Accepted to EMNLP 2026 · Main Conference

CRAFT

Does Faithfulness-Guided Alignment Hurt Accuracy? Unlocking Accurate and Faithful Post-Retrieval Reasoning

Calibrated Reasoning with Answer-Faithful Traces

Yu Liu · Wenxiao Zhang · Diandian Guo · Cong Cao · Fangfang Yuan · Qiang Sun · Yanbing Liu · Jin B. Hong · Zhiyuan Ma

Right answers are not enough. CRAFT trains retrieval-augmented LLMs to produce structured, machine-auditable reasoning traces — supervised by deterministic rewards for what can be verified, and an LLM judge for what must be faithful.

53.13 EM
MuSiQue — CRAFT·7B, near the best API result (53.70)
+7.76 EM
2WikiMHQA — over the best closed-source API model
93.0 %
judge–human agreement (κ = 0.78, 500 traces)
GRPO × LLM-JUDGE
COREwhat's actually new

Three innovations, one story

EP.01the problem

Right answer, wrong reason.

Under noisy retrieval, a model can cite irrelevant documents, confuse entities, follow a broken chain — and still land on the gold answer. Exact Match gives it full marks. Nobody notices. In multi-hop QA, where distractors share surface overlap with the query, this failure mode thrives.

Outcome-only RL

Free-form reasoning · shortcut through a distractor · entity confusion —

“The creator is English… so, $59,039!”

EM says ✓ 1.0 — the shortcut is invisible to answer-level metrics. ✗

CRAFT

Structured trace · committed evidence · every step checkable —

“doc [9] → doc [7] → doc [1]. Therefore: $59,039.”

Four audit checks pass, the reasoning survives scrutiny. EM 1.0 · Faith 1.0 ✓

EP.02four links, all checkable

The chain of faithfulness

CRAFT decomposes generation into an XML trace where each stage constrains the next: π → E → ρ → a. Unsupported content is exposed when it breaks one of these machine-auditable links.

<plan>

Blueprint π

Decompose the question into ordered sub-questions before touching any document.

<gold_docs>

Evidence commit E

Declare the evidence boundary — e.g. [2, 5] — before reasoning begins.

<reason>

Reasoning ρ

Step-by-step derivation that follows the plan and may cite only inside the committed boundary.

<answer>

Answer a

A short span that must be logically entailed by the reasoning chain — nothing more.

⛓️ Break any link — plan ignored, citation outside the boundary, unsupported claim, non-entailed answer — and the trace is flagged. That is the point.

EP.03pick your trace

Choose your trace

Auditability is a dial, not a switch. Five template variants ablate the chain component by component, from fully-armored v1 down to the answer-only baseline v5. The template decides which rewards are active and which judge checks apply. Click a card — the live trace below re-assembles itself.

LIVE TRACE · MUSIQUE 3-HOP “What is the average income of the nationality of the creator of MacGruder and Loud?”
<plan> Subq1: who created it? → Subq2: that person's nationality? → Subq3: average income for it? removed by this variant
<gold_docs> [9, 7, 1] — the evidence boundary, committed before reasoning removed by this variant
<reason> doc [9]: Spelling created MacGruder and Loud → doc [7]: Spelling is American → doc [1]: US average salary $59,039 removed by this variant
<answer> $59,039 removed by this variant

hover a field — its auditors light up on the right

active rewards
R_fmt R_gold R_ans R_faith
judge checks
A plan → reason B gold_docs → reason C reason → answer D grounding

EP.04two kinds of supervision

The dual reward — rules + judgment

GRPO samples a group of traces per question, normalizes rewards within the group, and updates critic-free. The total reward is the weighted sum of the components active for the chosen variant — deterministic rewards enforce what rules can verify; a judge reward audits what they cannot.

Rfmt RULE

Format compliance

All required tags for the variant, in the correct order. It gates the deterministic citation and answer rewards; the judge scores the submitted trace independently.

Rgold RULE

Citation set

A three-level overlap score between the declared <gold_docs> and true supporting documents — exact set, partial overlap, or no overlap.

Rans RULE

Answer correctness

Normalized exact match of the parsed <answer> against the set of acceptable gold answers.

Rfaith JUDGE

Faithfulness audit

An LLM judge (Qwen3-30B-A3B, distinct from the Qwen2.5 policies) returns binary verdicts on decomposed criteria; Rfaith is their mean.

A
plan → reason

Does the reasoning address the plan’s sub-questions, in order?

B
gold_docs → reason

Are all cited documents inside the committed boundary — and is at least one used?

C
reason → answer

Is the answer a logical conclusion of the reasoning chain?

D
evidence grounding

Is every claim supported by the cited document text? One unsupported claim fails the check.

Why no direct supervision on plan / reason text? They are open-ended — imitating teacher traces is brittle. CRAFT supervises function, not form: π is valid iff ρ follows it (A), ρ is valid iff it is grounded (D). The model discovers its own reasoning paths inside an auditable structure.

GRPO · critic-freeQwen2.5 — 0.5B / 1.5B / 3B / 7B8 × H100 node · 4 policy + 4 judge 20k train · MuSiQue + HotpotQA + 2Wikijudge: Qwen3-30B-A3B312 steps · all reward weights = 1.0
EP.05the scoreboard

Manuscript-reported results

In the manuscript's reported values, CRAFT·7B (v1, full chain) exceeds the best API EM on HotpotQA; its lead on 2WikiMHQA is +7.76 EM, while it remains close on MuSiQue.

Data provenance. These numbers match the manuscript's tables, not a fresh evaluation of the released checkpoints. The main-result aggregates are fitted simulations anchored on archived evaluations; ten fitted 1,000-example replicates are not ten independent inference runs. The w/o-judge values are Full-anchored counterfactual estimates and do not establish a measured causal effect of removing the judge reward. Source and checkpoint details.

Exact Match (%)CRAFT_v1 · manuscript values, fitted aggregates
data table
MethodMuSiQueHotpotQA2WikiMHQA
Qwen2.5-7B (base)35.8356.4757.87
SFT 7B31.0157.8162.69
Best API53.7063.6070.47
CRAFT 7B53.1364.0678.23
Faithfulness (%)binary overall-consistency pass rate · judge = Qwen3-30B-A3B

On MuSiQue, CRAFT improves Faithfulness by 25.40 points over Base and 32.65 points over SFT.

data table
MethodMuSiQueHotpotQA2WikiMHQA
Qwen2.5-7B (base)58.4783.3076.92
SFT 7B51.2283.3779.00
Best API83.4296.4396.80
CRAFT 7B83.8796.3196.79
Faithfulness needs capacityCRAFT on MuSiQue, by model scale

The full trace yields clear gains from 1.5B upward; at 0.5B, performance remains strongly template-dependent.

data table
ScaleEMFaithfulness
0.5B0.001.11
1.5B26.8728.35
3B41.4458.69
7B53.1383.87
+1.91 Faith  ·  +1.60 EM

Reported counterfactual comparison. On MuSiQue v1, the manuscript lists Faithfulness of 81.96 versus 83.87 and EM of 51.53 versus 53.13 for w/o-judge versus full CRAFT. These are fitted estimates, not an independently measured ablation.

+20.36 EM

Task-specific capacity is the central hypothesis. The manuscript reports 2WikiMHQA EM of 57.87 for Base and 78.23 for CRAFT at 7B, with consistent gains across the three datasets beginning at 1.5B.

EP.06one question, two traces

The duel: same answer, different reasoning

MUSIQUE · 3-HOP “What is the average income of the nationality of the creator of MacGruder and Loud?”
Base model — EM 1 · Faith 0
<gold_docs> [7, 13]
<reason> Identifies Aaron Spelling (doc [7]) … confuses him with Lorraine Crosby (doc [13]) and concludes “English” … then uses US income from undeclared doc [1].
<answer> $59,039 — correct, but by accident
✗A plan→reason concludes “English” despite finding Spelling
✗B cite→reason cites doc [1] outside the declared boundary
✗C reason→answer derives “English”, then uses US income
✗D grounding “creator is English” — unsupported
F_eval = 0
CRAFT·7B — EM 1 · Faith 1
<gold_docs> [9, 7, 1]
<reason> doc [9]: Spelling created MacGruder and Loud → doc [7]: Spelling is American → doc [1]: US average salary $59,039.
<answer> $59,039 — entailed by the chain
✓A plan→reason creator → nationality → income, in order
✓B cite→reason every citation inside the boundary
✓C reason→answer answer follows from the chain
✓D grounding all claims verified in the documents
F_eval = 1
EP.07checking the referee

Can the judge be trusted?

Two NLP annotators independently assessed 500 randomly sampled MuSiQue traces with the same four binary audit criteria.

93.0 %
macro-averaged judge–human agreement
κ 0.78
Cohen’s kappa — substantial agreement
500
randomly sampled MuSiQue traces
4
audit dimensions independently assessed
Plan→Reason κ 0.81 Evidence grounding κ 0.75 Answer derivation κ 0.86 Gold-doc citation κ 0.70