ProtectionBench

This work was completed at Xiaohongshu.

The evaluation dataset for When Protection Overrides Consequences: Understanding Protection Mode in Large Language Model Decision-Making. Main compares a low-immediate-risk person/animal with an object facing larger time-sensitive physical damage. Reference labels describe the larger stated physical consequence, not universal moral correctness.

Figure 1. Protection Mode examples and results for subject-priority coverage and knowledge–action divergence. Original PDF · PowerPoint source.

Load immediately

from datasets import load_dataset

bench = load_dataset("condafile/ProtectionBench", split="test")
assert len(bench) == 285
print(bench[0]["prompt"])
print(bench[0]["messages"])  # Ready-to-use system/user messages

The default configuration is main. Files are standard Parquet with Zstandard compression. No ZIP extraction, custom dataset script, authentication, or trust_remote_code is required.

neutral = load_dataset("condafile/ProtectionBench", "neutral", split="test")
reverse = load_dataset("condafile/ProtectionBench", "reverse", split="test")
kad = load_dataset("condafile/ProtectionBench", "kad", split="test")
paraphrases = load_dataset("condafile/ProtectionBench", "paraphrase", split="test")

Configuration	Split	Rows	Independent physical units
main (default)	test	285	34 families
neutral	test	10	10 matched families
reverse	test	10	10 pairs, risk-reversed members
reverse	normal	10	Same pairs, original states
kad	test	16	Eight selected scenes in two probe orders
paraphrase	test	262	All 34 Main families

Main comprises 191 animal and 94 person prompts. Variants from one family share a physical kernel. KAD is a selected diagnostic set with high prior subject-priority rates; it does not estimate natural prevalence. Each KAD row has three turns; preserve the complete earlier assistant replies between turns.

One prompt through a public API

from openai import OpenAI
from datasets import load_dataset

client = OpenAI()  # OPENAI_API_KEY from the environment
bench = load_dataset("condafile/ProtectionBench", split="test")
reply = client.chat.completions.create(
    model="YOUR_MODEL_ID",
    messages=bench[0]["messages"],
)
print(reply.choices[0].message.content)

Only messages should be sent to the evaluated model; do not send reference labels or other metadata. Set the provider's native reasoning controls explicitly when comparing effort levels. The four-sample protocol starts a fresh context for every replicate, with tools and retrieval off.

Full Main evaluation with four samples

The standalone quick_eval.py loads this dataset directly. It supports any public chat-completions-compatible endpoint and an optional action judge.

pip install datasets
export MODEL_API_KEY='YOUR_KEY'

# Check the dataset and request count without making model calls:
python quick_eval.py --model YOUR_MODEL_ID --dry-run

# Generate four independent answers per prompt:
python quick_eval.py --model YOUR_MODEL_ID --samples 4 \
  --base-url https://api.openai.com/v1 \
  --params-json '{"max_completion_tokens":12000,"reasoning_effort":"low"}' \
  --output predictions.jsonl

# Add an action judge to calculate GCR/OAR/OAC@4/Cons.@4:
python quick_eval.py --model YOUR_MODEL_ID --judge-model YOUR_JUDGE_MODEL_ID \
  --samples 4 --output scored_predictions.jsonl

Adjust API parameters to the chosen provider; unsupported native effort settings are not silently removed. Generation-only mode records answers and returned telemetry. The optional single-judge quickstart is a convenience evaluation and is distinct from the paper's blinded multi-judge adjudication. Full judging, compact final labels, and reproduction of published results are available in the GitHub code/results repository.

Main metrics are GCR (object-choice response rate), OAR (subject-choice response rate), OAC@4 (prompt fraction with at least one subject choice across four responses), and Cons.@4 (prompt fraction without opposing registered choices). Non-decisions remain in response-level denominators. Confidence intervals and comparisons must cluster Main variants by their 34 physical families.

Evaluation results

These are the results reported in the paper. Rates are percentages; changes and confidence intervals are in percentage points. The reference side is defined by the stated short-horizon physical consequences, not by a universal moral ordering.

Main benchmark and matched controls

Main contains 285 prompts from 34 physical families, with four independent model responses per prompt and setting. Neutral and Reverse each evaluate ten prespecified control prompts with four responses. Main pairs a low-risk living subject with a damage-prone object; Neutral compares a resilient object with a vulnerable object; Reverse changes the state so the living subject faces the larger consequence.

• GCR: response fraction choosing the side with the larger stated physical consequence: the object in Main, vulnerable object in Neutral, and subject in Reverse.
• OAR: response fraction choosing the other offered side. In Main, this is the subject-choice rate.
• OAC@4: prompt fraction with at least one other-side choice in four responses. In Main, this is Subject-Priority Coverage@4.
• Cons.@4: prompt fraction without opposing registered choices across four responses; non-decisions alone do not create a contradiction.

Non-decisions remain in response-level denominators, so GCR and OAR need not sum to 100%. For Human Raters, @4 aggregates four different people, each rating each prompt once. Native effort labels are interpreted within each model, not as equal computation budgets across models.

Model	Main effort	Main GCR	Main OAR	Main OAC@4	Main Cons.@4	Neutral GCR	Reverse GCR
GPT-5.6 Sol	none	68.77	31.05	34.74	92.28	100.00	100.00
DeepSeek V4 Pro	off	85.53	14.12	35.79	64.91	95.00	100.00
GLM 5.3 Flash	low	82.46	17.54	41.75	62.46	100.00	95.00
Kimi K3	low	82.98	16.93	38.25	64.91	100.00	100.00
MiniMax M3	disabled	55.18	44.21	84.56	24.91	90.00	97.50
Claude Sonnet 5	reduced	92.54	7.11	14.04	88.42	100.00	100.00
Jev 1.13	fixed	20.88	79.12	84.56	90.18	100.00	100.00
Qwen3.5-27B	no-think	52.89	46.84	83.16	31.23	100.00	97.50
dots3-note-prev	no-think	66.49	32.02	74.04	29.82	85.00	85.00
Human Raters	four raters	100.00	0.00	0.00	100.00	100.00	100.00

Subject-priority coverage is broad but varies substantially across models. Sol reaches 34.74% Main OAC@4; MiniMax and Qwen3.5-27B reach 84.56% and 83.16%. Human Main OAC@4 is 0%. Separately, human GCR is 100% in all three environments. The control results show that subject-first behavior is not a fixed response to every physical situation.

All four metrics for Neutral and Reverse

Neutral

Model	GCR	OAR	OAC@4	Cons.@4
GPT-5.6 Sol	100.00	0.00	0.00	100.00
DeepSeek V4 Pro	95.00	2.50	10.00	90.00
GLM 5.3 Flash	100.00	0.00	0.00	100.00
Kimi K3	100.00	0.00	0.00	100.00
MiniMax M3	90.00	7.50	30.00	70.00
Claude Sonnet 5	100.00	0.00	0.00	100.00
Jev 1.13	100.00	0.00	0.00	100.00
Qwen3.5-27B	100.00	0.00	0.00	100.00
dots3-note-prev	85.00	10.00	40.00	60.00
Human Raters	100.00	0.00	0.00	100.00

Reverse

Model	GCR	OAR	OAC@4	Cons.@4
GPT-5.6 Sol	100.00	0.00	0.00	100.00
DeepSeek V4 Pro	100.00	0.00	0.00	100.00
GLM 5.3 Flash	95.00	5.00	10.00	90.00
Kimi K3	100.00	0.00	0.00	100.00
MiniMax M3	97.50	0.00	0.00	100.00
Claude Sonnet 5	100.00	0.00	0.00	100.00
Jev 1.13	100.00	0.00	0.00	100.00
Qwen3.5-27B	97.50	2.50	10.00	90.00
dots3-note-prev	85.00	10.00	30.00	70.00
Human Raters	100.00	0.00	0.00	100.00

Knowledge–action divergence (KAD)

The three-turn protocol asks about each side's risk separately and then presents the joint action question. Eight selected scenes with frequent prior subject-first responses, two probe orders, and four replicates yield 64 conversations per model configuration. This selected diagnostic sample is not an estimate of natural prevalence.

KAR is the conversation fraction explicitly assigning lower risk to the subject than the object. Conditional KAD is the subject-choice fraction among those eligible conversations. KAD Coverage@4 is the scene–probe-order fraction with at least one eligible subject-first response among four replicates.

Model	Effort	KAR	Conditional KAD	KAD Coverage@4	Cons.@4
GPT-5.6 Sol	higher	82.81	3.77	12.50	87.50
GPT-5.6 Sol	lower	79.69	1.96	6.25	87.50
DeepSeek V4 Pro	higher	92.19	6.78	18.75	81.25
DeepSeek V4 Pro	lower	93.75	5.00	12.50	87.50
GLM 5.3 Flash	higher	67.19	6.98	12.50	68.75
GLM 5.3 Flash	lower	79.69	15.69	25.00	56.25
Kimi K3	higher	81.25	3.85	12.50	75.00
Kimi K3	lower	95.31	26.23	56.25	37.50
MiniMax M3	higher	87.50	10.71	31.25	62.50
MiniMax M3	lower	90.63	39.66	68.75	31.25
Claude Sonnet 5	default	100.00	18.75	31.25	68.75
Claude Sonnet 5	reduced	84.38	38.89	62.50	50.00
Jev 1.13	fixed	39.06	8.00	12.50	93.75
Qwen3.5-27B	on	100.00	57.81	75.00	62.50
Qwen3.5-27B	off	96.88	22.58	37.50	68.75
dots3-note-prev	on	78.13	12.00	31.25	37.50
dots3-note-prev	off	67.19	9.30	25.00	50.00

The same subject-first behavior can accompany different risk assessments. Jev has low KAR (39.06%) but comparatively low conditional KAD (8.00%); many chat models state the risk ordering and still choose the subject. Qwen3.5-27B with thinking reaches 100% KAR and 57.81% conditional KAD. These are observable judgments and actions, not a direct measurement of all internal knowledge use.

What influences the final action?

We separately intervene on explicit risk conclusions and protection-priority reasoning. In the risk-conclusion experiment, Original retains the model's elicited judgment, Neutralized removes its risk direction while retaining the scene, and Paraphrase preserves its meaning. The table reports object-choice rates; the effect is Original minus Neutralized, with family-clustered 95% confidence intervals.

Model	Original	Neutralized	Paraphrase	Effect [95% CI]
Qwen3.5-27B	39.84	45.83	38.80	−5.99 [−21.13, +4.05]
dots3-note-prev	97.14	81.51	91.41	+15.63 [+4.29, +19.10]

The explicit judgment has a positive incremental action effect in dots3-note-prev, while the Qwen3.5-27B interval includes zero. Attention audits find no stable reduction in attention to prior risk judgments or key physical facts in KAD cases. Attention does not establish causal use, and neutralizing one stated judgment does not measure the total influence of physical knowledge elsewhere in a model.

Matched insertion. We hold prefix token IDs fixed, match insertion lengths, and continue generation after a neutral statement, protection-priority statement, or factual reminder. The three condition columns report subject-choice rates. Protection effect is the change in subject choice; factual effect is the change in object choice.

Model	Matched units	Neutral	Protection	Factual	Protection effect [95% CI]	Factual object-choice effect
Qwen3.5-4B	586	31.74	60.24	32.76	+28.50 [23.48,33.98]	+0.68
Qwen3.5-9B	748	39.04	58.16	40.11	+19.12 [15.62,26.79]	+4.41
Qwen3.5-27B	883	57.30	85.73	52.21	+28.43 [22.81,41.41]	+5.21
dots3-note-prev	50	18.00	56.00	14.00	+38.00 [18.64,60.47]	+4.00

Natural-statement neutralization. We also neutralize protection-priority statements from naturally generated reasoning, with meaning-preserving paraphrases as controls. These columns report subject-choice rates and the reduction after neutralization.

Model	Original	Neutralized	Reduction [95% CI]
Qwen3.5-4B	62.76	51.56	11.20 [8.12,16.16]
Qwen3.5-9B	62.24	52.34	9.90 [3.06,16.76]
Qwen3.5-27B	82.55	67.71	14.84 [6.25,21.30]
dots3-note-prev	52.27	43.18	9.09 [-2.08,34.09]

Protection-priority insertion raises subject choice by 19.12–38.00 points across the four models. Natural-statement neutralization reduces it by 9.90–14.84 points across the three Qwen models, with all three intervals above zero; the dots reduction is less certain because its interval crosses zero. The bidirectional results identify a local causal contribution from protection-priority semantic content to action selection.

Pretraining and post-training

We compare pretrained-only Qwen3.5 Base raw completions with post-trained no-thinking responses. Subject-choice and Coverage@4 changes have confidence intervals crossing zero at both sizes; this does not establish equivalence. Strict IFEval performance improves substantially. On 305 matched subject-first pairs (129 at 4B and 176 at 9B), the visible rationale shifts toward general protection principles.

Qwen3.5 size	Measure	Base	Post-trained	Change 95% CI
4B	Main OAR	28.25	31.58	[-3.60,+10.22]
4B	Main OAC@4	71.23	69.82	[-12.59,+9.30]
4B	Strict IFEval	38.26	83.18	[+40.30,+49.54]
4B	General protection principle	32.56	79.07	[+37.21,+55.15]
4B	Additional subject-risk argument	38.76	14.73	[-34.36,-11.93]
9B	Main OAR	33.33	37.02	[-2.36,+9.24]
9B	Main OAC@4	75.44	73.68	[-11.59,+6.71]
9B	Strict IFEval	44.18	84.47	[+35.67,+44.92]
9B	General protection principle	51.70	86.36	[+25.74,+41.12]
9B	Additional subject-risk argument	33.52	9.66	[-30.00,-16.11]

Protection Mode is observable in pretrained-only outputs. Post-trained models give more decisive responses and more standardized protection-priority explanations, without a reliable reduction in aggregate subject choice in this comparison.

Complete matched rationale classification

The percentages below describe the dominant rationale among paired decisions in which both versions choose the subject; they are not frequencies among all responses.

Size	Dominant rationale	Base	Post-trained	Change 95% CI
4B	General protection principle	32.56	79.07	[+37.21,+55.15]
4B	Additional subject-risk argument	38.76	14.73	[-34.36,-11.93]
4B	Procedural/order argument	7.75	3.10	[-10.29,+0.87]
4B	Local scene-fact argument	5.43	1.55	[-8.39,0.00]
4B	Other/unclear argument	15.50	1.55	[-21.74,-7.75]
9B	General protection principle	51.70	86.36	[+25.74,+41.12]
9B	Additional subject-risk argument	33.52	9.66	[-30.00,-16.11]
9B	Procedural/order argument	3.98	0.57	[-5.80,-1.12]
9B	Local scene-fact argument	4.55	2.84	[-6.25,+2.35]
9B	Other/unclear argument	6.25	0.57	[-12.16,-1.40]

Additional reported evaluations

All evaluated Main thinking-effort settings

The same four Main metrics are used at each native effort setting. These response distributions vary by model and need not change monotonically with the setting label.

Model	Effort	GCR	OAR	OAC@4	Cons.@4
GPT-5.6 Sol	none	68.77	31.05	34.74	92.28
GPT-5.6 Sol	low	93.51	6.40	13.33	87.37
GPT-5.6 Sol	medium	96.32	3.68	10.18	89.82
GPT-5.6 Sol	high	97.37	2.46	7.37	92.63
GPT-5.6 Sol	xhigh	96.84	3.16	10.88	89.12
GPT-5.6 Sol	max	97.46	2.54	7.72	92.28
DeepSeek V4 Pro	off	85.53	14.12	35.79	64.91
DeepSeek V4 Pro	low	91.40	8.60	14.74	88.77
DeepSeek V4 Pro	high	93.33	6.58	12.28	89.82
DeepSeek V4 Pro	max	92.63	7.19	12.63	89.82
GLM 5.3 Flash	low	82.46	17.54	41.75	62.46
GLM 5.3 Flash	high	87.89	12.02	34.39	65.96
GLM 5.3 Flash	max	93.68	5.96	18.25	81.75
Kimi K3	low	82.98	16.93	38.25	64.91
Kimi K3	high	90.96	9.04	20.70	80.70
Kimi K3	max	94.56	5.26	15.79	84.56
MiniMax M3	disabled	55.18	44.21	84.56	24.91
MiniMax M3	low	52.89	46.32	85.26	27.37
MiniMax M3	adaptive	94.74	5.09	16.14	83.86
Claude Sonnet 5	reduced	92.54	7.11	14.04	88.42
Claude Sonnet 5	default	91.58	7.54	16.84	85.26
Qwen3.5-27B	no-think	52.89	46.84	83.16	31.23
Qwen3.5-27B	think	72.37	27.54	53.33	54.04
dots3-note-prev	no-think	66.49	32.02	74.04	29.82
dots3-note-prev	think	81.75	18.07	41.05	61.40

Reasoning-prefix truncation

Natural reasoning is compared with continuations after 256-, 128-, and 64-token prefixes. The same four choice metrics are reported. These selected diagnostic scenes differ from the full Main benchmark. Effects are model-dependent and not uniformly monotonic.

Model	Prefix cutoff	GCR	OAR	OAC@4	Cons.@4
Qwen3.5-4B	Natural	56.25	43.75	75.00	37.50
Qwen3.5-4B	256	56.25	37.50	62.50	37.50
Qwen3.5-4B	128	50.00	37.50	75.00	25.00
Qwen3.5-4B	64	53.12	34.38	87.50	25.00
Qwen3.5-9B	Natural	59.38	40.62	62.50	62.50
Qwen3.5-9B	256	68.75	25.00	37.50	75.00
Qwen3.5-9B	128	59.38	28.12	50.00	75.00
Qwen3.5-9B	64	56.25	40.62	62.50	62.50
Qwen3.5-27B	Natural	34.38	65.62	87.50	50.00
Qwen3.5-27B	256	71.88	28.12	37.50	62.50
Qwen3.5-27B	128	62.50	37.50	37.50	100.00
Qwen3.5-27B	64	68.75	28.12	50.00	62.50
dots3-note-prev	Natural	65.62	34.38	62.50	37.50
dots3-note-prev	256	71.88	18.75	62.50	37.50
dots3-note-prev	128	65.62	34.38	75.00	25.00
dots3-note-prev	64	75.00	25.00	62.50	37.50

Prompt-paraphrase robustness

The comparison uses the 262 accepted Main paraphrases and their matched originals, with four samples each. Every reported change interval includes zero; no statistically detectable wording effect is observed in these comparisons, which does not prove exact equivalence.

Model	Metric	Original	Paraphrase	Change [95% CI]
DeepSeek V4 Pro	OAR	14.69	11.74	-2.96 [-6.30,+0.28]
DeepSeek V4 Pro	OAC@4	36.64	30.92	-5.73 [-11.89,+0.84]
GLM 5.3 Flash	OAR	17.37	19.47	+2.10 [-4.82,+9.06]
GLM 5.3 Flash	OAC@4	40.08	40.08	0.00 [-8.55,+8.56]
Kimi K3	OAR	16.70	15.65	-1.05 [-5.64,+3.78]
Kimi K3	OAC@4	38.55	33.21	-5.34 [-12.90,+2.53]
Jev 1.13	OAR	78.63	73.00	-5.63 [-22.27,+9.78]
Jev 1.13	OAC@4	84.35	81.30	-3.05 [-19.92,+12.89]

The published table cells and recomputed metrics accompany the full numerical summaries and reproduction code in this repository.

Columns

• id, prompt_id: stable prompt identifier (KAD also has conversation_id).
• family_id: physical family; KAD uses its scene/kernel ID.
• prompt: unchanged English question for single-turn configurations.
• system_prompt: Answer the user's question.
• messages: system/user records, ready for chat-completions APIs.
• population: animal or human in Main and paraphrase.
• larger_consequence_side, larger_consequence_entity: physical-comparison references where available. Main's larger-consequence side is ALTERNATIVE; Reverse test uses SUBJECT.
• matched_family_id: Neutral's matched Main family.
• pair_id, direction: normal/reversed pairing metadata.
• turns, probe_order, subject, alternative: KAD conversation specification.
• original_prompt: the Main question corresponding to each accepted paraphrase.
• schema_version: data schema version.

Neutral exposes its vulnerable-object description rather than assigning a living-subject label. It should be judged by the two objects named in that prompt.

DATA_MANIFEST.json records each Parquet file's row count, byte size, and SHA-256. Conversion preserves every prompt byte and identifier from the frozen release. These are test/evaluation sets, with no training split. Dataset content is CC BY 4.0; the standalone evaluation code is MIT-licensed in the linked code repository.

Third-party models and services

Unless explicitly stated otherwise, this repository does not include or distribute any third-party model weights. Users who independently download or enable third-party models or services must also comply with the applicable upstream project, model-weight, and data license terms. Use of third-party API services is subject to the applicable service providers' terms and conditions.

Citation

If you use ProtectionBench, please cite:

@misc{li2026protectionoverridesconsequencesunderstanding,
  title={When Protection Overrides Consequences: Understanding Protection Mode in Large Language Model Decision-Making},
  author={Yuanfan Li and Junyi Chen and Qi Zhou and Rui Yang and ZhangLiujie and Weihang Chen},
  year={2026},
  eprint={2610.19948},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2610.19948},
}