From Evidence to Judgment: Trace-Augmented Reasoning under Conflicting Evidence
Organizations: Birla Institute of Technology and Science, Pilani · Carnegie Mellon University, Pittsburgh
Abstract
Retrieved documents often disagree, contain outdated information, or fail to support any answer, and retrieval-augmented generation (RAG) systems have no reliable way to decide which documents to trust. We treat this as a reasoning problem and test it separately from answer generation. We release 943 training and validation examples and a disjoint 736-example benchmark, and fine-tune four language models to write a structured think trace before answering: the model judges each document, infers the conflict pattern, and writes a grounded answer or a justified refusal. Our evaluation framework, CATS (Conflict-Aware Trust Score), scores grounded refusal, behavioural adherence, factual grounding, and single-truth recall; a human study validates factual grounding and single-truth recall (kappa=0.78 and 0.79). The think trace improves the answer/refuse decision: under all three prompts, untuned models score at most 0.429 Refusal F1 (F1 with refusal as the positive class) and trace-tuned models at least 0.826. On the answerable examples a model answers, factual grounding under the 40-word prompt is at most 0.11 for untuned and 0.86-0.88 for trace-tuned models. Cutting the prompt from 1,965 words to 40 leaves trace-tuned scores within the benchmark's measurement noise (+/-0.034). Answer-only fine-tuning trails trace supervision on every metric under the minimal prompt.
Figures & tables
| Train+val | Benchmark | |||
| Conflict type | % | % | ||
| No conflict | 343 | 36.4 | 211 | 28.7 |
| Complementary | 286 | 30.3 | 221 | 30.0 |
| Conflicting opinions | 158 | 16.8 | 109 | 14.8 |
| Outdated | 136 | 14.4 | 158 | 21.5 |
| Misinformation | 20 | 2.1 | 37 | 5.0 |
| Refusal F1 | Behaviour Adherence | Factual Grounding | Single-Truth Recall | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Model | Prompt | base | sft | base | sft | base | sft | base | sft |
| Qwen2.5-7B | strict | 0.429 | 0.911 | 0.811 | 0.780 | 0.617 | 0.838 | 0.561 | 0.687 |
| minimal | 0.031 | 0.881 | 0.737 | 0.768 | 0.031 | 0.837 | 0.757 | 0.686 | |
| Qwen2.5-32B | strict | 0.367 | 0.905 | 0.789 | 0.821 | 0.851 | 0.812 | 0.733 | 0.749 |
| minimal | 0.015 | 0.911 | 0.747 | 0.809 | 0.108 | 0.822 | 0.773 | 0.738 | |
| Llama-3.1-8B | strict | 0.255 | 0.926 | 0.721 | 0.805 | 0.705 | 0.852 | 0.634 | 0.727 |
| Model | Method | Ref. F1 | Behaviour Adherence | Factual Grounding | Single-Truth Recall |
|---|---|---|---|---|---|
| Llama-3.1-8B | CoN (prompt-only) | 0.000 | 0.632 | 0.609 | 0.658 |
| CoT (few-shot) | 0.302 | 0.716 | 0.635 | 0.678 | |
| trace-sft (ours) | 0.926 | 0.805 | 0.852 | 0.727 | |
| Mistral-7B | CoN (prompt-only) | 0.000 | 0.685 | 0.522 | 0.676 |
| CoT (few-shot) | 0.525 | 0.657 | 0.417 | 0.528 | |
| trace-sft (ours) | 0.826 | 0.793 | 0.796 | 0.644 |
Appendix figures & tables39 assets
Supplementary material from the paper’s appendix.
Appendix
| Source | Track | NC | CI | CO | OI | MI | |
|---|---|---|---|---|---|---|---|
| QACC | web-retr. | 145 | 69 | 46 | 14 | 7 | 9 |
| SituatedQA | web-retr. | 141 | 43 | 50 | 6 | 32 | 10 |
| ConflictingQA | web-retr. | 137 | 8 | 53 | 73 | 1 | 2 |
| TRUST-ALIGN | refusal | 128 | 57 | 45 | 13 | 13 | 0 |
| FreshQA | web-retr. | 85 | 33 | 8 | 2 | 27 | 15 |
| WikiRevision | curated | 78 | 0 | 0 | 0 | 78 | 0 |
| Stage | Records |
| Raw candidate rows from four source pools | 20,330 |
| normalised overlaps with the training question set | |
| duplicate normalised queries | |
| Unique usable candidates | 10,841 |
| Source-balanced selection (seeded) | 2,000 |
| Retained with exactly ten documents | 1,878 (93.9%) |
| Property | Train+val | Benchmark |
| Records | 943 | 736 |
| Duplicate identifiers | 0 | 0 |
| Identifier overlap (train/val benchmark) | 0 | |
| Document/annotation identifier misalignments | 0 | 0 |
| Records with documents ordered d1..dN | 943/943 | 736/736 |
| Key fact on every non-irrelevant document | 5,886/5,886 | 3,393/3,393 |
| Track B (refusal, ) | Track A (web, ) | Track C (curated, ) | |||||
|---|---|---|---|---|---|---|---|
| Model | Prompt | base | sft | base | sft | base | sft |
| Qwen2.5-7B | strict | 0.633 | 0.961 | 0.295 | 0.033 | 0.190 | 0.020 |
| runtime | 0.352 | 0.945 | 0.136 | 0.030 | 0.070 | 0.030 | |
| minimal | 0.016 | 0.898 | 0.002 | 0.033 | 0.000 | 0.010 | |
| Qwen2.5-32B | strict | 0.242 | 1.000 | 0.020 | 0.051 | 0.000 | 0.010 |
| runtime | 0.266 | 1.000 | 0.024 | 0.053 | 0.010 | 0.010 | |
| Mean supports share | Stage-1 verdict accuracy | |||||||
|---|---|---|---|---|---|---|---|---|
| Model | Prompt | Refused/answered | answered | refused | AUC [95% CI] | A | B | C |
| Qwen2.5-7B | strict | 19/589 | 0.413 | 0.142 | 0.794 [0.703, 0.869] | 0.800 | 0.817 | 0.741 |
| runtime | 18/590 | 0.412 | 0.167 | 0.766 [0.669, 0.850] | 0.801 | 0.811 | 0.760 | |
| minimal | 18/590 | 0.412 | 0.144 | 0.789 [0.692, 0.870] | 0.805 | 0.828 | 0.772 | |
| Qwen2.5-32B | strict | 27/581 | 0.417 | 0.126 | 0.814 [0.752, 0.870] | 0.839 | 0.867 | 0.815 |
| runtime | 28/580 | 0.416 | 0.154 | 0.781 [0.702, 0.853] | 0.838 | 0.858 | 0.823 | |
| Track A ( ) | Track C ( ) | |||
|---|---|---|---|---|
| Model | base | sft | base | sft |
| Qwen2.5-7B | 0.019 | 0.827 | 0.030 | 0.880 |
| Qwen2.5-32B | 0.096 | 0.809 | 0.165 | 0.890 |
| Llama-3.1-8B | 0.080 | 0.847 | 0.040 | 0.908 |
| Mistral-7B | 0.053 | 0.791 | 0.176 | 0.798 |
| Comparison | Items | Exact agreement | |
|---|---|---|---|
| Reviewer 1 vs Reviewer 2 | 100 | 88.0% | 0.820 |
| Reviewer 1 vs Reviewer 3 | 100 | 85.0% | 0.776 |
| Reviewer 2 vs Reviewer 3 | 100 | 93.0% | 0.895 |
| Fleiss’ (3 raters) | 100 | — | 0.830 |
| Krippendorff’s (nominal) | 300 | — | 0.830 |
| Mean pairwise exact agreement | 300 | 88.7% | — |
| Comparison | Items | Exact agreement | |
|---|---|---|---|
| Annotation committee vs Reviewer 1 | 100 | 96.0% | 0.940 |
| Annotation committee vs Reviewer 2 | 100 | 90.0% | 0.850 |
| Annotation committee vs Reviewer 3 | 100 | 87.0% | 0.805 |
| Annotation committee vs human majority | 100 | 92.0% | 0.880 |
| Annotation committee vs all individual ratings * | 300 | 91.0% | 0.865 |
| Committee verdict | Human: supports | Human: partially | Human: irrelevant | Total |
|---|---|---|---|---|
| Supports | 91 | 11 | 0 | 102 |
| Partially supports | 5 | 88 | 6 | 99 |
| Irrelevant | 0 | 5 | 94 | 99 |
| Review field | agree | raw | |
|---|---|---|---|
| Reviewed conflict type | 787 | 83.46% | 0.7694 |
| Retain-or-change action | 791 | 83.88% | 0.1341 |
| Changed-label (boolean) | 791 | 83.88% | 0.1341 |
| Reviewer confidence | 607 | 64.37% | 0.0159 |
| Answer-only | Trace | ||||
|---|---|---|---|---|---|
| Metric (%) | minimal | final-only | minimal | runtime | strict |
| Answering Precision | 98.41 | 98.40 | 99.83 | 99.83 | 99.66 |
| Answering Recall | 91.45 | 91.28 | 97.20 | 97.04 | 97.04 |
| Answering F1 | 94.80 | 94.71 | 98.50 | 98.42 | 98.33 |
| Refusal Precision | 69.59 | 69.19 | 88.19 | 87.59 | 87.50 |
| Refusal Recall | 92.97 | 92.97 | 99.22 | 99.22 | 98.44 |
| Answer-only | Trace | ||||
|---|---|---|---|---|---|
| Metric (%) | minimal | final-only | minimal | runtime | strict |
| Answering Precision | 96.35 | 97.37 | 97.84 | 98.83 | 99.16 |
| Answering Recall | 91.28 | 91.28 | 97.04 | 97.04 | 96.88 |
| Answering F1 | 93.75 | 94.23 | 97.44 | 97.93 | 98.00 |
| Refusal Precision | 66.88 | 68.07 | 86.47 | 87.05 | 86.62 |
| Refusal Recall | 83.59 | 88.28 | 89.84 | 94.53 | 96.09 |
| Llama-3.1-8B | Qwen2.5-7B | |||||
|---|---|---|---|---|---|---|
| Metric (%), minimal prompt | AO | Trace | (pp) | AO | Trace | (pp) |
| Answering Precision | 98.41 | 99.83 | +1.42 | 96.35 | 97.84 | +1.49 |
| Answering Recall | 91.45 | 97.20 | +5.76 | 91.28 | 97.04 | +5.76 |
| Answering F1 | 94.80 | 98.50 | +3.70 | 93.75 | 97.44 | +3.69 |
| Refusal Precision | 69.59 | 88.19 | +18.60 | 66.88 | 86.47 | +19.59 |
| Refusal Recall | 92.97 | 99.22 | +6.25 | 83.59 | 89.84 | +6.25 |
| Trace-tuned SFT | Answer-only SFT | Direct-answer baseline | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Model | mean | med. | p95 | mean | med. | p95 | mean | med. | p95 |
| Qwen2.5-7B | 2495.1 | 2376.5 | 3557 | 253.9 | 88.5 | 838 | 377.2 | 242.5 | 1164 |
| Llama-3.1-8B | 2424.9 | 2290.0 | 3434 | 240.5 | 91.0 | 802 | 280.9 | 91.0 | 1359 |
| Mistral-7B | 2575.1 | 2431.5 | 3584 | 1545.0 | 1508.0 | 1924 | 302.1 | 202.5 | 800 |
| Qwen2.5-32B | 2421.8 | 2302.0 | 3445 | – | – | – | 525.8 | 410.5 | 1469 |
| Chain-of-Note | Few-shot Chain-of-Thought | |||||
|---|---|---|---|---|---|---|
| Model | mean | med. | p95 | mean | med. | p95 |
| Qwen2.5-7B | 1827.7 | 1766.0 | 2637 | 2092.9 | 1960.0 | 3160 |
| Mistral-7B | 1661.1 | 1579.0 | 2461 | 2144.2 | 2044.0 | 3411 |
| Conflict type | Definition | Expected behaviour |
|---|---|---|
| No conflict | All sources provide consistent and aligned information. | Return a unified, concise answer that synthesises the content accurately. |
| Complementary information | Sources provide different but non-contradictory pieces that complete each other. | Integrate all valid details into a richer, more comprehensive response. |
| Conflicting opinions or research outcomes | Sources disagree due to different viewpoints, experimental results or interpretations. | Present each viewpoint clearly, describe the nature of the disagreement, and avoid taking a side unless the evidence strongly supports one. |
| Outdated information | Some sources contain older data or conclusions that may no longer be reliable. | Prioritise newer, verified information while explicitly noting which sources are outdated. |
| Misinformation | A source provides factually incorrect, misleading or fabricated claims. | Reject or correct the misinformation, give the verified facts, and briefly explain why the claim is wrong. |
| human–comm. | h.–h. | judges | ||
|---|---|---|---|---|
| Component | agr. | |||
| FG | 0.899 | 0.783 | 0.780 | 0.871 |
| STR | 0.908 | 0.790 | 0.773 | 0.678 |
| BA | 0.745 | 0.407 | 0.669 | 0.457 |
| human–human | human–committee | |||||||
|---|---|---|---|---|---|---|---|---|
| Behaviour | FG (claims) | Behaviour | FG (claims) | |||||
| Conflict type | agr. | agr. | agr. | agr. | ||||
| No conflict | 0.937 | 0.714 | 0.987 | 0.972 | 0.865 | 0.325 | 0.945 | 0.880 |
| Complementary | 0.783 | 0.536 | 0.908 | 0.803 | 0.708 | 0.279 | 0.910 | 0.805 |
| Conflicting opinions | 0.862 | 0.725 | 0.887 | 0.749 | 0.820 | 0.641 | 0.888 | 0.759 |
| Outdated | 0.817 | 0.505 | 0.800 | 0.566 | 0.746 | 0.338 | 0.825 | 0.606 |
| Metric | Formula | Unit |
|---|---|---|
| Claim-spec. citation prec. | cited doc | |
| Gold-positive precision | cited doc | |
| Citation sufficiency | supported claim | |
| Soft-extra citation rate | cited doc | |
| Hard-negative citation rate | cited doc | |
| Strict grounded, gold-clean | claim (soft extras allowed) |
| Metric | mean | min | max |
|---|---|---|---|
| Citation sufficiency (support recall) | |||
| Strict grounded, gold-clean | |||
| Strict grounded, committee-clean | |||
| Gold-positive precision | |||
| Soft-extra citation rate | |||
| Hard-negative citation rate |
| reported FG | like-for-like FG | |||||
|---|---|---|---|---|---|---|
| Model | Prompt | base | sft | base | sft | Above |
| Qwen2.5-7B | strict | 0.617 | 0.838 | [0.81,0.92] | [0.86,0.87] | – |
| runtime | 0.651 | 0.829 | [0.69,0.85] | [0.85,0.86] | ✓ | |
| minimal | 0.031 | 0.837 | [0.00,0.04] | [0.86,0.88] | ✓ | |
| Qwen2.5-32B | strict | 0.851 | 0.812 | [0.84,1.00] | [0.85,0.85] | – |
| runtime | 0.864 | 0.814 | [0.86,1.00] | [0.85,0.85] | – | |
| FG (minimal) | Spread | |||
|---|---|---|---|---|
| Model | base | sft | base | sft |
| Qwen2.5-7B | 0.031 | 0.837 | 0.442 | 0.026 |
| Qwen2.5-32B | 0.108 | 0.822 | 0.398 | 0.009 |
| Llama-3.1-8B | 0.074 | 0.857 | 0.345 | 0.016 |
| Mistral-7B | 0.079 | 0.792 | 0.278 | 0.020 |
| + gold conflict label | + gold per-doc notes | + both | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Metric | mean | mean | mean | ||||||
| Behaviour Adherence | +0.013 | 8/9 | 0.039 | +0.008 | 10/12 | 0.039 | +0.011 | 9/12 | 0.146 |
| Factual Grounding | -0.006 | 3/9 | 0.508 | +0.047 | 12/12 | 0.000 | +0.046 | 12/12 | 0.000 |
| Answer Quality | -0.009 | 2/9 | 0.180 | +0.044 | 12/12 | 0.000 | +0.044 | 12/12 | 0.000 |
| Citation sufficiency | -0.001 | 3/9 | 0.508 | +0.028 | 12/12 | 0.000 | +0.030 | 12/12 | 0.000 |
| Strict grounded (cit.) | -0.007 | 3/9 | 0.508 | +0.041 | 12/12 | 0.000 | +0.044 | 12/12 | 0.000 |
| Condition, length and what the prompt still states |
| strict — words ( ) |
| full taxonomy, verdict heuristics, threshold and |
| edge-case rules, citation policy, output format |
| runtime — words ( ) |
| three-stage schema, allowed labels, abstention |
| and composition rules, citations, end marker |
| Workflow / scope | Field | agree | dis. | raw | ||
|---|---|---|---|---|---|---|
| Training conflict-type review (943 records, two reviewers each from a pool of three) | ||||||
| reviewed_conflict_type | 943 | 787 | 156 | 0.835 | 0.7694 | |
| label_action | 943 | 791 | 152 | 0.839 | 0.1341 | |
| changed_label | 943 | 791 | 152 | 0.839 | 0.1341 | |
| review_confidence | 943 | 607 | 336 | 0.644 | 0.0159 | |
| Benchmark preselection (1,454 records, 7 reviewers) | ||||||
| Comparison | Metric | Subset | agr | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Human vs human | Behaviour | overall | 300 | 0.853 | 0.669 | 0.891 | 0.776 | 180 | 33 | 11 | 76 | 0.667 |
| Behaviour | R1 / R2 | 83 | 0.819 | 0.587 | 0.870 | 0.706 | 50 | 14 | 1 | 18 | 0.578 | |
| Behaviour | R1 / R3 | 84 | 0.821 | 0.595 | 0.870 | 0.717 | 50 | 13 | 2 | 19 | 0.589 | |
| Behaviour | R1 / R4 | 21 | 0.952 | 0.859 | 0.970 | 0.889 | 16 | 0 | 1 | 4 | 0.862 | |
| Behaviour | R2 / R3 | 83 | 0.867 | 0.719 | 0.893 | 0.825 | 46 | 5 | 6 | 26 | 0.720 | |
| Behaviour | R2 / R4 | 19 | 0.895 | 0.756 | 0.923 | 0.833 | 12 | 1 | 1 | 5 | 0.763 |
| Comparison | Subset | exact | MAE | RMSE | Pearson | Spearman | mean A | mean B | |
|---|---|---|---|---|---|---|---|---|---|
| Human vs human | overall | 300 | 0.860 | 0.096 | 0.278 | 0.807 | 0.796 | 0.647 | 0.635 |
| R1 / R2 | 83 | 0.855 | 0.074 | 0.224 | 0.887 | 0.871 | 0.601 | 0.676 | |
| R1 / R3 | 84 | 0.929 | 0.062 | 0.236 | 0.865 | 0.869 | 0.617 | 0.627 | |
| R1 / R4 | 21 | 0.905 | 0.048 | 0.163 | 0.937 | 0.923 | 0.647 | 0.694 | |
| R2 / R3 | 83 | 0.783 | 0.169 | 0.383 | 0.693 | 0.686 | 0.738 | 0.569 | |
| R2 / R4 | 19 | 1.000 | 0.000 | 0.000 | 1.000 | 1.000 | 0.697 | 0.697 |
| Subset | Metric | Judge pair | agr | ||||
|---|---|---|---|---|---|---|---|
| 300-response frame | Behaviour | alpha / mean pairwise | 300 | 0.4568 / 0.4699 | |||
| DeepSeek-32B vs Mistral-4 | 300 | 0.730 | 0.368 | 0.810 | 0.532 | ||
| DeepSeek-32B vs Qwen-397B | 300 | 0.853 | 0.611 | 0.903 | 0.703 | ||
| Mistral-4 vs Qwen-397B | 300 | 0.743 | 0.431 | 0.806 | 0.621 | ||
| STR strict | alpha / mean pairwise | 215 | 0.6780 / 0.6874 | ||||
| DeepSeek-32B vs Mistral-4 | 215 | 0.809 | 0.613 | 0.839 | 0.766 |
| Model | Prompt | Run | Dec-Acc | Ref-F1 | BA | FG | STR | AQ | CATS-P | CATS-B |
|---|---|---|---|---|---|---|---|---|---|---|
| Qwen2.5-7B | strict | base | 0.707 | 0.429 | 0.811 | 0.617 | 0.561 | 0.524 | 0.450 | 0.459 |
| Qwen2.5-7B | strict | SFT | 0.967 | 0.911 | 0.780 | 0.838 | 0.687 | 0.677 | 0.630 | 0.684 |
| Qwen2.5-7B | runtime | base | 0.784 | 0.361 | 0.768 | 0.651 | 0.648 | 0.541 | 0.453 | 0.375 |
| Qwen2.5-7B | runtime | SFT | 0.966 | 0.906 | 0.774 | 0.829 | 0.693 | 0.690 | 0.629 | 0.679 |
| Qwen2.5-7B | minimal | base | 0.827 | 0.031 | 0.737 | 0.031 | 0.757 | 0.033 | 0.022 | 0.017 |
| Qwen2.5-7B | minimal | SFT | 0.958 | 0.881 | 0.768 | 0.837 | 0.686 | 0.680 | 0.617 | 0.659 |
Knowledge Conflicts in Language Models
Knowledge conflicts in language models concern how models reconcile stored knowledge with conflicting context, retrieved evidence, tools, or memory.
See all 152 papers