id stringlengths 8 8 | category stringclasses 8
values | subcategory stringclasses 9
values | difficulty stringclasses 3
values | question stringlengths 54 376 | solution stringlengths 53 584 | tags stringclasses 9
values |
|---|---|---|---|---|---|---|
arb_0001 | causal_reasoning | diagnosis | medium | Observed effect: service downtime. Possible causes include software bug and others. What evidence would support or refute that software bug is the actual cause? Propose a simple diagnostic plan. | Supporting evidence for 'software bug': temporal correlation (cause preceded effect), presence of characteristic symptoms of software bug, absence of alternative explanations.
Refuting evidence: the effect occurred without software bug, or the effect is fully explained by another cause.
Diagnostic plan:
1. Check timest... | causal,diagnosis,agentic |
arb_0002 | agentic_planning | multi_step | medium | You are an agent assisting Alex. Goal: retrieve the umbrella from the bedroom and deliver it to the lab. Available tools include browser and file_system. Provide a numbered step-by-step plan with at least 3 steps. Include checks for success and recovery actions. | 1. Use browser or file_system to locate the current position of the umbrella (confirm it is in bedroom).
2. Navigate or request access to the bedroom.
3. Pick up / retrieve the umbrella.
4. Verify possession of the umbrella (success check).
5. Navigate to the lab.
6. Deliver / place the umbrella in the lab and confirm ... | planning,agent,multi-step,recovery |
arb_0003 | logical_reasoning | deduction | medium | No Rileys are in the library. Avery is in the library. Is Avery a Riley? Explain. | No. The premise states that no Rileys are in the library. Since Avery is in the library, Avery cannot be a Riley. | deduction,syllogism,logic |
arb_0004 | logical_reasoning | deduction | medium | Either Jamie or Cameron took the pen, but not both. Jamie did not take the pen. Who took it? | Cameron took the pen. This follows from the exclusive disjunction and the negation of one disjunct. | deduction,syllogism,logic |
arb_0005 | code_reasoning | debugging | medium | What is wrong with the following Python code?
```python
def add(a, b):
return a - b
```
Explain the error and provide a corrected version. | The function is named 'add' but performs subtraction. It should return a + b.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0006 | agentic_planning | long_horizon | hard | You are a long-horizon agent. Your ultimate goal is to prepare a weekly status report for Jordan. Sub-goals: (1) collect metrics from three different systems, (2) analyze trends, (3) draft the report, (4) request review, (5) publish. Describe how you would decompose this into an executable plan, including which tools y... | Decomposition:
1. Metrics collection: call database_query / terminal / API tools for each system; store results in files.
2. Analysis: use code_interpreter to compute trends, averages, anomalies.
3. Draft: generate structured text (markdown) from analysis results.
4. Review: send via email_client or create a review tic... | long-horizon,decomposition,agent,robustness |
arb_0007 | mathematical_reasoning | arithmetic | easy | Start with 32. Subtract 25. Then add 9. What is the final value? Show all steps. | Step 1: 32 - 25 = 7
Step 2: 7 + 9 = 16
Final answer: 16 | math,arithmetic,step-by-step |
arb_0008 | code_reasoning | debugging | medium | What is wrong with the following Python code?
```python
d = {'a': 1}
print(d['b'])
```
Explain the error and provide a corrected version. | KeyError: 'b' is not a key in the dictionary. Use d.get('b') or check membership.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0009 | logical_reasoning | deduction | easy | No Taylors are in the bedroom. Avery is in the bedroom. Is Avery a Taylor? Explain. | No. The premise states that no Taylors are in the bedroom. Since Avery is in the bedroom, Avery cannot be a Taylor. | deduction,syllogism,logic |
arb_0010 | mathematical_reasoning | arithmetic | easy | Start with 24. Add 30. Then multiply the result by 7. What is the final value? Show all steps. | Step 1: 24 + 30 = 54
Step 2: 54 * 7 = 378
Final answer: 378 | math,arithmetic,step-by-step |
arb_0011 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0012 | commonsense_reasoning | everyday | easy | A blue rabbit is observed in the server room. Is it more likely that the rabbit lives there permanently or is only visiting? Give a reasoned answer based on typical knowledge about rabbits. | Most rabbits do not permanently live in a server room (which is a human-built environment). It is therefore more likely that the rabbit is only visiting or was brought there temporarily. Exceptions exist (e.g. pets), but the default assumption is temporary presence. | commonsense,plausibility |
arb_0013 | code_reasoning | debugging | easy | What is wrong with the following Python code?
```python
x = [1, 2, 3]
print(x[3])
```
Explain the error and provide a corrected version. | IndexError: list index out of range. Valid indices are 0, 1, 2.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0014 | causal_reasoning | diagnosis | medium | Observed effect: delayed delivery. Possible causes include rain and others. What evidence would support or refute that rain is the actual cause? Propose a simple diagnostic plan. | Supporting evidence for 'rain': temporal correlation (cause preceded effect), presence of characteristic symptoms of rain, absence of alternative explanations.
Refuting evidence: the effect occurred without rain, or the effect is fully explained by another cause.
Diagnostic plan:
1. Check timestamps of rain indicators.... | causal,diagnosis,agentic |
arb_0015 | agentic_planning | long_horizon | hard | You are a long-horizon agent. Your ultimate goal is to prepare a weekly status report for Quinn. Sub-goals: (1) collect metrics from three different systems, (2) analyze trends, (3) draft the report, (4) request review, (5) publish. Describe how you would decompose this into an executable plan, including which tools yo... | Decomposition:
1. Metrics collection: call database_query / terminal / API tools for each system; store results in files.
2. Analysis: use code_interpreter to compute trends, averages, anomalies.
3. Draft: generate structured text (markdown) from analysis results.
4. Review: send via email_client or create a review tic... | long-horizon,decomposition,agent,robustness |
arb_0016 | code_reasoning | debugging | easy | What is wrong with the following Python code?
```python
x = [1, 2, 3]
print(x[3])
```
Explain the error and provide a corrected version. | IndexError: list index out of range. Valid indices are 0, 1, 2.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0017 | code_reasoning | debugging | medium | What is wrong with the following Python code?
```python
for i in range(5):
print(i)
```
Explain the error and provide a corrected version. | IndentationError: the print statement must be indented under the for-loop.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0018 | mathematical_reasoning | arithmetic | easy | Start with 27. Subtract 25. Then add 15. What is the final value? Show all steps. | Step 1: 27 - 25 = 2
Step 2: 2 + 15 = 17
Final answer: 17 | math,arithmetic,step-by-step |
arb_0019 | tool_use | selection | easy | Available tools: code_interpreter, email_client, calendar, database_query.
User goal: find the current weather in Berlin.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: code_interpreter
Approximate call: code_interpreter(query="find the current weather in Berlin")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0020 | logical_reasoning | deduction | medium | Either Morgan or Casey took the umbrella, but not both. Morgan did not take the umbrella. Who took it? | Casey took the umbrella. This follows from the exclusive disjunction and the negation of one disjunct. | deduction,syllogism,logic |
arb_0021 | commonsense_reasoning | everyday | easy | A blue cat is observed in the kitchen. Is it more likely that the cat lives there permanently or is only visiting? Give a reasoned answer based on typical knowledge about cats. | Most cats do not permanently live in a kitchen (which is a human-built environment). It is therefore more likely that the cat is only visiting or was brought there temporarily. Exceptions exist (e.g. pets), but the default assumption is temporary presence. | commonsense,plausibility |
arb_0022 | commonsense_reasoning | everyday | easy | A white dog is observed in the library. Is it more likely that the dog lives there permanently or is only visiting? Give a reasoned answer based on typical knowledge about dogs. | Most dogs do not permanently live in a library (which is a human-built environment). It is therefore more likely that the dog is only visiting or was brought there temporarily. Exceptions exist (e.g. pets), but the default assumption is temporary presence. | commonsense,plausibility |
arb_0023 | mathematical_reasoning | arithmetic | easy | Compute (14 * 19) + 45. Explain the order of operations. | Multiplication first: 14 * 19 = 266
Then addition: 266 + 45 = 311
Final answer: 311 | math,arithmetic,step-by-step |
arb_0024 | commonsense_reasoning | everyday | easy | A white owl is observed in the kitchen. Is it more likely that the owl lives there permanently or is only visiting? Give a reasoned answer based on typical knowledge about owls. | Most owls do not permanently live in a kitchen (which is a human-built environment). It is therefore more likely that the owl is only visiting or was brought there temporarily. Exceptions exist (e.g. pets), but the default assumption is temporary presence. | commonsense,plausibility |
arb_0025 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0026 | agentic_planning | multi_step | medium | You are an agent assisting Morgan. Goal: retrieve the laptop from the warehouse and deliver it to the lab. Available tools include calendar and file_system. Provide a numbered step-by-step plan with at least 3 steps. Include checks for success and recovery actions. | 1. Use calendar or file_system to locate the current position of the laptop (confirm it is in warehouse).
2. Navigate or request access to the warehouse.
3. Pick up / retrieve the laptop.
4. Verify possession of the laptop (success check).
5. Navigate to the lab.
6. Deliver / place the laptop in the lab and confirm del... | planning,agent,multi-step,recovery |
arb_0027 | code_reasoning | debugging | easy | What is wrong with the following Python code?
```python
x = [1, 2, 3]
print(x[3])
```
Explain the error and provide a corrected version. | IndexError: list index out of range. Valid indices are 0, 1, 2.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0028 | agentic_planning | long_horizon | hard | You are a long-horizon agent. Your ultimate goal is to prepare a weekly status report for Alex. Sub-goals: (1) collect metrics from three different systems, (2) analyze trends, (3) draft the report, (4) request review, (5) publish. Describe how you would decompose this into an executable plan, including which tools you... | Decomposition:
1. Metrics collection: call database_query / terminal / API tools for each system; store results in files.
2. Analysis: use code_interpreter to compute trends, averages, anomalies.
3. Draft: generate structured text (markdown) from analysis results.
4. Review: send via email_client or create a review tic... | long-horizon,decomposition,agent,robustness |
arb_0029 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0030 | agentic_planning | multi_step | medium | You are an agent assisting Morgan. Goal: retrieve the pen from the lobby and deliver it to the bedroom. Available tools include calendar and file_system. Provide a numbered step-by-step plan with at least 3 steps. Include checks for success and recovery actions. | 1. Use calendar or file_system to locate the current position of the pen (confirm it is in lobby).
2. Navigate or request access to the lobby.
3. Pick up / retrieve the pen.
4. Verify possession of the pen (success check).
5. Navigate to the bedroom.
6. Deliver / place the pen in the bedroom and confirm delivery.
Recov... | planning,agent,multi-step,recovery |
arb_0031 | logical_reasoning | deduction | medium | If someone owns a charger, then they have access to the garage. Riley owns a charger. What can we conclude about Riley? | We can conclude that Riley has access to the garage, by modus ponens. | deduction,syllogism,logic |
arb_0032 | mathematical_reasoning | arithmetic | easy | Compute (48 * 9) + 34. Explain the order of operations. | Multiplication first: 48 * 9 = 432
Then addition: 432 + 34 = 466
Final answer: 466 | math,arithmetic,step-by-step |
arb_0033 | commonsense_reasoning | everyday | easy | A green owl is observed in the kitchen. Is it more likely that the owl lives there permanently or is only visiting? Give a reasoned answer based on typical knowledge about owls. | Most owls do not permanently live in a kitchen (which is a human-built environment). It is therefore more likely that the owl is only visiting or was brought there temporarily. Exceptions exist (e.g. pets), but the default assumption is temporary presence. | commonsense,plausibility |
arb_0034 | agentic_planning | multi_step | medium | You are an agent assisting Jordan. Goal: retrieve the umbrella from the lobby and deliver it to the meeting room. Available tools include code_interpreter and file_system. Provide a numbered step-by-step plan with at least 5 steps. Include checks for success and recovery actions. | 1. Use code_interpreter or file_system to locate the current position of the umbrella (confirm it is in lobby).
2. Navigate or request access to the lobby.
3. Pick up / retrieve the umbrella.
4. Verify possession of the umbrella (success check).
5. Navigate to the meeting room.
6. Deliver / place the umbrella in the me... | planning,agent,multi-step,recovery |
arb_0035 | tool_use | selection | easy | Available tools: calendar, web_search, file_system, code_interpreter, browser.
User goal: check disk usage on /var.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: file_system
Approximate call: file_system(command="df -h /var")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0036 | logical_reasoning | deduction | easy | If someone owns a umbrella, then they have access to the meeting room. Alex owns a umbrella. What can we conclude about Alex? | We can conclude that Alex has access to the meeting room, by modus ponens. | deduction,syllogism,logic |
arb_0037 | agentic_planning | multi_step | medium | You are an agent assisting Taylor. Goal: retrieve the pen from the library and deliver it to the meeting room. Available tools include terminal and file_system. Provide a numbered step-by-step plan with at least 3 steps. Include checks for success and recovery actions. | 1. Use terminal or file_system to locate the current position of the pen (confirm it is in library).
2. Navigate or request access to the library.
3. Pick up / retrieve the pen.
4. Verify possession of the pen (success check).
5. Navigate to the meeting room.
6. Deliver / place the pen in the meeting room and confirm d... | planning,agent,multi-step,recovery |
arb_0038 | agentic_planning | multi_step | hard | You are an agent assisting Riley. Goal: retrieve the badge from the office and deliver it to the garage. Available tools include file_system and file_system. Provide a numbered step-by-step plan with at least 3 steps. Include checks for success and recovery actions. | 1. Use file_system or file_system to locate the current position of the badge (confirm it is in office).
2. Navigate or request access to the office.
3. Pick up / retrieve the badge.
4. Verify possession of the badge (success check).
5. Navigate to the garage.
6. Deliver / place the badge in the garage and confirm deli... | planning,agent,multi-step,recovery |
arb_0039 | agentic_planning | long_horizon | hard | You are a long-horizon agent. Your ultimate goal is to prepare a weekly status report for Casey. Sub-goals: (1) collect metrics from three different systems, (2) analyze trends, (3) draft the report, (4) request review, (5) publish. Describe how you would decompose this into an executable plan, including which tools yo... | Decomposition:
1. Metrics collection: call database_query / terminal / API tools for each system; store results in files.
2. Analysis: use code_interpreter to compute trends, averages, anomalies.
3. Draft: generate structured text (markdown) from analysis results.
4. Review: send via email_client or create a review tic... | long-horizon,decomposition,agent,robustness |
arb_0040 | tool_use | selection | medium | Available tools: terminal, code_interpreter, web_search.
User goal: restart the nginx service.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: terminal
Approximate call: terminal(appropriate_arguments)
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0041 | tool_use | selection | medium | Available tools: web_search, browser, terminal, file_system.
User goal: search for recent papers on transformer architectures.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: web_search
Approximate call: web_search(query="search for recent papers on transformer architectures")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0042 | mathematical_reasoning | arithmetic | easy | Start with 37. Subtract 41. Then add 46. What is the final value? Show all steps. | Step 1: 37 - 41 = -4
Step 2: -4 + 46 = 42
Final answer: 42 | math,arithmetic,step-by-step |
arb_0043 | tool_use | selection | easy | Available tools: code_interpreter, file_system, email_client, calendar.
User goal: send a meeting invitation for tomorrow at 10am.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: calendar
Approximate call: calendar(title="Meeting", time="tomorrow 10:00")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0044 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0045 | tool_use | selection | easy | Available tools: file_system, database_query, calendar, web_search, terminal.
User goal: search for recent papers on transformer architectures.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: web_search
Approximate call: web_search(query="search for recent papers on transformer architectures")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0046 | agentic_planning | multi_step | hard | You are an agent assisting Riley. Goal: retrieve the charger from the bedroom and deliver it to the lobby. Available tools include web_search and file_system. Provide a numbered step-by-step plan with at least 6 steps. Include checks for success and recovery actions. | 1. Use web_search or file_system to locate the current position of the charger (confirm it is in bedroom).
2. Navigate or request access to the bedroom.
3. Pick up / retrieve the charger.
4. Verify possession of the charger (success check).
5. Navigate to the lobby.
6. Deliver / place the charger in the lobby and confi... | planning,agent,multi-step,recovery |
arb_0047 | causal_reasoning | diagnosis | hard | Observed effect: delayed delivery. Possible causes include power outage and others. What evidence would support or refute that power outage is the actual cause? Propose a simple diagnostic plan. | Supporting evidence for 'power outage': temporal correlation (cause preceded effect), presence of characteristic symptoms of power outage, absence of alternative explanations.
Refuting evidence: the effect occurred without power outage, or the effect is fully explained by another cause.
Diagnostic plan:
1. Check timest... | causal,diagnosis,agentic |
arb_0048 | code_reasoning | debugging | easy | What is wrong with the following Python code?
```python
s = 'hello'
print(s.append('!'))
```
Explain the error and provide a corrected version. | AttributeError: 'str' object has no attribute 'append'. Strings are immutable; use s + '!'.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0049 | agentic_planning | multi_step | hard | You are an agent assisting Quinn. Goal: retrieve the wallet from the warehouse and deliver it to the lobby. Available tools include browser and file_system. Provide a numbered step-by-step plan with at least 6 steps. Include checks for success and recovery actions. | 1. Use browser or file_system to locate the current position of the wallet (confirm it is in warehouse).
2. Navigate or request access to the warehouse.
3. Pick up / retrieve the wallet.
4. Verify possession of the wallet (success check).
5. Navigate to the lobby.
6. Deliver / place the wallet in the lobby and confirm ... | planning,agent,multi-step,recovery |
arb_0050 | code_reasoning | debugging | medium | What is wrong with the following Python code?
```python
x = [1, 2, 3]
print(x[3])
```
Explain the error and provide a corrected version. | IndexError: list index out of range. Valid indices are 0, 1, 2.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0051 | causal_reasoning | diagnosis | hard | Observed effect: error logs. Possible causes include rain and others. What evidence would support or refute that rain is the actual cause? Propose a simple diagnostic plan. | Supporting evidence for 'rain': temporal correlation (cause preceded effect), presence of characteristic symptoms of rain, absence of alternative explanations.
Refuting evidence: the effect occurred without rain, or the effect is fully explained by another cause.
Diagnostic plan:
1. Check timestamps of rain indicators.... | causal,diagnosis,agentic |
arb_0052 | tool_use | selection | easy | Available tools: web_search, file_system, email_client, database_query, code_interpreter.
User goal: query the sales table for Q3 revenue.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: database_query
Approximate call: database_query(sql="SELECT SUM(revenue) FROM sales WHERE quarter='Q3'")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0053 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0054 | agentic_planning | multi_step | hard | You are an agent assisting Casey. Goal: retrieve the phone from the lab and deliver it to the lobby. Available tools include database_query and file_system. Provide a numbered step-by-step plan with at least 5 steps. Include checks for success and recovery actions. | 1. Use database_query or file_system to locate the current position of the phone (confirm it is in lab).
2. Navigate or request access to the lab.
3. Pick up / retrieve the phone.
4. Verify possession of the phone (success check).
5. Navigate to the lobby.
6. Deliver / place the phone in the lobby and confirm delivery.... | planning,agent,multi-step,recovery |
arb_0055 | logical_reasoning | deduction | easy | All Alexs own a umbrella. Jamie is a Alex. Does Jamie own a umbrella? Explain step by step. | Yes. From the universal statement 'All Alexs own a umbrella' and the particular 'Jamie is a Alex', it follows by universal instantiation that Jamie owns a umbrella. | deduction,syllogism,logic |
arb_0056 | code_reasoning | debugging | easy | What is wrong with the following Python code?
```python
def add(a, b):
return a - b
```
Explain the error and provide a corrected version. | The function is named 'add' but performs subtraction. It should return a + b.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0057 | tool_use | selection | easy | Available tools: email_client, file_system, code_interpreter, terminal, calendar.
User goal: list all Python files in the current directory.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: terminal
Approximate call: terminal(command="ls *.py")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0058 | code_reasoning | debugging | easy | What is wrong with the following Python code?
```python
for i in range(5):
print(i)
```
Explain the error and provide a corrected version. | IndentationError: the print statement must be indented under the for-loop.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0059 | causal_reasoning | diagnosis | medium | Observed effect: error logs. Possible causes include software bug and others. What evidence would support or refute that software bug is the actual cause? Propose a simple diagnostic plan. | Supporting evidence for 'software bug': temporal correlation (cause preceded effect), presence of characteristic symptoms of software bug, absence of alternative explanations.
Refuting evidence: the effect occurred without software bug, or the effect is fully explained by another cause.
Diagnostic plan:
1. Check timest... | causal,diagnosis,agentic |
arb_0060 | logical_reasoning | deduction | easy | If someone owns a key, then they have access to the lobby. Jordan owns a key. What can we conclude about Jordan? | We can conclude that Jordan has access to the lobby, by modus ponens. | deduction,syllogism,logic |
arb_0061 | tool_use | selection | medium | Available tools: web_search, file_system, code_interpreter, terminal.
User goal: list all Python files in the current directory.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: terminal
Approximate call: terminal(command="ls *.py")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0062 | mathematical_reasoning | arithmetic | easy | Start with 46. Subtract 4. Then add 29. What is the final value? Show all steps. | Step 1: 46 - 4 = 42
Step 2: 42 + 29 = 71
Final answer: 71 | math,arithmetic,step-by-step |
arb_0063 | agentic_planning | long_horizon | hard | You are a long-horizon agent. Your ultimate goal is to prepare a weekly status report for Quinn. Sub-goals: (1) collect metrics from three different systems, (2) analyze trends, (3) draft the report, (4) request review, (5) publish. Describe how you would decompose this into an executable plan, including which tools yo... | Decomposition:
1. Metrics collection: call database_query / terminal / API tools for each system; store results in files.
2. Analysis: use code_interpreter to compute trends, averages, anomalies.
3. Draft: generate structured text (markdown) from analysis results.
4. Review: send via email_client or create a review tic... | long-horizon,decomposition,agent,robustness |
arb_0064 | tool_use | selection | medium | Available tools: code_interpreter, file_system, email_client.
User goal: send a meeting invitation for tomorrow at 10am.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: email_client
Approximate call: email_client(title="Meeting", time="tomorrow 10:00")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0065 | code_reasoning | debugging | easy | What is wrong with the following Python code?
```python
x = [1, 2, 3]
print(x[3])
```
Explain the error and provide a corrected version. | IndexError: list index out of range. Valid indices are 0, 1, 2.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0066 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0067 | commonsense_reasoning | everyday | easy | A green wolf is observed in the server room. Is it more likely that the wolf lives there permanently or is only visiting? Give a reasoned answer based on typical knowledge about wolfs. | Most wolfs do not permanently live in a server room (which is a human-built environment). It is therefore more likely that the wolf is only visiting or was brought there temporarily. Exceptions exist (e.g. pets), but the default assumption is temporary presence. | commonsense,plausibility |
arb_0068 | agentic_planning | multi_step | medium | You are an agent assisting Jamie. Goal: retrieve the umbrella from the kitchen and deliver it to the office. Available tools include database_query and file_system. Provide a numbered step-by-step plan with at least 4 steps. Include checks for success and recovery actions. | 1. Use database_query or file_system to locate the current position of the umbrella (confirm it is in kitchen).
2. Navigate or request access to the kitchen.
3. Pick up / retrieve the umbrella.
4. Verify possession of the umbrella (success check).
5. Navigate to the office.
6. Deliver / place the umbrella in the office... | planning,agent,multi-step,recovery |
arb_0069 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0070 | logical_reasoning | deduction | medium | If someone owns a notebook, then they have access to the warehouse. Morgan owns a notebook. What can we conclude about Morgan? | We can conclude that Morgan has access to the warehouse, by modus ponens. | deduction,syllogism,logic |
arb_0071 | tool_use | selection | easy | Available tools: web_search, file_system, email_client, browser, code_interpreter.
User goal: search for recent papers on transformer architectures.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: web_search
Approximate call: web_search(query="search for recent papers on transformer architectures")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0072 | mathematical_reasoning | arithmetic | easy | Start with 20. Subtract 30. Then add 10. What is the final value? Show all steps. | Step 1: 20 - 30 = -10
Step 2: -10 + 10 = 0
Final answer: 0 | math,arithmetic,step-by-step |
arb_0073 | tool_use | selection | easy | Available tools: browser, calendar, code_interpreter, database_query.
User goal: send a meeting invitation for tomorrow at 10am.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: calendar
Approximate call: calendar(title="Meeting", time="tomorrow 10:00")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0074 | code_reasoning | debugging | easy | What is wrong with the following Python code?
```python
for i in range(5):
print(i)
```
Explain the error and provide a corrected version. | IndentationError: the print statement must be indented under the for-loop.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0075 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0076 | mathematical_reasoning | arithmetic | easy | Start with 17. Add 23. Then multiply the result by 38. What is the final value? Show all steps. | Step 1: 17 + 23 = 40
Step 2: 40 * 38 = 1520
Final answer: 1520 | math,arithmetic,step-by-step |
arb_0077 | code_reasoning | debugging | easy | What is wrong with the following Python code?
```python
def add(a, b):
return a - b
```
Explain the error and provide a corrected version. | The function is named 'add' but performs subtraction. It should return a + b.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0078 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0079 | agentic_planning | multi_step | hard | You are an agent assisting Casey. Goal: retrieve the phone from the kitchen and deliver it to the bedroom. Available tools include calendar and file_system. Provide a numbered step-by-step plan with at least 4 steps. Include checks for success and recovery actions. | 1. Use calendar or file_system to locate the current position of the phone (confirm it is in kitchen).
2. Navigate or request access to the kitchen.
3. Pick up / retrieve the phone.
4. Verify possession of the phone (success check).
5. Navigate to the bedroom.
6. Deliver / place the phone in the bedroom and confirm del... | planning,agent,multi-step,recovery |
arb_0080 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0081 | code_reasoning | debugging | easy | What is wrong with the following Python code?
```python
x = [1, 2, 3]
print(x[3])
```
Explain the error and provide a corrected version. | IndexError: list index out of range. Valid indices are 0, 1, 2.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0082 | agentic_planning | long_horizon | hard | You are a long-horizon agent. Your ultimate goal is to prepare a weekly status report for Cameron. Sub-goals: (1) collect metrics from three different systems, (2) analyze trends, (3) draft the report, (4) request review, (5) publish. Describe how you would decompose this into an executable plan, including which tools ... | Decomposition:
1. Metrics collection: call database_query / terminal / API tools for each system; store results in files.
2. Analysis: use code_interpreter to compute trends, averages, anomalies.
3. Draft: generate structured text (markdown) from analysis results.
4. Review: send via email_client or create a review tic... | long-horizon,decomposition,agent,robustness |
arb_0083 | code_reasoning | debugging | medium | What is wrong with the following Python code?
```python
s = 'hello'
print(s.append('!'))
```
Explain the error and provide a corrected version. | AttributeError: 'str' object has no attribute 'append'. Strings are immutable; use s + '!'.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0084 | agentic_planning | multi_step | hard | You are an agent assisting Quinn. Goal: retrieve the pen from the meeting room and deliver it to the bedroom. Available tools include file_system and file_system. Provide a numbered step-by-step plan with at least 4 steps. Include checks for success and recovery actions. | 1. Use file_system or file_system to locate the current position of the pen (confirm it is in meeting room).
2. Navigate or request access to the meeting room.
3. Pick up / retrieve the pen.
4. Verify possession of the pen (success check).
5. Navigate to the bedroom.
6. Deliver / place the pen in the bedroom and confir... | planning,agent,multi-step,recovery |
arb_0085 | agentic_planning | long_horizon | hard | You are a long-horizon agent. Your ultimate goal is to prepare a weekly status report for Quinn. Sub-goals: (1) collect metrics from three different systems, (2) analyze trends, (3) draft the report, (4) request review, (5) publish. Describe how you would decompose this into an executable plan, including which tools yo... | Decomposition:
1. Metrics collection: call database_query / terminal / API tools for each system; store results in files.
2. Analysis: use code_interpreter to compute trends, averages, anomalies.
3. Draft: generate structured text (markdown) from analysis results.
4. Review: send via email_client or create a review tic... | long-horizon,decomposition,agent,robustness |
arb_0086 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0087 | tool_use | selection | easy | Available tools: code_interpreter, database_query, file_system.
User goal: search for recent papers on transformer architectures.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: code_interpreter
Approximate call: code_interpreter(query="search for recent papers on transformer architectures")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional too... | tool-use,function-calling,agent |
arb_0088 | agentic_planning | long_horizon | hard | You are a long-horizon agent. Your ultimate goal is to prepare a weekly status report for Taylor. Sub-goals: (1) collect metrics from three different systems, (2) analyze trends, (3) draft the report, (4) request review, (5) publish. Describe how you would decompose this into an executable plan, including which tools y... | Decomposition:
1. Metrics collection: call database_query / terminal / API tools for each system; store results in files.
2. Analysis: use code_interpreter to compute trends, averages, anomalies.
3. Draft: generate structured text (markdown) from analysis results.
4. Review: send via email_client or create a review tic... | long-horizon,decomposition,agent,robustness |
arb_0089 | mathematical_reasoning | arithmetic | easy | Start with 17. Add 22. Then multiply the result by 23. What is the final value? Show all steps. | Step 1: 17 + 22 = 39
Step 2: 39 * 23 = 897
Final answer: 897 | math,arithmetic,step-by-step |
arb_0090 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0091 | logical_reasoning | deduction | easy | All Morgans own a laptop. Taylor is a Morgan. Does Taylor own a laptop? Explain step by step. | Yes. From the universal statement 'All Morgans own a laptop' and the particular 'Taylor is a Morgan', it follows by universal instantiation that Taylor owns a laptop. | deduction,syllogism,logic |
arb_0092 | agentic_planning | long_horizon | hard | You are a long-horizon agent. Your ultimate goal is to prepare a weekly status report for Jordan. Sub-goals: (1) collect metrics from three different systems, (2) analyze trends, (3) draft the report, (4) request review, (5) publish. Describe how you would decompose this into an executable plan, including which tools y... | Decomposition:
1. Metrics collection: call database_query / terminal / API tools for each system; store results in files.
2. Analysis: use code_interpreter to compute trends, averages, anomalies.
3. Draft: generate structured text (markdown) from analysis results.
4. Review: send via email_client or create a review tic... | long-horizon,decomposition,agent,robustness |
arb_0093 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0094 | tool_use | selection | easy | Available tools: email_client, web_search, database_query, calendar, terminal.
User goal: list all Python files in the current directory.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: terminal
Approximate call: terminal(command="ls *.py")
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0095 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0096 | code_reasoning | debugging | medium | What is wrong with the following Python code?
```python
for i in range(5):
print(i)
```
Explain the error and provide a corrected version. | IndentationError: the print statement must be indented under the for-loop.
Corrected version depends on the intended behavior; e.g. fix the operator, index, indentation, key, or method accordingly. | code,debugging,python |
arb_0097 | tool_use | selection | medium | Available tools: email_client, browser, database_query.
User goal: restart the nginx service.
Which tool(s) should the agent call first, and with what approximate arguments? Explain the reasoning and the expected observation. | Primary tool: email_client
Approximate call: email_client(appropriate_arguments)
Reasoning: The goal requires information or an action that this tool is designed for. After receiving the observation, the agent should verify completeness and, if needed, call additional tools. | tool-use,function-calling,agent |
arb_0098 | agentic_interaction | clarification | medium | User request: 'Get me the latest numbers.' The context is ambiguous (could mean financial metrics, sports scores, version numbers, etc.). How should an agent respond? Provide a clarification strategy and example clarifying questions. | An agent should not guess when critical ambiguity exists. Strategy: ask targeted clarifying questions while offering the most likely interpretations.
Example questions:
- Are you referring to financial / sales numbers, system metrics, or something else?
- For which time period and which entity?
- In which format do you... | ambiguity,clarification,user-interaction |
arb_0099 | agentic_planning | long_horizon | hard | You are a long-horizon agent. Your ultimate goal is to prepare a weekly status report for Taylor. Sub-goals: (1) collect metrics from three different systems, (2) analyze trends, (3) draft the report, (4) request review, (5) publish. Describe how you would decompose this into an executable plan, including which tools y... | Decomposition:
1. Metrics collection: call database_query / terminal / API tools for each system; store results in files.
2. Analysis: use code_interpreter to compute trends, averages, anomalies.
3. Draft: generate structured text (markdown) from analysis results.
4. Review: send via email_client or create a review tic... | long-horizon,decomposition,agent,robustness |
arb_0100 | mathematical_reasoning | arithmetic | easy | Compute (31 * 22) + 19. Explain the order of operations. | Multiplication first: 31 * 22 = 682
Then addition: 682 + 19 = 701
Final answer: 701 | math,arithmetic,step-by-step |
End of preview. Expand in Data Studio
Agentic & Reasoning Benchmark (ARB) – Expanded
Ein synthetischer Benchmark mit 2.550 Fragen und Lösungen, optimiert für die Evaluation von Agentic Capabilities und Reasoning.
Überblick
| Eigenschaft | Wert |
|---|---|
| Anzahl Beispiele | 2.550 |
| Kategorien | 8 |
| Schwierigkeitsgrade | easy / medium / hard |
| Formate | CSV + JSON |
| Reproduzierbarkeit | Generator-Skript (seed=42) enthalten |
| Lizenz | CC-BY-4.0 |
Kategorien
| Kategorie | Anzahl | Beschreibung |
|---|---|---|
agentic_planning |
550 | Multi-Step- und Long-Horizon-Planung, Decomposition, Recovery |
tool_use |
350 | Tool-Selection, Function-Calling, Argument-Design |
logical_reasoning |
350 | Deduktion, Syllogismen, logische Schlussfolgerungen |
mathematical_reasoning |
350 | Multi-Step-Arithmetik und Operator-Reihenfolge |
causal_reasoning |
250 | Ursachenanalyse und diagnostische Pläne |
code_reasoning |
250 | Debugging von Python-Snippets |
commonsense_reasoning |
250 | Plausibilitäts- und Alltagsschlussfolgerungen |
agentic_interaction |
200 | Umgang mit Ambiguität und Klärungsstrategien |
Dateistruktur (Hugging Face ready)
agentic-reasoning-benchmark/
├── benchmark.csv # 2.550 Beispiele
├── benchmark.json # Identische Daten im JSON-Format
├── generate_benchmark.py # Reproduzierbarer Generator
└── README.md # Dataset Card
Spalten
| Spalte | Typ | Beschreibung |
|---|---|---|
id |
string | Eindeutige ID (arb_0001 …) |
category |
string | Hauptkategorie |
subcategory |
string | Unterkategorie |
difficulty |
string | easy / medium / hard |
question |
string | Die Frage / Aufgabe |
solution |
string | Musterlösung |
tags |
string | Komma-separierte Tags |
Verwendung
Hugging Face Dataset laden
from datasets import load_dataset
ds = load_dataset("YOUR_USERNAME/agentic-reasoning-benchmark")
print(len(ds["train"])) # 2550
print(ds["train"][0])
Lokal
import pandas as pd
df = pd.read_csv("benchmark.csv")
print(df.category.value_counts())
print(df.difficulty.value_counts())
Evaluationsvorschläge
- Kurze Antworten (Math, Logic): Exact Match oder Soft Match
- Offene Aufgaben (Planning, Tool-Use, Causal): LLM-as-a-Judge mit Rubrik
- Rubrik-Kriterien:
- Korrekte Tool-Auswahl
- Vollständigkeit der Schritte
- Vorhandensein von Recovery-/Fehlerbehandlung
- Klarheit der Begründung
Design-Prinzipien
- Starker Fokus auf agentische Fähigkeiten (Tool-Use, Multi-Step, Long-Horizon, Recovery)
- Solide Abdeckung klassischer Reasoning-Typen (Logik, Mathematik, Kausalität, Code)
- Jede Frage besitzt eine nachvollziehbare Musterlösung
- Generator ist deterministisch (seed=42) und im Repository enthalten
Einschränkungen
- Synthetisch und templatisiert erzeugt – ersetzt keine realen Human-Annotated Benchmarks (GAIA, Terminal-Bench, SWE-bench, OSWorld usw.)
- Für produktionsreife Evaluation empfohlen: Kombination mit menschlichen Annotationen oder starken Judge-Modellen
- Long-Horizon- und Ambiguitäts-Aufgaben sind bewusst offen formuliert
Zitation
Agentic & Reasoning Benchmark (ARB), Expanded Edition, 2026.
Synthetic dataset for evaluating agentic and reasoning capabilities of LLMs.
Lizenz
CC-BY-4.0
- Downloads last month
- 7