Prompt Optimization Is a Coin Flip: Diagnosing When It Helps in Compound AI Systems
Paper • 2604.14585 • Published
run_id string | cell_id string | trace_index int32 | task_name string | target_model string | actual_model string | optimizer string | seed int32 | baseline_prompt_file string | optimized_prompt string | trace string | completion_tokens int32 | is_correct bool | inclusion_threshold string | threshold_fallback string | baseline_acc float64 | baseline_std float64 | optimized_acc float64 | gap_std float64 | primary_tag string | all_tags list | rationale string | confidence float64 | source string | judge_actual_model string | judge_cost_usd float64 | manual_label_subset string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
cell_dspy_mipro_gsm8k_ds_v4flash_seed43_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed43 | 68 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 43 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let's break it down step by step.
**Step 1:** Initial likes after 1 week = 2000.
**Step 2:** Three weeks later, the number of likes was 70 times the initial number.
So,
\( 2000 \times 70 = 140,000 \) likes at that point.
**Step 3:** Recently, she received 20,000 more new likes.
So,
\( 140,000 + 20,000 = 160,... | 123 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | instruction_collapse | [
"instruction_collapse"
] | The optimized prompt drops the few-shot exemplars and the explicit answer format ("The answer is X") present in the baseline, replacing them with a generic instruction to "state the final answer." This causes the model to output a boxed answer without the expected extraction marker, and the lack of worked examples like... | 0.85 | llm_judge | deepseek-v4-flash | 0.000906 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed43_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed43 | 85 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 43 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let’s go step by step.
1. Josh starts with **3 balls** in the first week.
2. Each week he adds **1 ball**.
- Week 1: 3 balls
- Week 2: 4 balls
- Week 3: 5 balls
- Week 4: 6 balls
3. After the fourth week, he has **6 balls**.
4. On stage, he slips and drops **3 balls**.
- 6 − 3 = **3 bal... | 190 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | instruction_collapse | [
"instruction_collapse"
] | The optimized prompt drops the few-shot exemplars and the explicit answer format ("The answer is ...") present in the baseline, replacing them with a vague instruction to "state the final answer." The model's trace shows correct reasoning but outputs a boxed answer without the required "####" marker, causing the scorer... | 0.95 | llm_judge | deepseek-v4-flash | 0.000928 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed43_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed43 | 127 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 43 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let’s break it down step by step.
**Step 1: Understand the rates**
- Jack: 1 quarter = 20 minutes.
- Two worse friends: half as long → 1 quarter = 10 minutes each.
- One better friend: 1.5 times as long → 1 quarter = 30 minutes.
**Step 2: Total play time**
They play for 4 hours.
4 hours = 4 × 60 = 240 minut... | 300 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | over_specification | [
"over_specification"
] | The optimized prompt adds a rigid step-by-step procedure that assumes all players play simultaneously for the full duration, which works for many problems but fails on this input where the correct interpretation (likely sequential play or shared machine) yields a different answer. The prompt's structure forces a specif... | 0.85 | llm_judge | deepseek-v4-flash | 0.000928 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed43_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed43 | 137 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 43 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let’s go step by step.
1. Christina invited 16 friends.
2. She expects that 1/4 of them won’t show up.
That means she expects \( 16 \times \frac{1}{4} = 4 \) guests to be absent.
3. So the number of guests who actually attend is \( 16 - 4 = 12 \).
4. She needs 0.75 gift bags per attending guest.
That m... | 162 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | format_break | [
"format_break"
] | The model output uses a boxed LaTeX expression (\boxed{18}) instead of the expected plain-number format (e.g., '#### 18' or just '18'), so the answer extraction fails to parse the final answer correctly. The reasoning is otherwise correct, but the format mismatch causes the scorer to miss the answer, leading to underpe... | 0.95 | llm_judge | deepseek-v4-flash | 0.000885 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed43_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed43 | 144 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 43 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let's break it down step by step.
**Step 1: Time spent recording videos**
Britany records 18 videos per week, each 4 minutes long.
18 × 4 = 72 minutes per week.
**Step 2: Time spent writing songs**
She spends 2 hours per week writing songs.
2 hours = 120 minutes.
**Step 3: Time spent on makeup**
She spends... | 243 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | format_break | [
"format_break"
] | The optimized prompt instructs the model to 'state the final answer' but does not specify a machine-parseable format. The model outputs a boxed answer in LaTeX (oxed{1128 ext{ minutes}}) and an alternative in hours, which the GSM8K answer extractor cannot parse to a plain number. The reasoning is correct, but the sco... | 0.95 | llm_judge | deepseek-v4-flash | 0.000926 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed43_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed43 | 159 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 43 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let's break it down step by step.
**Step 1: Initial amounts**
- Blue shoe boxes: 7
- Red shoe boxes: 9
**Step 2: Blue boxes used**
He uses 3 blue shoe boxes.
Blue boxes left:
\( 7 - 3 = 4 \)
**Step 3: Red boxes used**
He uses \( \frac{1}{3} \) of his red shoe boxes.
\( \frac{1}{3} \times 9 = 3 \) red... | 181 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | format_break | [
"format_break"
] | The optimized prompt instructs the model to 'state the final answer' but does not specify a machine-parseable format. The model outputs a boxed expression '4 blue and 6 red' instead of a single number, causing the answer extractor to fail even though the reasoning is correct. The baseline prompt's few-shot examples imp... | 0.95 | llm_judge | deepseek-v4-flash | 0.000905 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed43_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed43 | 181 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 43 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let’s go step by step.
1. **Understand the time frame**
They worked for 3 weeks.
1 week = 7 days, so 3 weeks = \( 3 \times 7 = 21 \) days.
2. **Work per day**
Each day, they worked for \( \frac{3}{8} \) of a day.
Assuming a full day is 24 hours:
\( \frac{3}{8} \times 24 = 9 \) hours per day f... | 206 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | format_break | [
"format_break"
] | The optimized prompt instructs the model to 'state the final answer' but does not specify the required format (e.g., '#### N'). The model outputs a boxed answer, which the GSM8K answer-extraction parser cannot parse, causing a correct reasoning trace to be scored as incorrect. This format mismatch explains the accuracy... | 0.95 | llm_judge | deepseek-v4-flash | 0.000889 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed44_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed44 | 19 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 44 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let’s go step by step.
**Step 1:** Initial amount of water
Day 1: 4000 gallons.
**Step 2:** After a day of heavy rain, the amount doubles.
Day 2: \( 4000 \times 2 = 8000 \) gallons.
**Step 3:** On the third day, the volume increases by 6000 gallons from the previous day.
Day 3: \( 8000 + 6000 = 14000 \) gallon... | 153 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | instruction_collapse | [
"instruction_collapse"
] | The optimized prompt drops the few-shot exemplars and the explicit answer-formatting cue ("The answer is ...") from the baseline, replacing them with a vague instruction to "state the final answer." This causes the model to output a boxed expression and compute a total over three days instead of the single day's volume... | 0.95 | llm_judge | deepseek-v4-flash | 0.000897 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed44_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed44 | 24 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 44 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let's go step by step.
**Step 1: Find the initial number of corn.**
We know: "At Allan's house, there is twice as much corn as cannolis."
He has 40 cannolis initially.
So, initial corn = \( 2 \times 40 = 80 \).
**Step 2: Add the new cannolis.**
He buys 60 more cannolis.
New total cannolis = \( 40 + 60 = 100... | 220 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | instruction_collapse | [
"instruction_collapse"
] | The optimized prompt drops the few-shot exemplars and the explicit answer-formatting cue (the '####' marker) present in the baseline, reducing the task to a vague instruction. The model produces a correct reasoning chain but outputs a boxed answer instead of the required '#### N' format, causing the scorer to fail extr... | 0.95 | llm_judge | deepseek-v4-flash | 0.000925 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed44_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed44 | 44 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 44 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | We know the car’s initial price in 2007 is $20,000, and it depreciates by 21% each year.
Depreciation of 21% means the car retains \( 100\% - 21\% = 79\% \) of its value each year.
**Step 1: Determine the number of years from 2007 to 2010.**
\[
2010 - 2007 = 3 \text{ years}
\]
**Step 2: Apply the depreciation for... | 278 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | instruction_collapse | [
"instruction_collapse"
] | The optimized prompt drops the few-shot exemplars from the baseline, which provided worked examples of the required reasoning format and answer extraction pattern. Without these exemplars, the model produces a free-form solution with a boxed answer (\boxed{9860.78}) instead of the expected '#### 9860.78' format, and mo... | 0.9 | llm_judge | deepseek-v4-flash | 0.000981 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed44_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed44 | 120 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 44 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let’s break this down step by step.
**Step 1: Total monthly allowance**
Colby has $150 for the month.
**Step 2: What he already spent**
- Movies on Friday/Saturday: 5 movies × $10 = $50
- Movies on other days: 8 movies × $7 = $56
- Popcorn: 2 tubs × $8 = $16
- Candy: 4 boxes × $2 = $8
**Step 3: Total spe... | 278 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | format_break | [
"format_break"
] | The optimized prompt instructs the model to 'state the final answer' but does not specify a machine-parseable format like '#### N'. The model outputs a boxed answer (\boxed{14}) which the GSM8K answer extractor cannot parse, causing a false negative even though the reasoning is correct. The baseline prompt uses explici... | 0.95 | llm_judge | deepseek-v4-flash | 0.000949 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed44_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed44 | 124 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 44 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let’s go step-by-step.
**Step 1: Starting length**
The chalk starts at 5 inches.
**Step 2: Monday’s usage**
He used 45% of the chalk on Monday.
45% of 5 inches = \( 0.45 \times 5 = 2.25 \) inches used.
Remaining after Monday:
\( 5 - 2.25 = 2.75 \) inches.
**Step 3: Normal daily usage**
Normally he uses 2... | 270 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | over_specification | [
"over_specification"
] | The optimized prompt instructs the model to 'state the final answer' without specifying the required format (e.g., '#### N'), and the model outputs a boxed answer that the scorer cannot parse. The baseline prompt uses a consistent 'The answer is N' pattern that the scorer expects, but the optimized prompt's vague outpu... | 0.9 | llm_judge | deepseek-v4-flash | 0.000955 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed44_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed44 | 125 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 44 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let's break it down step by step.
**Step 1: Initial counts**
- Blue shoe boxes: 7
- Red shoe boxes: 9
**Step 2: Blue boxes used**
He uses 3 blue shoe boxes.
Blue boxes left:
\( 7 - 3 = 4 \)
**Step 3: Red boxes used**
He uses \( \frac{1}{3} \) of his red shoe boxes.
\( \frac{1}{3} \times 9 = 3 \) red ... | 181 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | format_break | [
"format_break"
] | The optimized prompt instructs the model to 'state the final answer' but does not specify a machine-parseable format. The model outputs a boxed expression '4 blue and 6 red' instead of the expected single number (10), causing the answer extraction to fail even though the reasoning is correct. | 0.95 | llm_judge | deepseek-v4-flash | 0.000881 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed44_1784252377 | cell_dspy_mipro_gsm8k_ds_v4flash_seed44 | 156 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 44 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let’s break it down step by step.
1. **Alexandra** raises $430.
Kim raises $320 more than Alexandra:
\( 430 + 320 = 750 \)
So Kim raises $750.
2. **Sarah** raises $300.
Maryam raises $400 more than Sarah:
\( 300 + 400 = 700 \)
So Maryam raises $700.
3. Now add all amounts:
Alexandr... | 201 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | format_break | [
"format_break"
] | The optimized prompt instructs the model to 'state the final answer' but does not specify a machine-parseable format (e.g., '#### 2180'). The model outputs a boxed answer using LaTeX oxed{2180}, which the GSM8K answer extractor likely fails to parse, causing a false negative even though the reasoning is correct. The b... | 0.95 | llm_judge | deepseek-v4-flash | 0.000924 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed46_1784252573 | cell_dspy_mipro_gsm8k_ds_v4flash_seed46 | 3 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 46 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let's break it down step by step.
**Step 1: Points in the first game**
Sasha scored 14 points.
Julie scored 4 fewer points than Sasha:
Julie’s first game points = 14 − 4 = 10.
**Step 2: Points in the second game**
Sasha scored 6 fewer points than Julie’s score in the first game.
Julie’s first game score was... | 165 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | format_break | [
"format_break"
] | The optimized prompt instructs the model to 'state the final answer' but does not specify a parseable format. The model outputs a boxed answer (\boxed{18}) instead of the expected '#### 18' marker, causing the answer extractor to fail even though the reasoning is correct. The baseline prompt uses explicit 'The answer i... | 0.95 | llm_judge | deepseek-v4-flash | 0.000917 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed46_1784252573 | cell_dspy_mipro_gsm8k_ds_v4flash_seed46 | 11 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 46 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let's go step-by-step.
**Step 1: Find the daily energy use before reduction.**
The AC uses 900 watts per hour.
If used for 8 hours a day:
\( 900 \times 8 = 7200 \) watt-hours per day.
**Step 2: Convert to kilowatt-hours.**
Since 1 kilowatt = 1000 watts:
\( 7200 \div 1000 = 7.2 \) kWh per day.
**Step 3: Fin... | 242 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | over_specification | [
"over_specification"
] | The optimized prompt instructs the model to 'state the final answer' without specifying the required format (e.g., '#### N'), and the model outputs a boxed answer. However, the scorer expects the answer in the baseline format (e.g., 'The answer is 81'). The trace shows correct reasoning but the final answer is wrong be... | 0.85 | llm_judge | deepseek-v4-flash | 0.000953 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed46_1784252573 | cell_dspy_mipro_gsm8k_ds_v4flash_seed46 | 36 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 46 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let’s go step by step.
**Step 1:** Initial amount of water
Day 1: 4000 gallons.
**Step 2:** After a day of heavy rain, the amount doubles.
Day 2: \( 4000 \times 2 = 8000 \) gallons.
**Step 3:** On the third day, the volume increases by 6000 gallons from the previous day.
Day 3: \( 8000 + 6000 = 14000 \) gallon... | 153 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | instruction_collapse | [
"instruction_collapse"
] | The optimized prompt drops the few-shot exemplars and the explicit answer format ("The answer is X") from the baseline, replacing them with a vague instruction to "state the final answer." The model then interprets the task as computing the total over three days rather than the amount on the third day, which is the act... | 0.9 | llm_judge | deepseek-v4-flash | 0.000913 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed46_1784252573 | cell_dspy_mipro_gsm8k_ds_v4flash_seed46 | 62 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 46 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let the cost of the lot be \( x \) dollars.
Then the house costs \( 3x \) dollars.
Together, they cost $120,000:
\[
x + 3x = 120{,}000
\]
\[
4x = 120{,}000
\]
\[
x = 30{,}000
\]
So the house costs:
\[
3x = 3 \times 30{,}000 = 90{,}000
\]
**Final answer:**
\[
\boxed{90{,}000}
\] | 122 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | format_break | [
"format_break"
] | The model output uses a boxed LaTeX expression (\boxed{90{,}000}) instead of the expected plain number or '#### 90000' format. The scorer likely fails to extract the answer from this non-standard format, causing a correct reasoning trace to be marked incorrect. | 0.95 | llm_judge | deepseek-v4-flash | 0.000854 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed46_1784252573 | cell_dspy_mipro_gsm8k_ds_v4flash_seed46 | 79 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 46 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let’s go step-by-step.
1. There are 54 students total.
2. They are separated into 6 equal groups.
So each group has:
\( 54 \div 6 = 9 \) students per group.
3. The activity requires 12 groups.
Currently, there are 6 groups.
4. Number of additional groups needed:
\( 12 - 6 = 6 \).
**Final answe... | 107 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | over_specification | [
"over_specification"
] | The optimized prompt instructs the model to 'state the final answer' without specifying the required format (e.g., '#### N'), and the model outputs a boxed answer. The scorer likely expects the '####' marker, so the correct reasoning is not extracted. This is a format break, but the root cause is the prompt's over-spec... | 0.85 | llm_judge | deepseek-v4-flash | 0.000883 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed46_1784252573 | cell_dspy_mipro_gsm8k_ds_v4flash_seed46 | 88 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 46 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let’s break it down step by step.
1. **Alexandra** raises $430.
Kim raises $320 more than Alexandra:
\( 430 + 320 = 750 \)
So Kim raises $750.
2. **Sarah** raises $300.
Maryam raises $400 more than Sarah:
\( 300 + 400 = 700 \)
So Maryam raises $700.
3. Now add all amounts:
Alexandr... | 203 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | instruction_collapse | [
"instruction_collapse"
] | The optimized prompt drops the few-shot exemplars and the explicit answer format ("The answer is ...") present in the baseline, leaving only a vague instruction to "state the final answer." The model outputs a boxed answer instead of the expected plain number, and the lack of structured reasoning guidance leads to an a... | 0.85 | llm_judge | deepseek-v4-flash | 0.000918 | null |
cell_dspy_mipro_gsm8k_ds_v4flash_seed46_1784252573 | cell_dspy_mipro_gsm8k_ds_v4flash_seed46 | 117 | gsm8k | deepseek-v4-flash | deepseek-v4-flash | dspy_mipro | 46 | null | Solve a grade-school math problem step by step, then state the final answer.
Q: {question}
A: | Let's go step-by-step.
**Step 1: Total land**
Abraham starts with 80 square meters.
**Step 2: First sale**
He sold half of the land.
Half of 80 = 40 square meters.
He sold this for $50.
**Step 3: Second sale**
After that, he sold another 1/4 of his land.
1/4 of 80 = 20 square meters.
He sold this for $... | 232 | false | d1b | null | 0.969 | 0.004183 | 0.965 | 0.956183 | instruction_collapse | [
"instruction_collapse"
] | The optimized prompt drops the few-shot exemplars from the baseline, which provided worked solutions and a clear answer format ("The answer is X"). Without these exemplars, the model produces a correct reasoning chain but outputs the answer in a \(\boxed{140}\) format that the scorer cannot parse, leading to a false ne... | 0.9 | llm_judge | deepseek-v4-flash | 0.00094 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 1 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that overfit to the validation examples. The representative trace shows a news article about Google's business success being misclassified as 'Business' instead of 'Sci/Tech', likely because the added rul... | 0.85 | llm_judge | deepseek-v4-flash | 0.000746 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 3 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that overfit the validation distribution. The representative trace shows a news article about Apple's financial earnings being misclassified as 'Business' instead of 'Sci/Tech', likely because the added r... | 0.85 | llm_judge | deepseek-v4-flash | 0.000736 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 7 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Sci/Tech | 3 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'World: international politics, diplomacy...') that over-constrain the classification. The trace shows a news article about an elephant receiving therapy—a human-interest/oddity story—which does not fit neatly into the provided subcategories, causing the... | 0.85 | llm_judge | deepseek-v4-flash | 0.000748 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 11 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory bullet points (e.g., 'not stock news about the company') that over-constrain the classification. The trace shows a news article about a company (SAP) winning a government contract, which the model classifies as 'Business' because it involves a corporate contract, but the g... | 0.85 | llm_judge | deepseek-v4-flash | 0.000758 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 15 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Sci/Tech | 3 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'World: international politics, diplomacy...') that overfit to the validation examples and cause the model to misclassify a news article about a study on online drug purchases as Sci/Tech instead of World, because the article's mention of 'Internet' and ... | 0.85 | llm_judge | deepseek-v4-flash | 0.00073 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 28 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Article: Alternative Energy: Ready or not? Even with a boost from higher oil prices and growing concern about global warming, the payoff on most alternative energy technologies seems a ways off.
Topic: **Sci/Tech** | 43 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory bullet points that over-specify the boundaries of each label. In the trace, an article about alternative energy economics is classified as Sci/Tech because the bullet for Sci/Tech includes 'engineering breakthroughs' and 'weather science', but the article is actually a Bus... | 0.85 | llm_judge | deepseek-v4-flash | 0.000743 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 29 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Sports | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'World: international politics, diplomacy...') that overfit to the validation examples. The trace shows a news article about betting on a charity single being misclassified as 'Sports' because the prompt's 'Sports' definition includes 'games, matches, at... | 0.85 | llm_judge | deepseek-v4-flash | 0.000759 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 32 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Sci/Tech | 3 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that over-constrain the classification. The trace shows a news article about consumer spyware being misclassified as Sci/Tech instead of Business, likely because the added Sci/Tech definition includes 'so... | 0.85 | llm_judge | deepseek-v4-flash | 0.000744 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 49 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'World: international politics, diplomacy, conflicts...') that overfit to the validation examples. The trace shows a news article about a debt ceiling increase being misclassified as 'Business' instead of 'World', likely because the prompt's subcategory ... | 0.85 | llm_judge | deepseek-v4-flash | 0.000754 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 51 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that overfit to the validation distribution. The trace shows a news article about a business deal involving a tech company (SAP) being misclassified as 'Business' instead of 'Sci/Tech', likely because the... | 0.85 | llm_judge | deepseek-v4-flash | 0.000745 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 54 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory bullet points that over-specify the classification boundaries. In the trace, an article about computer screen prices is misclassified as 'Business' because the bullet for Business includes 'stock news, M&A' and the article is from Reuters, but the article is actually about... | 0.85 | llm_judge | deepseek-v4-flash | 0.000733 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 57 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Sports | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'World: international politics, diplomacy...') that overfit to the validation examples. The trace shows a tennis article about the Australian Open being classified as 'Sports' when the gold label is 'World', likely because the article's focus on a tourna... | 0.85 | llm_judge | deepseek-v4-flash | 0.000751 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 65 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions for each label (e.g., 'World: international politics, diplomacy, conflicts...'). While these definitions help on typical examples, they can mislead the model on borderline cases: the article about oil prices and stocks is classified as 'Business' because it men... | 0.85 | llm_judge | deepseek-v4-flash | 0.000745 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 73 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that overfit to the validation examples. This causes the model to misclassify a news article about IBM acquiring a tech company as 'Business' instead of 'Sci/Tech', because the added rule incorrectly excl... | 0.85 | llm_judge | deepseek-v4-flash | 0.000725 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 79 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Sci/Tech | 3 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that overfit to the validation distribution. The representative trace shows a drug-safety article about Merck and Vioxx, which the model misclassifies as Sci/Tech (likely because it involves medical resea... | 0.85 | llm_judge | deepseek-v4-flash | 0.000788 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 87 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'World: international politics, diplomacy...') that overfit to the validation distribution. The trace shows a news article about Alstom signing contracts in China—a business-related event with geopolitical implications—which the model misclassifies as 'B... | 0.85 | llm_judge | deepseek-v4-flash | 0.000744 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 90 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'not stock news about the company') that over-constrain the classification. This causes the model to misclassify a news article about Intel's flash market gains as 'Business' instead of 'Sci/Tech', because the added rule incorrectly excludes tech-company... | 0.85 | llm_judge | deepseek-v4-flash | 0.000739 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 92 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'not stock news about the company') that overfit to the validation distribution. The trace shows a flat-panel LCD factory article being misclassified as 'Business' because the added rule incorrectly excludes such tech-manufacturing news from Sci/Tech, ca... | 0.85 | llm_judge | deepseek-v4-flash | 0.000715 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 94 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'World: international politics, diplomacy...') that overfit to the validation examples. The trace shows a news article about an Internet pharmacy closing due to economic factors—the model outputs 'Business' because the prompt's Business definition includ... | 0.85 | llm_judge | deepseek-v4-flash | 0.000757 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 109 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that overfit to the validation distribution. The trace shows a news article about Tokyo stocks being classified as 'Business' when the gold label is 'World', likely because the added specification incorre... | 0.85 | llm_judge | deepseek-v4-flash | 0.000724 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 113 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Sci/Tech | 3 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that overfit to the validation distribution. The representative trace shows a news article about Internet-based video and copyright issues, which the model classified as Sci/Tech due to the 'technology pr... | 0.85 | llm_judge | deepseek-v4-flash | 0.000749 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 118 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that overfit to the validation distribution. The trace shows a business-related article about Salesforce.com's profit being misclassified as 'Business' instead of 'Sci/Tech', likely because the added rule... | 0.85 | llm_judge | deepseek-v4-flash | 0.000732 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 126 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory bullet points (e.g., 'World: international politics, diplomacy...') that over-constrain the classification. The trace shows a news article about a legal ruling in a tech takeover case being misclassified as 'Business' because the bullet points emphasize 'M&A' under Busines... | 0.85 | llm_judge | deepseek-v4-flash | 0.000788 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 132 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Sci/Tech | 3 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that overfit to the validation distribution. The representative trace shows a news article about a sex toy shutting down an airport — a 'World' story — but the model outputs 'Sci/Tech', likely because the... | 0.85 | llm_judge | deepseek-v4-flash | 0.000761 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 135 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that overfit to the validation distribution. The representative trace shows a news article about a business appointment (Kellogg CEO nominated for Commerce Secretary) being misclassified as 'World' becaus... | 0.85 | llm_judge | deepseek-v4-flash | 0.000758 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 136 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that over-constrain the classification. The representative trace shows a Sci/Tech article about vegetarianism being misclassified as World, likely because the added Sci/Tech definition excludes non-stock,... | 0.85 | llm_judge | deepseek-v4-flash | 0.000737 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 141 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'World: international politics, diplomacy...') that overfit to the validation examples. The trace shows the model classifying a cybersecurity resignation as 'World' because the definitions bias it toward geopolitical framing, whereas the gold label is 'S... | 0.85 | llm_judge | deepseek-v4-flash | 0.000728 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 160 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that over-constrain the classifier. The trace shows a Forbes article about a marketing campaign using text messages and bottle caps—a topic that could be interpreted as Business (marketing) or Sci/Tech (t... | 0.85 | llm_judge | deepseek-v4-flash | 0.000771 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 183 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'Sci/Tech: ... not stock news about the company') that overfit to the validation distribution. In the trace, an article about TheStreet.com being up for sale involves financial news and a company, which the prompt's Sci/Tech definition explicitly exclude... | 0.85 | llm_judge | deepseek-v4-flash | 0.000752 | null |
cell_gepa_ag_news_ds_v4flash_seed44_1784252660 | cell_gepa_ag_news_ds_v4flash_seed44 | 193 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 44 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
- World: international politics, diplomacy, conflicts, treaties, elections, geopolitics.
- Sports: games, matches, athletes, leagues, tournaments, scores.
- Business: finance, markets, corporate earnings, regulation, stock news, M&A.
- Sci/Tech:... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.85 | 0.67082 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed subcategory definitions (e.g., 'World: international politics, diplomacy...') that overfit to the validation distribution. The representative trace shows a news article about online political behavior (voter site preferences) which is not clearly covered by the rigid subcategory list,... | 0.85 | llm_judge | deepseek-v4-flash | 0.000754 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 9 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Sports | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (about Google's quantum computing chip) that is not present in the baseline. This extra example is specific to Sci/Tech and does not help generalization; it may bias the model toward Sci/Tech predictions or away from World topics. The trace shows a sports-related article being ... | 0.85 | llm_judge | deepseek-v4-flash | 0.000715 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 10 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Sports | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (about a quantum computing chip) that is not present in the baseline. This extra example is from the Sci/Tech category, which may bias the model toward Sci/Tech predictions or away from World/Sports/Business in ambiguous cases. The representative trace shows a sports-related ar... | 0.85 | llm_judge | deepseek-v4-flash | 0.000727 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 29 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip → Sci/Tech) that is not present in the baseline. This extra example biases the model toward Sci/Tech predictions and does not improve generalization, as shown by the misclassification of a World article as Business. The added example leaks a speci... | 0.85 | llm_judge | deepseek-v4-flash | 0.00071 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 39 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Sports | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip → Sci/Tech) that is not present in the baseline. This extra example overfits to a specific Sci/Tech pattern (technology company + product), which does not generalize to the test article about Sandia Motor Speedway being sold on eBay — a story that... | 0.9 | llm_judge | deepseek-v4-flash | 0.000747 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 47 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Sports | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip → Sci/Tech) that is not present in the baseline. This extra example overfits to a specific Sci/Tech pattern from the validation set, but on the test article about BASS moving headquarters (which is actually Business, not Sports), the model is bias... | 0.85 | llm_judge | deepseek-v4-flash | 0.000734 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 52 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is specific to Sci/Tech, which biases the model toward Sci/Tech predictions for articles that are not clearly Sci/Tech, as seen in the trace where a Microsoft-Time Warner deal is misclassified as Business instead of Sci/Tech. This added exam... | 0.9 | llm_judge | deepseek-v4-flash | 0.000715 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 61 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is also Sci/Tech, reinforcing the existing Sci/Tech pattern from the baseline. However, the representative trace shows a Business article about Microsoft licensing fees being misclassified as Business when the gold label is Sci/Tech, suggest... | 0.7 | llm_judge | deepseek-v4-flash | 0.000745 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 62 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example about a quantum computing chip labeled Sci/Tech, which overfits to the validation set's pattern of technology-related articles. The representative trace shows a Sci/Tech article about global warming and hurricanes being misclassified as World, indicating the added example did n... | 0.85 | llm_judge | deepseek-v4-flash | 0.000703 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 70 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is specific to Sci/Tech, which overfits the validation distribution and does not generalize. The representative trace shows a Sci/Tech article misclassified as Business, indicating the added example skewed the model's decision boundary. | 0.85 | llm_judge | deepseek-v4-flash | 0.000691 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 74 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example ('Google unveils new quantum computing chip with 1000 qubits. Topic: Sci/Tech') that is not present in the baseline. This extra example overfits the prompt to the validation distribution, making the model more likely to classify tech-related articles as Sci/Tech but not improvi... | 0.85 | llm_judge | deepseek-v4-flash | 0.000735 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 78 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is not present in the baseline. This extra example is a Sci/Tech article, which may bias the model toward predicting Sci/Tech for similar articles but does not help with the actual test distribution. The representative trace shows a news art... | 0.85 | llm_judge | deepseek-v4-flash | 0.000733 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 79 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example ('Google unveils new quantum computing chip with 1000 qubits. Topic: Sci/Tech') that is not present in the baseline. This extra example overfits to a specific Sci/Tech pattern (tech company + product announcement), which does not generalize to the test article about Unisys layo... | 0.9 | llm_judge | deepseek-v4-flash | 0.000749 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 82 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Sports | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip → Sci/Tech) that is not present in the baseline. This extra example overfits to a specific Sci/Tech pattern (quantum computing) and does not improve generalization; the representative trace shows a sports-related article being misclassified as Spo... | 0.85 | llm_judge | deepseek-v4-flash | 0.00072 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 91 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is specific to Sci/Tech, which overfits the validation distribution and does not generalize. The representative trace shows a Sci/Tech article (about hawks being evicted) being misclassified as World, suggesting the added example biased the ... | 0.85 | llm_judge | deepseek-v4-flash | 0.000717 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 98 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds an extra example about Google's quantum computing chip labeled Sci/Tech, which overfits to the validation set. The representative trace shows a news article about Google founders selling stock, which the model misclassifies as Business (likely because the added example biases it toward Sci/Tec... | 0.85 | llm_judge | deepseek-v4-flash | 0.000722 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 99 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Sports | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is not present in the baseline. This extra example biases the model toward Sci/Tech predictions and does not improve generalization; the representative trace shows a sports article misclassified as Sports when the gold label is World, indica... | 0.85 | llm_judge | deepseek-v4-flash | 0.000704 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 100 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | trivial_restatement | [
"trivial_restatement"
] | The optimized prompt is nearly identical to the baseline, only adding one extra example (Google quantum computing chip) that does not improve generalization and may slightly shift the distribution. The core instruction and format are unchanged, so the underperformance is due to a trivial restatement that fails to add t... | 0.9 | llm_judge | deepseek-v4-flash | 0.000693 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 110 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Sports | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example ('Google unveils new quantum computing chip with 1000 qubits. Topic: Sci/Tech') that is not present in the baseline. This extra example overfits the prompt to the validation distribution, making the model more likely to classify tech-related articles as Sci/Tech while not impro... | 0.85 | llm_judge | deepseek-v4-flash | 0.000731 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 125 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is specific to Sci/Tech, which overfits the validation distribution and does not generalize. The representative trace shows a Business prediction for a Sci/Tech article, indicating the added example skewed the model's behavior away from the ... | 0.85 | llm_judge | deepseek-v4-flash | 0.000695 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 126 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is very similar to the test article about PayPal technical issues, both being Sci/Tech. This overfits the prompt to the validation distribution, causing the model to misclassify a Sci/Tech article as Business because the added example does n... | 0.85 | llm_judge | deepseek-v4-flash | 0.000702 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 132 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is not present in the baseline. This extra example is a Sci/Tech article, which may bias the model toward Sci/Tech predictions for similar tech-related articles, but the representative trace shows a tech-related article (Cisco/Fujitsu networ... | 0.85 | llm_judge | deepseek-v4-flash | 0.00074 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 135 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Sports | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example ('Google unveils new quantum computing chip with 1000 qubits. Topic: Sci/Tech') that is not present in the baseline. This extra example overfits the prompt to the validation distribution, causing the model to bias toward Sci/Tech or to rely on surface patterns from the added ex... | 0.85 | llm_judge | deepseek-v4-flash | 0.000752 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 137 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Sports | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (about Google's quantum chip) that is not present in the baseline. This extra example biases the model toward Sci/Tech topics, as seen in the trace where a cricket article about a player missing a test match is misclassified as Sports instead of World. The added example does no... | 0.95 | llm_judge | deepseek-v4-flash | 0.000704 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 138 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is not present in the baseline. This extra example is from the Sci/Tech category, which may bias the model toward predicting Sci/Tech for articles that are similar in topic (e.g., technology-related), but the representative trace shows a mis... | 0.85 | llm_judge | deepseek-v4-flash | 0.000726 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 144 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is not present in the baseline. This extra example is a Sci/Tech article, which may bias the model toward predicting Sci/Tech for similar tech-related articles, but the representative trace shows a Business article being misclassified as Sci... | 0.7 | llm_judge | deepseek-v4-flash | 0.000856 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 158 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip → Sci/Tech) that is not present in the baseline. This extra example overfits to a specific Sci/Tech pattern (technology company + new product), which does not generalize to the test article about Thomson and ContentGuard (a business/tech crossover... | 0.85 | llm_judge | deepseek-v4-flash | 0.000731 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 164 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum chip → Sci/Tech) that is not present in the baseline. This extra example overfits to a specific Sci/Tech pattern, but the representative trace shows a Business article misclassified as World, indicating the added example did not help generalization and may have ... | 0.85 | llm_judge | deepseek-v4-flash | 0.0007 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 170 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is not present in the baseline. This extra example is a Sci/Tech article, which may bias the model toward predicting Sci/Tech for similar inputs, but the representative trace shows a Business article misclassified as World. The added example... | 0.85 | llm_judge | deepseek-v4-flash | 0.000716 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 176 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip) that is specific to Sci/Tech, which biases the model toward that label for technology-related articles. However, the representative trace shows a business-related article about IBM hiring being misclassified as Business when the gold label is Sci... | 0.8 | llm_judge | deepseek-v4-flash | 0.000743 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 196 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | World | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt is nearly identical to the baseline but adds a fifth example ('Google unveils new quantum computing chip with 1000 qubits. Topic: Sci/Tech') that over-represents the Sci/Tech category. This biases the model toward Sci/Tech predictions and does not improve generalization, as evidenced by the misclas... | 0.9 | llm_judge | deepseek-v4-flash | 0.000706 | null |
cell_gepa_ag_news_ds_v4flash_seed45_1784252712 | cell_gepa_ag_news_ds_v4flash_seed45 | 197 | ag_news | deepseek-v4-flash | deepseek-v4-flash | gepa | 45 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Article: Diplomats meet at UN to discuss climate policy.
Topic: World
Article: Lakers beat Celtics in overtime thriller.
Topic: Sports
Article: Federal Reserve hints at interest rate hike next quarter.
Topic: Business
Article: New AI model clai... | Business | 1 | false | d1b | null | 0.862 | 0.017889 | 0.845 | 0.950329 | exemplar_leak | [
"exemplar_leak"
] | The optimized prompt adds a fifth example (Google quantum computing chip → Sci/Tech) that is not present in the baseline. This extra example biases the model toward Sci/Tech predictions on test articles that mention technology, but does not improve overall accuracy and may cause misclassifications like the representati... | 0.85 | llm_judge | deepseek-v4-flash | 0.000701 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 2 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | Topic: Business
This article discusses a corporate asset sale and financial transaction between two telecommunications companies, which falls under corporate finance and regulatory/business dealings rather than technology product innovation. | 38 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | format_break | [
"format_break"
] | The optimized prompt's added definitions and examples cause the model to output a verbose explanation after the label, breaking the expected 'Topic: <label>' format. The scorer likely expects only the label after 'Topic:', and the extra text prevents correct extraction, even though the predicted label 'Business' is cor... | 0.9 | llm_judge | deepseek-v4-flash | 0.000726 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 3 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | **Business**
This article focuses on corporate finance and regulatory issues - specifically a pharmaceutical company's decision to sponsor a clinical trial in response to regulatory concerns about drug safety. While it involves a medical/pharmaceutical product, the news angle is about the business decision, corporate ... | 77 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed definitions that over-specify the Sci/Tech category, explicitly excluding technology company products and digital innovations from Business. This causes the model to misclassify a business-related pharmaceutical trial as Business (which is correct) but the trace shows the model's reas... | 0.7 | llm_judge | deepseek-v4-flash | 0.000817 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 14 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | **Topic: Business**
This article discusses labor market trends, employment predictions, and economic policy implications regarding the future job market for unskilled workers in Britain. While it touches on economic forecasting, it's fundamentally about labor economics, workforce trends, and business/employment policy... | 76 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed definitions and extra examples that over-specify the Sci/Tech category, explicitly excluding technology company products and digital innovations from Business. This causes the model to misclassify a Business article about labor market trends as Business, but the trace shows the model'... | 0.7 | llm_judge | deepseek-v4-flash | 0.000825 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 17 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | **Topic: Business**
This article focuses on corporate finance and a major acquisition deal (Standard Chartered's bid for a majority stake in Bank Permata), which falls under business/corporate finance rather than technology company developments. | 49 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | format_break | [
"format_break"
] | The optimized prompt's added definitions and examples do not cause the model to produce a bare label; instead, the model outputs a bolded label followed by an explanatory sentence. The scorer likely expects only the label (e.g., 'Business') and fails to extract it from the verbose response, leading to a false negative ... | 0.9 | llm_judge | deepseek-v4-flash | 0.000762 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 21 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | This article appears to be incomplete or corrupted (it cuts off mid-sentence with "they just didn #39;t smell right").
Based on the fragment provided, it seems to be a personal narrative or lifestyle piece about Christmas trees, which doesn't clearly fit into any of the four categories (World, Sports, Business, Sci/T... | 95 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | instruction_collapse | [
"instruction_collapse",
"format_break"
] | The optimized prompt's added definitions and examples over-emphasize the 'Sci/Tech' category and introduce a restrictive exclusion ('but NOT technology company products...') that misleads the model on a Business article about a retail trend. Additionally, the model's response is a meta-commentary refusing to classify, ... | 0.85 | llm_judge | deepseek-v4-flash | 0.000783 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 22 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | **Topic: Business**
This article focuses on corporate financial performance (sales figures, earnings) and insurance industry operations, which are core business topics. While it mentions a tech company (Aviva), the article is about their financial results and pension sales—not about technology products or digital inno... | 63 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | over_specification | [
"over_specification"
] | The optimized prompt adds a detailed definition for Business that explicitly excludes technology company products, platforms, or digital innovations, and includes a counter-example (Sony, tech startup) that biases the model to classify any mention of a tech company as Sci/Tech. In the trace, the model correctly identif... | 0.85 | llm_judge | deepseek-v4-flash | 0.000802 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 25 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | Topic: World
This article deals with a human interest/social issue involving an individual and their family situation. While it doesn't fit neatly into the primary categories, it falls closest to "World" as it covers a general news event of international/human interest nature rather than sports, business, or technolog... | 70 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed definitions for each category, particularly for Business and Sci/Tech, which over-constrain the classification. The trace shows the model struggling with a human-interest story, incorrectly forcing it into 'World' because the definitions don't cover such cases, leading to a wrong clas... | 0.8 | llm_judge | deepseek-v4-flash | 0.000755 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 26 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | **Sci/Tech**
This article describes a software product launch and its features (network monitoring, compliance tools, security capabilities). It focuses on a technology company's product announcement and digital innovation, which falls under Sci/Tech rather than Business. | 54 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed definitions that over-specify the Sci/Tech category, explicitly excluding technology company products from Business and including them in Sci/Tech. This causes the model to classify a software product launch as Sci/Tech, but the gold label is Business, indicating the added constraints... | 0.85 | llm_judge | deepseek-v4-flash | 0.000766 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 32 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | Topic: Business
This article discusses a corporate acquisition (UBS buying SoundView from Charles Schwab) involving financial terms and market positioning. While it involves financial/banking companies, the focus is on the corporate transaction, deal value, and business strategy rather than technology products or digi... | 63 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed definitions that over-specify the Business category by explicitly excluding technology-related content ('but NOT technology company products, platforms, or digital innovations'). This causes the model to overthink and misclassify a straightforward business acquisition as not Business,... | 0.85 | llm_judge | deepseek-v4-flash | 0.000756 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 33 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | Topic: Business
This article discusses economic policy and economic conditions (strong dollar, oil prices, drought effects) affecting a country's economy, which falls under Business rather than World politics/diplomacy. | 43 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed definitions that over-narrow the Business category to exclude technology-related content and over-emphasize economic policy, causing the model to misclassify a World news article about a country's economic outlook as Business. The added specificity works for the validation distributio... | 0.85 | llm_judge | deepseek-v4-flash | 0.000743 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 35 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | Topic: **Sci/Tech**
This article discusses hiring trends in the IT services industry and job market conditions for tech workers. While it touches on employment, the focus is on the technology sector and tech worker skills/careers, making it a tech industry development rather than general business/economic policy. | 64 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed definitions that over-specify the Sci/Tech category, explicitly excluding technology company products and digital innovations from Business and including tech industry developments in Sci/Tech. This causes the model to misclassify a business/employment article about IT services hiring... | 0.85 | llm_judge | deepseek-v4-flash | 0.000771 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 37 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | **Sci/Tech**
This article discusses a security vulnerability in the Linux kernel, which is a technology product/platform issue. It covers a technical security breach in software, making it a technology industry development rather than a business, sports, or world news story. | 57 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | format_break | [
"format_break"
] | The optimized prompt's added definitions and examples do not change the classification task, but the model's output includes a bolded label followed by an explanatory sentence. The scorer likely expects the bare label 'Sci/Tech' and fails to extract it from the verbose response, causing a correct classification to be m... | 0.9 | llm_judge | deepseek-v4-flash | 0.000757 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 38 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | **Topic: Business**
This article discusses Nissan's production suspensions and financial impact due to parts shortages, which are corporate operational and financial matters. While it involves a manufacturing company, the focus is on production decisions, supply chain issues, and earnings impact—core business concerns... | 70 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed definitions and extra examples that over-specify the Sci/Tech category, explicitly excluding technology company products and digital innovations from Business. This causes the model to misclassify a business-related article about Nissan's production suspensions as Business (which is c... | 0.8 | llm_judge | deepseek-v4-flash | 0.000813 | null |
cell_gepa_ag_news_haiku45_seed43_1785570814 | cell_gepa_ag_news_haiku45_seed43 | 47 | ag_news | claude-haiku-4-5-20251001 | claude-haiku-4-5-20251001 | gepa | 43 | null | Classify each news article into one of: World, Sports, Business, Sci/Tech.
Definitions:
- World: International politics, conflicts, diplomacy, global events, international relations
- Sports: Athletic competition, teams, players, game results, sports organizations
- Business: Corporate finance, markets, earnings, econ... | **Sci/Tech**
This article discusses Microsoft's software update (SP2 for Windows XP), which is a technology product announcement and development. Even though it mentions a delay and corporate concerns about installation, the core subject is a tech company's software release, making it a Sci/Tech classification rather ... | 70 | false | d1b | null | 0.771 | 0.023292 | 0.7 | 3.048305 | over_specification | [
"over_specification"
] | The optimized prompt adds detailed definitions that over-specify the Sci/Tech category, explicitly excluding technology company products from Business and including tech company announcements in Sci/Tech. This causes the model to classify a Microsoft software update as Sci/Tech, but the gold label is Business, indicati... | 0.85 | llm_judge | deepseek-v4-flash | 0.000752 | null |
End of preview. Expand in Data Studio
APO Failure-Mode Dataset
A labelled dataset of failure modes from Automated Prompt Optimization (APO) methods. Each record is one (run, held-out evaluation trace) tuple from a sweep cell where the optimised prompt underperformed the manual baseline by at least 0.5 standard deviations (the D1b inclusion criterion described below). Records carry one primary failure-mode tag from a locked 5-tag taxonomy plus zero or more secondary tags.
This dataset is the primary deliverable of an ongoing USTH Master's thesis (defence target September 2026) titled Automated Prompt Optimization for Large Language Model Applications Using DSPy and OPRO. The dataset is the empirical substrate for Claim 1 of the thesis (descriptive — when APO underperforms, what kind of failed prompt does it produce?).
Dataset Summary
- Records: 1,695 per-trace labels across 27 included runs from 27 sweep cells.
- Sweep substrate: 32 realised cells (
optimiser × dataset × model) at N=5 seeds — the complete Path-B grid, closed out 2026-08-02/03 — across 3 optimisers, 5 datasets, and 3 model families (see Dataset Structure below). - Annotation pipeline: a locked LLM-as-judge classifier (DeepSeek V4-flash) running the prompt at
prompts/failure_mode_judge.txtof the source repository, with a deterministic short-circuit fortrivial_restatementcases (similarity ≥ 0.95 to the manual baseline). - Inclusion criterion: D1b (effect-size). A run is included iff
optim_acc < baseline_acc − 0.5 × baseline_stdwherebaseline_stdis the std across the N=5 seeds of the matchingoptimizer="none"cell. Fallback to D1a (point estimate) when the baseline cell has fewer than 3 successful seeds;threshold_fallbackflags these records. - Total judge cost for the labels: $0.98 USD (DeepSeek V4-flash, max_tokens=4096).
- Status: v0.2, public release (2026-08-30). Label validity is measured four ways (see Annotators): test–retest decoding stability κ = 0.925 (passes the pre-registered κ ≥ 0.7 gate), score-blind κ = 0.650, a 79% false-positive anchoring disposition, and the pre-registered seal — a blind human annotation of the frozen 50-trace audit form, measured 2026-08-02 at author–judge κ = 0.925 (47/50 agreement; author-rater under a blind protocol, judge labels withheld). The human audit that gated public release is cleared. See Considerations.
Supported Tasks
- Failure-mode classification. Given an (optimised prompt, baseline prompt, held-out trace, question, gold) tuple, predict the primary failure-mode tag from {
trivial_restatement,instruction_collapse,exemplar_leak,over_specification,format_break}. - APO reliability auditing. Each cell's underperformance pattern is a within-cell descriptive statistic; the dataset can be used to compare optimiser families' failure-mode profiles (e.g.\ MIPROv2's
instruction_collapse + format_breakvs.\ GEPA'strivial_restatement + instruction_collapse + exemplar_leakvs.\ OPRO's 100%trivial_restatement).
Languages
English (en) only. All prompts, traces, and rationales are English; the underlying APO benchmarks (GSM8K, HotpotQA, SST-5, AG News, HumanEval) are English-only.
Dataset Structure
Data Fields
| field | type | description |
|---|---|---|
run_id |
string | Cell run identifier (cell_<optimizer>_<task>_<model>_seed<N>_<unix_ts>). |
cell_id |
string | Cell identifier without timestamp (groups the 5 seeds of one cell). |
trace_index |
int32 | Position of this trace within the run's held-out evaluation slice. |
task_name |
string | One of gsm8k, hotpotqa, sst5, ag_news, humaneval. |
target_model |
string | The held-out evaluator model id (e.g.\ deepseek-v4-flash). |
actual_model |
string | The provider-returned model id per the X.PIN.1 audit trail. |
optimizer |
string | One of dspy_mipro, opro, gepa. (Baseline none cells are not in the dataset.) |
seed |
int32 | Seed from {42, 43, 44, 45, 46}. |
baseline_prompt_file |
string | Path of the frozen baseline file (prompts/wei2022_gsm8k.txt for GSM8K; prompts/<task>_baseline.txt for the others, per X.BASELINE.1). |
optimized_prompt |
string | The optimised prompt template the held-out evaluation ran with. |
trace |
string | The model's raw completion on this held-out example. |
completion_tokens |
int32 | Token count of the completion (per the provider's tokenizer). |
is_correct |
bool | Did the trace score correct on this example under the task's primary metric? |
inclusion_threshold |
string | Either d1b (effect-size) or d1a (fallback). |
threshold_fallback |
string | d1a iff the baseline cell had < 3 seeds; null otherwise. |
baseline_acc |
float64 | Mean accuracy across the matching none cell's N=5 seeds. |
baseline_std |
float64 | Std across the matching none cell's N=5 seeds. |
optimized_acc |
float64 | This run's accuracy on the held-out slice. |
gap_std |
float64 | (baseline_acc − optimized_acc) / baseline_std. |
primary_tag |
string | The primary failure-mode tag. One of {trivial_restatement, instruction_collapse, exemplar_leak, over_specification, format_break} or null. |
all_tags |
list | Multi-label tag set (subsumes primary_tag). |
rationale |
string | The judge's natural-language explanation. |
confidence |
float64 | Judge's reported confidence in primary_tag ∈ [0, 1]. |
source |
string | deterministic (similarity short-circuit), llm_judge (real LLM call), judge_parse_failure (empty / malformed JSON), or excluded_dspy_serialization_bug (substrate-level exclusion). |
judge_actual_model |
string | Provider-returned judge model id, null for deterministic / excluded paths. |
judge_cost_usd |
float64 | Per-record API cost in USD. |
manual_label_subset |
string | Null in the parquet. The blind human audit's 50 labels (§ Annotators) are versioned in the source repository as a keyed sidecar (outputs/audits/2026-07-28/labels.csv, keyed by audit-form row id with the mapping in form.md). |
Data Splits
This is a single-split dataset (no train/val/test partition); records are organised by (cell, seed) provenance rather than by inference split. The dataset is intended as labelled evaluation material for downstream APO-reliability research, not for training.
Per-source breakdown
source |
count | % |
|---|---|---|
llm_judge |
1004 | 59.2% |
deterministic |
691 | 40.8% |
Per-primary-tag breakdown
primary_tag |
count | % |
|---|---|---|
trivial_restatement |
708 | 41.8% |
format_break |
376 | 22.2% |
over_specification |
308 | 18.2% |
exemplar_leak |
253 | 14.9% |
instruction_collapse |
49 | 2.9% |
(none) |
1 | 0.1% |
Per-task breakdown
| task | count |
|---|---|
sst5 |
660 |
hotpotqa |
484 |
ag_news |
434 |
gsm8k |
117 |
Per-optimiser breakdown
| optimiser | count |
|---|---|
gepa |
1652 |
opro |
22 |
dspy_mipro |
21 |
Per-model breakdown
| model | count |
|---|---|
Qwen/Qwen2.5-7B-Instruct-Turbo |
847 |
claude-haiku-4-5-20251001 |
610 |
deepseek-v4-flash |
238 |
Dataset Creation
Curation Rationale
Two concurrent works frame the gap this dataset fills:
- Coin Flip (Zhang et al., 2026; arXiv:2604.14585) reports that 49% of APO runs on Claude Haiku 4.5 underperform zero-shot, but releases only ANOVA results, not labelled failure traces.
- ETGPO (Singh, Yadav & Blanco, 2026; arXiv:2602.00997) uses an error taxonomy to drive APO optimisation but does not release a labelled dataset of optimiser-side failures.
The labelled-dataset deliverable in this dataset is, to our knowledge, the first published collection of multi-method × multi-task × multi-model APO failures with a fixed taxonomy and per-record provenance metadata.
Source Data
The held-out evaluation traces are model completions on standard NLP benchmarks (GSM8K, HotpotQA, SST-5, AG News, HumanEval), produced by three APO methods (DSPy MIPROv2, OPRO, GEPA) and one baseline (none — the manual baseline prompt). Each trace's underlying dataset is publicly available under its original license:
- GSM8K: MIT License (Cobbe et al., 2021). Source: https://huggingface.co/datasets/gsm8k.
- HotpotQA: CC BY-SA 4.0 (Yang et al., 2018). Source: https://huggingface.co/datasets/hotpot_qa.
- SST-5: derived from the Stanford Sentiment Treebank (Socher et al., 2013) via the SetFit redistribution. Source: https://huggingface.co/datasets/SetFit/sst5.
- AG News: CC BY-SA 3.0 / public domain headlines. Source: https://huggingface.co/datasets/ag_news.
- HumanEval: MIT License (Chen et al., 2021). Source: https://huggingface.co/datasets/openai_humaneval.
Only the model's completions and the gold/predicted scoring are reproduced in this dataset; the underlying questions / inputs are referenced by (task_name, seed, trace_index) so the upstream datasets can be reloaded deterministically.
Annotations
The annotation pipeline is a hybrid deterministic + LLM-as-judge classifier:
- Deterministic short-circuit (
source == "deterministic"). If the optimised prompt has cosine similarity ≥ 0.95 (whitespace-normalised) or is byte-identical to the manual baseline, the record is taggedtrivial_restatementwithout an LLM call. This accounts for ~30% of dataset records. - LLM-as-judge (
source == "llm_judge"). The remaining records are scored by DeepSeek V4-flash running the locked judge prompt atprompts/failure_mode_judge.txt(max_tokens=4096 to accommodate the model's hidden reasoning channel, see Limitations). The judge returns a JSON object withtags,primary,rationale,confidence. - Parse-failure fallback (
source == "judge_parse_failure"). Records where the judge returned malformed JSON or empty content are tagged with a nullprimary_tagand a rationale preserving the raw response for human follow-up. - Substrate exclusion (
source == "excluded_dspy_serialization_bug"). Records affected by the X.DSPY-SERIAL.1 substrate bug were excluded from the dataset pre-v0.1; the bug was fixed on 2026-06-02 and the affected cells re-classified, so v0.1 has 0 records in this bucket.
Annotation Procedure
- Annotator: DeepSeek V4-flash (versioned id
deepseek-v4-flash, pinned per the X.PIN.1 audit).judge_actual_modelrecords the provider-returned id per record. - Prompt: frozen at
prompts/failure_mode_judge.txtonce the production sweep started on 2026-06-02. Editing the prompt invalidates the inter-rater κ measurement perAGENTS.md §2 rule 6. - Taxonomy: 5 base tags (
trivial_restatement,instruction_collapse,exemplar_leak,over_specification,format_break) — locked. Up to 2 reserved slots may be populated from an open-coding pass (P4.OPENCODE.1) before sweep lock; v0.1 ships with the 5 base tags only.
Who are the annotators?
The primary annotator is the DeepSeek V4-flash model running the locked judge prompt. Label reliability was measured four ways, in increasing order of independence:
- Test–retest decoding stability (locked fallback): the judge re-labelled a 50-trace stratified subsample of its own judge-assigned traces (10 per primary tag; deterministic
trivial_restatementlabels excluded as reproducible by construction). κ = 0.925 (observed 0.940, chance 0.200; PABAK 0.88) — passes the κ ≥ 0.7 gate. This is self-agreement at temperature 0, not inter-rater agreement. - Score-blind re-judge: masking the accuracy gap and the underperformance premise drops blind-vs-dataset agreement to κ = 0.650 — the judge's labels partly depend on seeing the scores.
- Anchoring false-positive arm: shown traces from runs that did not underperform, the judge still assigned a failure mode 79% of the time — an upper bound on its premise-following disposition.
- Blind human audit (the pre-registered seal): author–judge κ = 0.925 (measured 2026-08-02; observed 0.940, chance 0.200; PABAK 0.88) — passes the κ ≥ 0.7 gate. A human rater labelled the frozen 50-trace stratified form blind (judge labels withheld until all 50 labels were committed). Per-tag agreement:
exemplar_leak10/10,format_break10/10,instruction_collapse9/10,over_specification9/10,trivial_restatement9/10; all three disagreements lie within the interpretive trio (instruction_collapse↔over_specification/trivial_restatement). Two caveats: the rater is the source-thesis author (blind to labels, but not independent of the rubric — a second, independent rater would be the strongest version), and the audited sample is drawn from the two-family (DeepSeek + Qwen) dataset build, so traces contributed by the Haiku family after the audit carry judge labels only.
Net verdict: the judge is internally consistent, the deterministic backbone plus the content-visible categories survive score-blinding, and a blind human rater concurs at κ = 0.925 — so the labels carry measured author–judge agreement, one step short of independent inter-rater confirmation. instruction_collapse remains the most score-anchored category (both raters saw the underperformance premise).
Personal and Sensitive Information
None. The underlying benchmarks are pre-existing public NLP datasets; model completions do not contain personal information. Prompt content is purely task instructions plus optimisation artefacts.
Considerations for Using the Data
Social Impact
The dataset may be useful for researchers studying:
- when automated prompt optimisation fails and why (Claim 1 / RQ1 of the source thesis),
- whether published APO reliability benchmarks are reproducible across optimiser families (the cross-task profiles in this dataset answer this for the 32-cell substrate),
- whether human evaluators agree with LLM-as-judge classification of optimisation failures (judge-internal bounds plus a blind human audit at κ = 0.925 are reported — see Annotators).
It is not suitable for training stronger APO methods directly — the dataset is small (≈1k records), single-language, and was not constructed for generative training.
Discussion of Biases
The dataset reflects the substrate's biases:
- Task coverage is limited to 5 well-studied APO benchmarks. There is no evidence that the failure-mode profile generalises to alignment, dialogue, or multilingual tasks.
- Model coverage spans 3 model families (DeepSeek V4-flash, Claude Haiku 4.5, Qwen-2.5-7B-Instruct-Turbo); the Haiku tier was completed 2026-08-02/03 and carries GEPA runs only (see Optimiser coverage below). Haiku traces postdate the blind human audit and carry judge labels only.
- Optimiser coverage is asymmetric: GEPA runs at full 5 × 3 grid; DSPy MIPROv2 and OPRO run only on
gsm8k × DeepSeek V4-flash(Path B narrowing perP4.MULTITASK.1partial port). The full grid is a thesis-development item. - Judge bias: DeepSeek V4-flash labels the dataset. It has not been independently audited for systematic bias toward particular taxonomy tags. The M2 concurrent-validity pilot (planned) will quantify this.
Other Known Limitations
- Human audit is author-rater, not independent: the blind human audit (author–judge κ = 0.925, § Annotators) was annotated by the source-thesis author under a blind protocol. Blinding removes label leakage but not shared training on the rubric, so this is author–judge agreement under blinding rather than independent inter-rater reliability; the score-anchored
instruction_collapseshould still be read with the anchoring caveat in mind. - DeepSeek V4-flash hidden reasoning (X.PIN.2): V4-flash defaults to thinking-mode-on at the API layer, costing ~8pp accuracy on a GSM8K seed-42 probe. This dataset was produced from the reasoning-disabled substrate: all DeepSeek V4-flash cells were re-run with
extra_body={"thinking": {"type": "disabled"}}(theD-NEW-2re-run), and this rebuild deduplicates to those thinking-off runs. Earlier thinking-on aggregates are superseded. - Open-coding pass (
P4.OPENCODE.1): the 2 reserved taxonomy slots have not been populated. v0.1 ships with the 5 base tags only. - Coin Flip replication cell (
P4.REPLICATE.1): not included in v0.1. It will appear in v0.2 as a separatereplication_cellsubset.
Comparison to Concurrent Corpora
| dataset | unit of analysis | size | release |
|---|---|---|---|
| APO Failure-Mode Dataset (this dataset) | optimiser-output failure (prompt + trace) | ~1,695 | HF Hub |
| ETGPO (Singh et al., 2026; arXiv:2602.00997) | task-error (LLM output failure) | not released as dataset | (paper only) |
| Tian et al. (arXiv:2509.14404) | manually-written prompt defect | not released | (paper only) |
| MemAPO (arXiv:2603.21520) | optimiser-internal memory artefact | not released | (paper only) |
| Coin Flip (Zhang et al., 2026; arXiv:2604.14585) | ANOVA variance decomposition | aggregate only | (paper only) |
This dataset is the first publicly-released dataset that is keyed on (optimiser-output, failure-mode label) tuples with per-record provenance.
Reproducibility
The dataset is fully reproducible from the source repository (private during thesis development; public release with thesis defence):
- Run the 32-cell sweep via
scripts/run_sweep.py --config configs/sweep_main.yaml. Cost: ≈ $16 USD across all cells (the Anthropic tier dominates). - Classify the failure modes via
scripts/classify_failures.py --all --concurrency 4 --judge-max-tokens 4096. Cost: ≈ $2 USD. - Rebuild this card via
scripts/build_dataset_card.py.
Every record's parent run carries a run_metadata.json with git SHA, resolved cell config, provider-returned actual_model, and per-call cost. The substrate is git-pinned to deepseek-v4-flash (versioned id) per the X.PIN.1 audit and frozen baselines per X.BASELINE.1 (prompts/<task>_baseline.txt for each task).
Citation
TODO (defence September 2026): replace this placeholder with the thesis citation + an arXiv companion link.
@misc{apo_failure_modes_2026,
title = {APO Failure-Mode Dataset: A Labelled Dataset of Automated Prompt Optimization Failures},
author = {Dang Vu Son Tung},
year = {2026},
note = {USTH Master's thesis, v0.2},
howpublished = {Hugging Face Hub, \url{https://huggingface.co/datasets/tungdvs/apo-failure-modes}},
}
License
This dataset is released under CC BY 4.0. The underlying benchmark datasets remain under their original licenses (see Source Data); each record's task field identifies the upstream license that applies to that record's underlying input.
- Downloads last month
- 17