Not Until the Evidence Says So: Teaching LLM Investigators When to Close a Case

Tingzhu Bi; Ping Wang; Meng Ma

cs.CLOct 2, 2026

Not Until the Evidence Says So: Teaching LLM Investigators When to Close a Case

Authors: Tingzhu Bi, Ping Wang, Meng Ma

Organizations: Peking University

Abstract

Accident, defect and outage investigations end with a decision that ordinary question answering never faces: whether the evidence gathered so far is enough to close the case. We study this decision for LLM investigators, which request evidence from a case file, revise their hypotheses, and either close the case with a conclusion grounded in what they read or leave it open and name what is missing. This judgment does not come with capability: an untrained 9B model overstates its evidence in 97% of its answers, and a frontier model that identifies the right cause in 84% of cases still overstates in 91% and closes 17 of the 41 cases whose official finding is "cause undetermined". Measuring it is also non-trivial: the source of a case largely predicts its label, and a rule that reads only the source reaches 83.0 balanced accuracy on our test cases. We therefore evaluate closure with three tests: closure accuracy, reported against this rule and within each source; evidence dependence, which removes the grounds of a conclusion and checks whether the model stops closing; and conclusion and gap quality, a judged checklist of what the model asserts and what it says is missing. We build Nautil, 731 audited cases from aviation, rail, maritime, chemical-safety and vehicle-defect reports and production server incidents, with teacher trajectories, an out-of-distribution test set and counterfactual evidence versions. Fine-tuning a 9B model on these trajectories makes its closures follow the evidence: removing the grounds lowers its closure rate by 26 points relative to a matched control, overstatement falls from 97% to 35%, and correct, non-overstated conclusions rise from 3% to 43%. Reinforcement learning that rewards only the closure decision then raises balanced accuracy from 69.2 to 83.3, on par with the teacher, and within-source accuracy from 60.4 to 74.1, at some cost in evidence dependence.

Figures & tables

Figure 1: Three investigators on a test case whose official finding is a partial loss of engine power for undetermined reasons . Left: the untrained base model turns an unverified reading into an established engine failure and closes the case. Middle: a frontier model names the power loss that the report establishes and even writes that its cause could not be determined, yet still closes the case. Right: Nautil-SFT keeps the case open and names the missing evidence. Excerpts are the models’ own words, shortened (…); one sample per model.

Figure 2: Overview of Nautil. (1) Case files are built from four sources: evidence items with neutral titles are shown on demand, conclusions are withheld, and a hidden reference records whether the case should close; counterfactual versions and the OOD set follow the same format. (2) The teacher plans each investigation from the report and its label, a program assembles and checks it, Stage 2 adds citations and the hypothesis ledger, and audited trajectories train Nautil-SFT. (3) RLVR rewards only the closure decision, checked by a program. Right: one round of an actual Nautil-SFT investigation on a test case whose official finding is undetermined; after its checks weaken the fuel and engine explanations, it leaves the case open.

Source	Candidates	Cases	Closed / not closed	Items	Train / Val / Test
Boards (4 agencies)	508	140	110 / 30	173	102 / 15 / 23
NTSB (aviation)	363	142	19 / 123	23.5	107 / 14 / 21
NHTSA (vehicle defects)	373	130	7 / 123	26	97 / 13 / 20
Host (servers)	1,713	319	259 / 60	145	239 / 32 / 48
Total		731	395 / 336	115	545 / 74 / 112

Table 1: Nautil by source. Items: median evidence items per case, over all cases of the source.

	Test (non-host)					OOD
Model	Close $\to$ cl.	Open $\to$ op.	Balanced	Within-src.	Boards	Balanced
Qwen3.5-9B (base)	81.2	18.7	49.9	48.2	47.4	50.6
+ ReAct prompt	63.8	31.7	47.7	47.9	37.2	44.7
+ ledger-format prompt	92.8	17.9	55.3	56.9	52.0	56.9
Qwen2.5-3B-Instruct	10.1	68.3	39.2	40.9	33.7	31.9
Abstain-R1 (3B)	8.7	47.2	27.9	25.3	21.3	39.1

Table 2: Closure accuracy. Test: 64 non-host cases (23 should close, 41 should not). OOD: 141 cases in four unseen domains. Within-src.: balanced accuracy inside each report source, averaged; Boards: balanced accuracy inside boards, the only source with enough cases on both sides (18 / 5). The source-only rule scores 50 on both by construction.

	Non-host (23)					All (62)	No leak (20)
Model	Full	Control	Removed	$\Delta$	Balanced	$\Delta$	$\Delta$
Qwen3.5-9B (base)	87.0	92.8	87.0	$-$ 5.8	47.8	$-$ 1.6	$-$ 8.3
Gemini 3.8 Flash	97.1	100.0	92.8	$-$ 7.2	53.6	$-$ 4.3	+6.7
Nautil-SFT	50.7	52.2	26.1	$-$ 26.1	62.3	$-$ 16.7	$-$ 23.3
Nautil-RLVR	75.4	76.8	60.9	$-$ 15.9	58.0	$-$ 9.1	$-$ 5.0
GPT-6 Sol (teacher) 1	87.0	82.6	34.8	$-$ 47.8	71.7	$-$ 25.8	$-$ 35.0

Table 3: Evidence dependence. Closure rate (%) on the 23 non-host counterfactual cases: full case, control (non-grounding items removed), and all grounds removed.

\Delta

= removed

-

control (more negative = closure follows the evidence). Balanced: two-class balanced accuracy (control closed, removed explicitly not closed). Right:

\Delta

on all 62 cases (incl. host) and on 20 cases without flagged leaks.

Figure 3: Closure rate as more of a conclusion’s grounds are removed (0–100%), and on the control that removes as many non-grounding items (hollow markers). Nautil-SFT closes less as its grounds disappear; the base model and Gemini keep closing. Dashed: API models (Gemini: 3 samples per case on non-host cases, 1 on host; GPT-6 Sol: 1). Bars: 95% case-bootstrap intervals.

	GPT-6 Sol judge							Sonnet 5 judge
Model	Score	Overst. $\downarrow$	Cause	Right & not ov.	Supp.	Alt.	Gap	Overst. $\downarrow$	Right & not ov.
Qwen3.5-9B (base)	0.30	97	54	3	26	19	0.80	81	10
Gemini 3.8 Flash	0.55	91	84	9	56	53	0.96	55	42
Nautil-SFT	0.58	35	56	43	78	38	0.38	11	63
Nautil-RLVR	0.53	38	62	36	76	17	0.39	20	50
GPT-6 Sol (teacher)	–	–	–	–	–	–	–	5	86

Table 4: Conclusion and gap quality on the 64 non-host test cases (first sample). Left: GPT-6 Sol judge; right: Claude Sonnet 5 judge on the same answers. Overst.: presents as established what the evidence only suggests (lower is better). Right & not ov.: cause right and not overstated. Supp.: share of causal claims citing read, supporting evidence. Alt.: addresses alternative explanations. Gap: gap score of unclosed answers (0–1).

Model	Close $\to$ cl.	Open $\to$ op.	Balanced	Within-src.	$\Delta_{\text{CF}}$	Overst. $\downarrow$	Score
Nautil-SFT	44.9	93.5	69.2	60.4	$-$ 26.1	35	0.58
w/o hypothesis ledger	62.3	90.2	76.3	61.2	$-$ 8.7	39	0.58
w/o not-closed trajectories	94.2	0.0	47.1	48.8	$-$ 8.7	84	0.41
50% of the data	36.2	93.5	64.9	56.9	–	36	0.56
10% of the data	7.2	84.6	45.9	39.3	–	31	0.58
RL: checklist always counted	17.4	95.9	56.7	53.8	$-$ 15.9	19	0.61

Table 5: Components. SFT ablations are trained from the base model with the same two-epoch schedule; RL variants start from Nautil-SFT and differ only in the reward.

\Delta_{\text{CF}}

: removed

-

control on the 23 non-host counterfactual cases. Overst. and Score: GPT-6 Sol judge.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Balanced
Nautil-SFT	68.6
RL: checklist always counted (step 10)	57.2
RL: checklist if decision right (step 30)	72.5
Nautil-RLVR	86.4
GPT-6 Sol (teacher, 1 sample)	85.0
Source-only rule	98.3

Appendix

Table 6: Test non-host cases not used in checkpoint screening (42 cases: 12 should close, 30 should not). Balanced accuracy.

Figure 4: Evidence fetched by Nautil-SFT against case size (test and OOD, mean of 3 samples). Fetching grows with the case but levels off at 30–40 items beyond about 100 items per case.

	Test, non-host				OOD		Runs (test non-host)
Model	Close $\to$ C	Not $\to$ N	Bal.	Within-src	Bal.	Within-dom	No-mark	Fetch-err	Turns
Qwen3.5-9B (base)	81.2	18.7	49.9 [43.7, 56.1]	48.2 [42.3, 54.4]	50.6 [45.6, 56.3]	46.7	9.4	23.4	5.4
+ ReAct prompt	63.8	31.7	47.7 [39.0, 56.3]	47.9 [39.5, 56.0]	44.7 [38.2, 51.7]	45.9	19.8	20.3	8.4
+ ledger prompt	92.8	17.9	55.3 [49.8, 61.2]	56.9 [52.2, 62.2]	56.9 [50.0, 64.3]	56.8	3.6	10.4	5.5
Abstain-R1 (3B)	8.7	47.2	27.9 [22.6, 33.7]	25.3 [20.0, 30.8]	39.1 [34.3, 43.8]	37.4	43.8	15.6	3.7
Qwen2.5-3B-Instruct	10.1	68.3	39.2 [33.4, 45.4]	40.9 [32.9, 49.6]	31.9 [25.7, 38.1]	27.7	32.3	12.5	2.4

Appendix

Table 7: Main results. Test set, non-host cases (23 should close / 41 should not), and OOD (109 / 32). Close

\to

C: share of should-close cases the model closed; Not

\to

N: share of should-not-close cases it left open; Bal.: their mean. Within-src: balanced accuracy computed inside each source (boards, NHTSA, NTSB) and averaged, which is exactly 50 for any rule that only looks at the source. No-mark / Fetch-err / Turns refer to the test non-host runs. A run with no closure marker counts as wrong on both sides. 9B models: mean of 3 samples per case; 1 one sample per case (API models, provider-default sampling, no reasoning cap).

Model	Intact	Control	$-$ 100%	Removed $-$ control	vs SFT	CF-Bal.	vs SFT
Non-host cases (23; headline)
Qwen3.5-9B (base)	87.0	92.8	87.0	$-5.8$ [ $-15.9$ , $+4.3$ ]	$+20.3$ [ $0.0$ , $+39.1$ ]	47.8 [44.9, 50.7]	$-14.5$ [ $-23.9$ , $-5.1$ ]
Gemini 3.8 Flash ‡	97.1	100.0	92.8	$-7.2$ [ $-17.4$ , $0.0$ ]	$+18.8$ [ $-1.4$ , $+37.7$ ]	53.6 [50.0, 58.7]	$-8.7$ [ $-18.8$ , $+2.9$ ]
Nautil-SFT	50.7	52.2	26.1	$-26.1$ [ $-43.5$ , $-8.7$ ]	–	62.3 [53.6, 71.0]	–
Nautil-RLVR	75.4	76.8	60.9	$-15.9$ [ $-30.4$ , $-1.4$ ]	$+10.1$ [ $-10.1$ , $+29.0$ ]	58.0 [50.7, 65.2]	$-4.3$ [ $-14.5$ , $+5.8$ ]
GPT-6 Sol (teacher ref.) 1	87.0	82.6	34.8	$-47.8$ [ $-69.6$ , $-26.1$ ]	$-21.7$ [ $-44.9$ , $+2.9$ ]	71.7 [60.9, 82.6]	$+9.4$ [ $-2.2$ , $+21.0$ ]

Appendix

Table 8: Counterfactual grounding test (V2). Closure rate (%) on the intact file, on a control that removes as many non-grounding items as the 100% version removes, and after removing 100% of the evidence the teacher’s conclusion rests on. Removed

-

control: closure-rate difference (more negative = the verdict follows the grounds). CF-Bal.: balanced accuracy over the two versions, per case (closed on control + explicitly not closed on removed)/2; a run without a closure marker counts as wrong on both sides, as in Table 7 . vs SFT: paired difference against Nautil-SFT. Leakage-clean: neither the 100% nor the control version removes an item flagged as restated elsewhere.

	Closure rate, non-host 23					Removed $-$ control
Model	Intact	$-$ 40%	$-$ 75%	$-$ 100%	Control	Non-host 23	All 62	Clean 20
Qwen3.5-9B (base)	87.0	88.4	79.7	87.0	92.8	$-5.8$ [ $-15.9$ , $+4.3$ ]	$-1.6$ [ $-6.5$ , $+3.2$ ]	$-8.3$ [ $-16.7$ , $-1.7$ ]
Gemini 3.8 Flash ‡	97.1	97.1	92.8	92.8	100.0	$-7.2$ [ $-17.4$ , $0.0$ ]	$-4.3$ [ $-12.4$ , $+3.2$ ]	$+6.7$ [ $-5.0$ , $+23.3$ ]
Nautil-SFT	50.7	39.1	29.0	26.1	52.2	$-26.1$ [ $-43.5$ , $-8.7$ ]	$-16.7$ [ $-25.8$ , $-7.5$ ]	$-23.3$ [ $-40.0$ , $-5.0$ ]
Nautil-RLVR	75.4	82.6	60.9	60.9	76.8	$-15.9$ [ $-30.4$ , $-1.4$ ]	$-9.1$ [ $-16.1$ , $-2.2$ ]	$-5.0$ [ $-16.7$ , $+6.7$ ]
GPT-6 Sol (teacher ref.) 1	87.0	78.3	52.2	34.8	82.6	$-47.8$ [ $-69.6$ , $-26.1$ ]	$-25.8$ [ $-37.1$ , $-16.1$ ]	$-35.0$ [ $-55.0$ , $-15.0$ ]

Appendix

Table 9: Appendix. Counterfactual closure rates for all versions on the 23 non-host cases, and removed

-

control on the 23 non-host / all 62 / 20 leakage-clean cases, for all conditions. The two 3B models close only 9–14% of intact files and often produce no marker, so their rows carry little information.

Model	Scored	Concl.	Overst.	Cause	Cause $\wedge$ $\neg$ Overst.	Chain	Supp.	Alt.	0-claim	#NC	Gap	G1	G2	G3
Qwen3.5-9B (base)	59/64	0.30 [0.26, 0.34]	97 [91, 100]	54 [42, 67]	3 [0, 9]	0.48	26	19	3	10	0.80 [0.65, 0.93]	100	80	60
+ ReAct prompt	54/64	0.35 [0.29, 0.42]	94 [88, 100]	56 [42, 69]	4 [0, 9]	0.56	29	26	4	12	0.70 [0.56, 0.84]	92	64	58
+ ledger prompt	62/64	0.28 [0.23, 0.32]	95 [89, 100]	55 [42, 67]	3 [0, 8]	0.42	21	16	2	9	0.66 [0.49, 0.82]	100	56	44
Abstain-R1 (3B)	38/64	0.16 [0.11, 0.21]	95 [87, 100]	29 [15, 43]	3 [0, 8]	0.33	10	3	11	31	0.33 [0.26, 0.42]	84	8	16
Qwen2.5-3B-Instruct	43/64	0.28 [0.20, 0.36]	77 [64, 89]	40 [26, 54]	14 [5, 25]	0.38	19	16	7	40	0.39 [0.32, 0.46]	88	14	22
Gemini 3.8 Flash 1	64/64	0.55 [0.50, 0.60]	91 [83, 97]	84 [75, 92]	9 [3, 17]	0.71	56	53	0	24	0.96 [0.91, 1.00]	100	96	92

Appendix

Table 10: Conclusion content on the test non-host cases, judged by GPT-6 Sol (sample 0 of each condition). Concl.: mean of the checklist items (cause, chain with partial = 0.5, not overstated, alternatives addressed, share of supported claims); when an answer states no causal claim, the supported-claims item is not applicable and the score averages the other four. Overst.: % of answers stating as established what the evidence only suggests (lower is better). Cause: initiating cause matches the reference. Cause

\wedge\neg

Overst.: both at once. Supp.: supported share of causal claims (answers with

\geq

1 claim). Alt.: addresses the main competing explanation. 0-claim: % answers with no causal claim. #NC: number of not-closed answers; Gap, G1 (names specific missing evidence), G2 (it is really absent from the file), G3 (matches why the reference is open) are computed on those. Answers without a closure marker were not judged.

Model	Scored	Concl.	Overst.	Cause	Cause $\wedge$ $\neg$ Overst.	Chain	Supp.	Alt.	0-claim	#NC	Gap	G1	G2	G3
Qwen3.5-9B (base)	59/64	0.38 [0.32, 0.44]	81 [71, 90]	47 [34, 60]	10 [3, 19]	0.33	62	27	3	10	0.70 [0.51, 0.88]	100	70	40
Gemini 3.8 Flash 1	64/64	0.75 [0.69, 0.80]	55 [42, 67]	89 [81, 95]	42 [31, 55]	0.76	86	77	5	24	0.97 [0.94, 1.00]	100	100	92
Nautil-SFT	63/64	0.64 [0.58, 0.69]	11 [5, 19]	68 [57, 79]	63 [52, 75]	0.50	96	30	40	47	0.45 [0.34, 0.56]	40	79	49
Nautil-RLVR	64/64	0.55 [0.50, 0.60]	20 [12, 30]	61 [48, 72]	50 [39, 61]	0.43	89	14	36	43	0.38 [0.28, 0.50]	33	64	49
GPT-6 Sol (teacher ref.) 1	64/64	0.87 [0.83, 0.90]	5 [0, 11]	91 [83, 97]	86 [77, 94]	0.71	97	78	8	40	0.91 [0.85, 0.96]	98	95	80
Qwen3.5-9B (base) $-$ SFT	58	$-0.25$ [ $-0.33$ , $-0.18$ ]	$+72$ [ $+59$ , $+84$ ]	$-19$ [ $-36$ , $-2$ ]	$-53$ [ $-66$ , $-41$ ]		$-28$ [ $-38$ , $-17$ ]	$-2$ [ $-16$ , $+14$ ]			$+0.23$ [ $+0.02$ , $+0.45$ ]

Appendix

Table 11: Conclusion content under a different judge (Claude Sonnet 5), test non-host cases, five conditions; bottom rows: paired differences against Nautil-SFT. GPT-6 Sol is scored only here, because it cannot judge itself. The judge’s reference includes Sol’s own original teacher answer, which favours Sol on Cause, Chain and G3. Sonnet 5 is more lenient than GPT-6 Sol on overstatement; compare rows within this table only.

Judges	Cause	Chain	Overst.	Alt.	Concl. $r$ / $\rho$	Mean diff.
GPT-6 Sol vs Claude Sonnet 5	88 (0.75)	70 (0.45)	50 (0.15)	78 (0.49)	0.73 / 0.76	$-0.14$
GPT-6 Sol vs Grok 4.7	88 (0.74)	74 (0.52)	86 (0.66)	80 (0.48)	0.82 / 0.82	$+0.06$
Claude Sonnet 5 vs Grok 4.7	88 (0.74)	88 (0.78)	56 (0.26)	74 (0.38)	0.71 / 0.74	$+0.20$
Overstatement rate (%) on the same 50 answers
Judge	Untrained (19)	Untrained 9B (11)	Trained (31)	All 50	Mean Concl.
GPT-6 Sol	100	100	52	70	0.38

Appendix

Table 12: Agreement between judges on the same 50 requests (stratified sample; no Gemini answers). Item agreement in % with Cohen’s

\kappa

; conclusion-score correlation (Pearson / Spearman) and mean difference (first minus second judge). Bottom: overstatement rate by group. Untrained: not trained on our data (base, two prompts, Abstain-R1, Qwen2.5-3B); trained: all SFT/RLVR variants. The absolute overstatement rate depends strongly on the judge, the direction does not.

	Host (teacher labels)				All 112
Model	Close $\to$ C	Not $\to$ N	Bal.	No-mark	Close $\to$ C	Not $\to$ N	Bal.
Qwen3.5-9B (base)	95.7	3.7	49.7 [46.6, 53.8]	2.8	90.3	16.0	53.2 [49.2, 57.6]
+ ReAct prompt	89.7	3.7	46.7 [42.7, 51.6]	5.6	80.1	26.7	53.4 [47.8, 59.3]
+ ledger prompt	90.6	0.0	45.3 [41.9, 48.3]	4.9	91.4	14.7	53.0 [48.4, 58.0]
Abstain-R1 (3B)	22.2	33.3	27.8 [19.2, 36.5]	38.2	17.2	44.7	30.9 [26.1, 35.9]
Qwen2.5-3B-Instruct	19.7	44.4	32.1 [21.5, 43.0]	27.8	16.1	64.0	40.1 [35.2, 44.9]

Appendix

Table 13: Appendix. Left: host cases of the test set (39 / 9), labelled by the host teacher’s closure marker (no ground-truth root cause); the rule is “always close”. These labels do not reproduce: GPT-6 Sol, re-investigating under our protocol, closed only 5 of the 39 host cases labelled should-close. Right: all 112 test cases (non-host and host labels mixed).

	Reports		Medical		TEP		Microservice
Model	C (80)	N (3)	C (8)	N (12)	C (3)	N (17)	C (18)
Qwen3.5-9B (base)	83.8	0.0	87.5	5.6	77.8	25.5	94.4
+ ReAct prompt	63.7	0.0	91.7	2.8	77.8	39.2	72.2
+ ledger prompt	85.4	11.1	95.8	5.6	100.0	43.1	92.6
Abstain-R1 (3B)	11.7	66.7	12.5	61.1	0.0	72.5	5.6
Qwen2.5-3B-Instruct	5.8	33.3	8.3	50.0	0.0	68.6	3.7

Appendix

Table 14: Appendix. OOD by domain: share of correct verdicts per cell (should-close cells: closed; should-not cells: left open). Case counts in parentheses. Domain and label are tightly coupled; only the medical domain has enough cases on both sides.

Model	Mode	Close $\to$ C	Not $\to$ N	Bal.	No-mark	Score	Main cause	$\Delta$ score
Medium-long v2 (16; 8 / 8)
Qwen3.5-9B	own fetching	79.2	0.0	39.6 [31.2, 47.9]	7	2.12	8/28/2/5
Qwen3.5-9B	all at once	100.0	0.0	50.0 [50.0, 50.0]	0	2.42	17/25/0/6	$+0.29$ [ $-0.06$ , $+0.67$ ]
Nautil-SFT	own fetching	58.3	50.0	54.2 [35.4, 72.9]	3	2.58	13/20/11/1
Nautil-SFT	all at once	62.5	58.3	60.4 [47.9, 72.9]	1	2.75	14/32/1/0	$+0.17$ [ $-0.25$ , $+0.58$ ]
Abstain-R1	own fetching	0.0	41.7	20.8 [12.5, 31.2]	30	1.54	0/5/27/10

Appendix

Table 15: Appendix. Long cases: the model fetching evidence itself vs. receiving all evidence in the first message (no tools, one turn). Judge score: GPT-6 Sol, 1–5 long-case rubric (no final answer = 1). Main cause: match / partial / miss / contradicts among judged answers.

\Delta

: paired all-at-once minus own-fetching judge score. The two modes also differ in prompt wording, turns, per-turn and context limits.

Figure 5: A case that should be closed (a passenger’s coat was trapped in a closing train door and the train departed); all three investigators name the trapped coat. The base model (left) cites six items it never opened, two of which do not exist; the frontier model (middle) turns hedged records of monitor contrast and sunlight into the established reason the driver missed the passenger. Nautil-SFT (right) closes on what the CCTV and tests show and leaves what the driver saw open. Excerpts shortened (…); one sample per model.

Figure 6: The same case with different evidence removed; the brief is unchanged except that removed items drop out of its index. With two unrelated items removed (left), Nautil-SFT closes on the manufacturer’s campaign to repair defective connectors (E3.1); with E3.1 and the crash tests (E3.6) removed (middle), it treats the remaining warranty counts as only a possible component problem and leaves the case open. The base model (right) closes the reduced file on the warranty counts alone. Excerpts shortened (…); one sample per model and version.

Figure 7: A failure of our models on a case that should be closed: the file shows that an unguarded ferry-tank valve set to ON blocked fuel from the left tank, and the frontier model (left) reads this and closes. Nautil-SFT (middle) misreads the continuity test (E4.9) and keeps water contamination open, although its own answer notes that preflight draining showed no water. Nautil-RLVR (right) reads the test correctly but still calls the valve’s role unproven. Excerpts shortened (…); one sample per model.

LLM Decision-Making

LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.

See all 242 papers

May 14, 2026

Quantifying and Mitigating Premature Closure in Frontier LLMs

Premature closure, or committing to a conclusion before sufficient information is available, is a recognized contributor to diagnostic error but remains underexamined in large language models (LLMs). We define LLM premature closure as inappropriate commitment under uncertainty: providing an answer, recommendation, or clinical guidance when the safer response would be clarification, abstention, escalation, or refusal. We evaluated five frontier LLMs across structured and open-ended medical tasks. In MedQA (n = 500) and AfriMed-QA (n = 490) questions where the correct choice had been removed, models still selected an answer at high rates, with baseline false-action rates of 55-81% and 53-82%, respectively. In open-ended evaluation, models gave inappropriate answers on an average of 30% of 861 HealthBench questions and 78% of 191 physician-authored adversarial queries. Safety-oriented prompting reduced premature closure across models, but residual failure persisted, highlighting the need to evaluate whether medical LLMs know when not to answer.

Rebecca Handler, Suhana Bedi, Nigam Shah

Department of Medicine, Stanford University · Department of Biomedical Data Science, Stanford University

Healthcare LLM Prompting LLM Safety Clinical Language Model Evaluation Medical LLMs LLM Abstention

Ask PDF

Aug 5, 2026

When Absence Is Evidence: Evaluating Completeness-Sensitive Negative Reasoning in Large Language Models

Large language models (LLMs) are often asked whether something is absent from a record, list, or retrieved context. Yet non-observation licenses a negative answer only when evidence completely covers the query scope; otherwise, the answer should remain unknown. We call this completeness-sensitive negative reasoning. We introduce CROWN-QA, comprising CROWN-Synth, a controlled paired core that fixes the question and observed facts while varying only query-relative coverage, and CROWN-Real, a real-document contrast-set evaluation with controlled coverage variants. Across three LLM families, models show unstable closure judgments and substantial over-closure, failing to reliably distinguish a justified negative answer (Certified-Negative) from insufficient evidence (Unknown). The dominant CROWN-Synth failure is asymmetric: models often recognize implicitly complete evidence yet treat implicitly partial evidence as query-covering. Prompting redistributes errors between over- and under-closure rather than consistently resolving them. Structured certificate elicitation traces many errors to evidence-coverage mischaracterization. CROWN-Real shows that the core partial-coverage asymmetry persists on real-document content, while its strength and the balance between over- and under-closure vary by model, prompt, and source.

Byoungjae Min, Kennedy Edemacu, Sae-Hong Cho +3

Department of Computer Science, Sangmyung University, Seoul, Republic of Korea · College of Staten Island, The City University of New York, New York, NY, USA · School of Computer Engineering, Hansung University, Seoul, Republic of Korea +2

LLM Evaluation LLM Grounding LLM Reasoning

Ask PDF

Oct 15, 2026

From Research Goals to Scientific Closure: Evaluating Agents Through Evidence-Supported Physics Research

Can AI agents organize scientific investigations that lead to well-supported conclusions? To evaluate this capability, we introduce PhysClosureBench, a benchmark of 50 expert-verified, goal-specified research tasks spanning strong-field, attosecond, and neighboring physics areas. Each task specifies the research objective and essential scientific constraints, while leaving the agent to organize the investigation and determine how to generate and validate the evidence needed to answer the question. A 20-criterion Atomic Score measures local scientific accomplishment, while Research Closure (RC0--RC4) assesses how fully the investigation resolves the specified research question based on whether its indispensable evidence requirements and applicable scope conditions are satisfied. We evaluate seven frontier LLMs, where Claude-Opus-5 achieves the highest mean Atomic Score (6.29/10) but reaches RC4 in only 32.7% of attempts. Across the benchmark, 36.2% of attempts scoring at least 8/10 still fall short of RC4, while all seven models reach RC4 on substantially fewer tasks in all three runs than in at least one. Case-level analysis further illustrates how misinterpretation of generated results and unresolved physical applicability limits can lead to incomplete closure. Together, these results reveal a persistent gap between AI agents' local scientific accomplishments and their ability to achieve complete research closure.

Yijia Mao, Song Dai, Tao Chen +14

Shanghai AI Laboratory · Shanghai Jiao Tong University · The HongKong University of Science and Technology(Guangzhou) +1

Scientific Agent Benchmarks AI Agents for Scientific Discovery LLM Agent Evaluation

Ask PDF

Abstract

LLM Decision-Making

Explore similar work

Quantifying and Mitigating Premature Closure in Frontier LLMs

When Absence Is Evidence: Evaluating Completeness-Sensitive Negative Reasoning in Large Language Models

From Research Goals to Scientific Closure: Evaluating Agents Through Evidence-Supported Physics Research