
Declared in the spec; unused by the current node templates.
Appendix
Table 14: Input slots interpolated into every node prompt.
Label
Meaning
1
No citation support: core claims lack any relevant source backing
2
Single-source: claim supported by only one source, no multi-source verification
3
Redundant multi-source: multiple sources cited but all trace to the same origin
4
Conflict blindness: sources contain real conflicts but agent completely missed them (source blending)
5
Aggregation/attribution defect: consistency/conflict detected, but aggregation introduced framing errors or conflict attribution was insufficient
6
Deficient adjudication: reached L6 but adjudication was unsupported, overconfident, or incorrectly weighted
Search-augmented language agents should retrieve external information only when necessary and ground their answers in retrieved evidence. Yet terminal correctness assigns equal credit to direct answers from parametric knowledge, evidence-grounded answers, and answers recalled after redundant search. Process supervision and LLM judges provide richer feedback but can require high-cost annotation or inference; low-cost, non-contrastive policy confidence signals do not directly identify whether retrieved evidence changes an answer. We propose Search-G1, a reward framework that couples retrieval necessity with evidence reliance through two intervention-calibrated hidden-state readouts for selective grounded search. A prompt readout predicts closed-book sufficiency, whose complement proxies policy-relative retrieval necessity; an answer-commit readout predicts evidence reliance as sensitivity to evidence deletion conditional on the realized reasoning prefix. A correctness-first reward gates evidence-reliance credit by retrieval necessity, rewards sufficient direct answers, and penalizes repeated search. After calibration, scoring needs neither process annotations nor LLM-judge inference. As policy-relative targets, trajectory distributions, and representations may drift, Search-G1 periodically regenerates targets and refits its readouts. Across four QA benchmarks and two scales, Search-G1 achieves competitive task accuracy, with stronger counterfactual evidence dependence and fewer searches in the main comparisons, complemented by independent human audits of answer support.
Ruoxi Cheng, Haoxuan Ma, Hongyi Zhang +7
Fudan University · Tencent · Nanjing University +2
Outcome-reward reinforcement learning such as GRPO gives every token of a search-agent rollout the same advantage, although the rollout exercises two capabilities: finding evidence and using it. We show that accuracy improves without grounding: Search-R1 raises accuracy on all seven QA benchmarks, yet on Bamboogle the evidence-supported share of correct answers falls from 84.0% to 42.9%. We attribute this to GRPO's credit assignment: rollouts that retrieve answer-bearing evidence are favoured in aggregate only in proportion to their accuracy gain over the rest, a grounding gap that answering from memory or misusing evidence shrinks or reverses. To address this, we propose Grounded Credit Assignment (GCA), which uses verifiable sub-outcomes (evidence retrieval and answer correctness) to correct search and answer credit separately: a wrong answer is blamed more than the search that found the evidence, and a correct answer that bypasses retrievable evidence is credited less. Never negative on evidence-finding search, the correction does not shrink with the grounding gap; capped at 10% of each token's advantage, it never overrides the outcome signal. Without extra models, rollouts, or training stages, GCA improves average EM over Search-R1 by 5.7 (Qwen2.5-3B) and 6.3 points (Qwen3-30B-A3B) and raises this share from 76.9% to 93.4%.
Qirui Liu, Hao Chen, Yi Wen +7
South China University of Technology · FiT, Tencent · City University of Hong Kong +5
Search agents can use cross-document graphs to discover sources through multi-hop connections and choose subsequent actions based on candidate rankings and relationships within a limited observation window. However, documents may rank highly because their numerous connections make them easier to reach, without necessarily providing more valuable evidence. Graph diffusion methods such as Personalized PageRank (PPR) aggregate propagation contributions from multiple sources, allowing highly connected distractor nodes to crowd the observation window. Flat node scores also obscure cross-document evidence trails, limiting continued exploration. We propose Contrastive Evidence Paths (CEP), which scores candidates by their strongest individual paths on a query-conditioned document graph and calibrates these scores against the same graph's background visitation distribution to mitigate structural bias. CEP ranks candidates by the calibrated scores and retains the corresponding source paths. Building on CEP, we develop Structured Corpus Interaction (SCI), which organizes candidates and their connecting paths into a structured Map, enabling agents to verify source documents and pursue missing information through expansion from anchor documents. Experiments on BrowseComp-Plus and TREC-COVID show that SCI achieves state-of-the-art answer accuracy and nDCG@10 with the lowest generation cost. Compared with the strongest baselines by task performance, maximum reductions in tool calls and total generation costs exceed 50% and 40%, respectively. Further analysis confirms that SCI with CEP suppresses ranking inflation induced by structural connections and improves gold recall and exploration efficiency.
Junchen Li, Liang Xu, Rongzheng Wang +3
University of Electronic Science and Technology of China Chengdu, China