Over-Reflective Search Agents
Collection
Searching after Answering: Diagnosing and Repairing Over-Reflection in Search Agents • 8 items • Updated • 1
id stringlengths 8 26 | prompt stringlengths 1.24k 200k | tools stringclasses 1
value | label stringlengths 19 966 | metadata stringlengths 232 18.4k |
|---|---|---|---|---|
stoppivot_3843057473_3 | [{"role": "system", "content": "Tool for browsing.\nThe `cursor` appears in brackets before each browsing display: `[{cursor}]`.\nCite information from the tool using the following format:\n`【{cursor}†L{line_start}(-L{line_end})?】`, for example: `【6†L9-L11】` or `【8†L3】`.\nDo not quote more than 10 words directly from t... | [{"type": "function", "function": {"name": "search", "description": "Searches for information related to `query` and displays `topn` results.", "parameters": {"type": "object", "properties": {"query": {"type": "string"}, "topn": {"type": "number", "default": 10}, "source": {"type": "string"}}, "required": ["query"]}}},... | {"keys": [["arena", "arena league of legends", "league of legends"]]} | {"evidence_state": "grounded", "a_star": "finalize", "a_star_text": "## 정답: **`Arena (League of Legends)`**\n\n### 추론 근거 요약\n\n| 단서 | 확인 내용 |\n|------|-----------|\n| **5. MOBA 프랜차이즈 식별** | 좌측 탐색 항목에 **Runeterra**, **챔피언 목록**, **주간 무료 챔피언 순환**, **룬 개편** 등이 등장 → **League of Legends** |\n| **4. 대형 팬 위키 플랫폼** | 페이지 곳곳에 **... |
stoppivot_3959413707_2 | [{"role": "system", "content": "Tool for browsing.\nThe `cursor` appears in brackets before each browsing display: `[{cursor}]`.\nCite information from the tool using the following format:\n`【{cursor}†L{line_start}(-L{line_end})?】`, for example: `【6†L9-L11】` or `【8†L3】`.\nDo not quote more than 10 words directly from t... | [{"type": "function", "function": {"name": "search", "description": "Searches for information related to `query` and displays `topn` results.", "parameters": {"type": "object", "properties": {"query": {"type": "string"}, "topn": {"type": "number", "default": 10}, "source": {"type": "string"}}, "required": ["query"]}}},... | {"keys": [["gloucester city council elections"]]} | {"evidence_state": "grounded", "a_star": "finalize", "a_star_text": "The process described is the **Gloucester City Council elections**, specifically the **By-election results** section of its Wikipedia article.\n\nHere's how each clue maps to the article:\n\n| Clue | How it matches |\n|---|---|\n| **Non-metropolitan d... |
studpivot_chunk_006_43_178 | [{"role": "system", "content": "Tool for browsing.\nThe `cursor` appears in brackets before each browsing display: `[{cursor}]`.\nCite information from the tool using the following format:\n`【{cursor}†L{line_start}(-L{line_end})?】`, for example: `【6†L9-L11】` or `【8†L3】`.\nDo not quote more than 10 words directly from t... | [{"type": "function", "function": {"name": "search", "description": "Searches for information related to `query` and displays `topn` results.", "parameters": {"type": "object", "properties": {"query": {"type": "string"}, "topn": {"type": "number", "default": 10}, "source": {"type": "string"}}, "required": ["query"]}}},... | {"keys": [["goreville"]]} | {"evidence_state": "ungrounded", "a_star": "toolcall", "a_star_text": "", "sft_action_name": null, "trajectory_id": "chunk_006_43", "assistant_depth": 178, "source": "student_dpi", "next_action": "finalize", "mu_profiled": null, "allow_seed": false, "track": "C_nextfinal"} |
stoppivot_1493681930_7 | [{"role": "system", "content": "Tool for browsing.\nThe `cursor` appears in brackets before each browsing display: `[{cursor}]`.\nCite information from the tool using the following format:\n`【{cursor}†L{line_start}(-L{line_end})?】`, for example: `【6†L9-L11】` or `【8†L3】`.\nDo not quote more than 10 words directly from t... | [{"type": "function", "function": {"name": "search", "description": "Searches for information related to `query` and displays `topn` results.", "parameters": {"type": "object", "properties": {"query": {"type": "string"}, "topn": {"type": "number", "default": 10}, "source": {"type": "string"}}, "required": ["query"]}}},... | {"keys": [["가족성 고콜레스테롤혈증 familial hypercholesterolaemia", "가족성 고콜레스테롤혈증", "familial hypercholesterolaemia"]]} | {"evidence_state": "grounded", "a_star": "finalize", "a_star_text": "## 추론 과정\n\n### 첫 번째 전략 분석\n\n| 단서 | 해석 |\n|------|------|\n| \"DLin-MC3-DMA 기반 2세대 지질나노입자에 실린 siRNA\" | Alnylam사의 2세대 LNP 플랫폼 |\n| \"32명의 건강한 지원자, 단회 정맥주사 0.4 mg/kg\" | ALN-PCS (후일 inclisiran) 제1상 임상시험(Fitzgerald et al., *Lancet* 2014) |\n| \"간에서 만들어... |
stoppivot_1149427659_13 | "[{\"role\": \"system\", \"content\": \"Tool for browsing.\\nThe `cursor` appears in brackets before(...TRUNCATED) | "[{\"type\": \"function\", \"function\": {\"name\": \"search\", \"description\": \"Searches for info(...TRUNCATED) | {"keys": [["alexandra strait"]]} | "{\"evidence_state\": \"grounded\", \"a_star\": \"finalize\", \"a_star_text\": \"**Alexandra Strait*(...TRUNCATED) |
stoppivot_2563988023_4 | "[{\"role\": \"system\", \"content\": \"Tool for browsing.\\nThe `cursor` appears in brackets before(...TRUNCATED) | "[{\"type\": \"function\", \"function\": {\"name\": \"search\", \"description\": \"Searches for info(...TRUNCATED) | {"keys": [["battle of prilep"]]} | "{\"evidence_state\": \"grounded\", \"a_star\": \"finalize\", \"a_star_text\": \"The battle you're d(...TRUNCATED) |
stoppivot_1996946444_3 | "[{\"role\": \"system\", \"content\": \"Tool for browsing.\\nThe `cursor` appears in brackets before(...TRUNCATED) | "[{\"type\": \"function\", \"function\": {\"name\": \"search\", \"description\": \"Searches for info(...TRUNCATED) | {"keys": [["방책 coa", "coa"]]} | "{\"evidence_state\": \"grounded\", \"a_star\": \"finalize\", \"a_star_text\": \"## 중심 개념 X:(...TRUNCATED) |
stoppivot_2691996652_2 | "[{\"role\": \"system\", \"content\": \"Tool for browsing.\\nThe `cursor` appears in brackets before(...TRUNCATED) | "[{\"type\": \"function\", \"function\": {\"name\": \"search\", \"description\": \"Searches for info(...TRUNCATED) | {"keys": [["mario testino"], ["cecil beaton"]]} | "{\"evidence_state\": \"grounded\", \"a_star\": \"finalize\", \"a_star_text\": \"**Ellis Faas** is t(...TRUNCATED) |
stoppivot_1471553970_3 | "[{\"role\": \"system\", \"content\": \"Tool for browsing.\\nThe `cursor` appears in brackets before(...TRUNCATED) | "[{\"type\": \"function\", \"function\": {\"name\": \"search\", \"description\": \"Searches for info(...TRUNCATED) | "{\"keys\": [[\"vegetable margooga a slow-cooked vegetable and bread stew\", \"a slow-cooked vegetab(...TRUNCATED) | "{\"evidence_state\": \"grounded\", \"a_star\": \"finalize\", \"a_star_text\": \"Based on the inform(...TRUNCATED) |
stoppivot_823665821_3 | "[{\"role\": \"system\", \"content\": \"Tool for browsing.\\nThe `cursor` appears in brackets before(...TRUNCATED) | "[{\"type\": \"function\", \"function\": {\"name\": \"search\", \"description\": \"Searches for info(...TRUNCATED) | {"keys": [["first lady of senegal"]]} | "{\"evidence_state\": \"grounded\", \"a_star\": \"finalize\", \"a_star_text\": \"**First Lady of Sen(...TRUNCATED) |
End of preview. Expand in Data Studio
PivotRL-Asis-collection
Turn-level GRPO training data for Qwen3.5-35B-A3B-Asis-GRPO: 4,395 stop-decision pivot states (4,355 distinct questions) harvested from web-search agent trajectories of the Asis SFT policy. Each row is a serialized agent-conversation prefix that ends at a decision point where the agent must choose between stopping to answer and continuing to search, together with gold answer keysets used by a deterministic stop-verifier reward.
Construction
Pivot states were selected by K=8 rollout profiling: from each candidate mid-trajectory state, 8 continuations were rolled out with the SFT policy and scored by the deterministic stop-verifier; only mixed-outcome states (profiled success rate 0 < μ̂ < 1; observed range 0.125–0.875) were kept, so every state carries learning signal under group-relative advantage. Hard states (μ̂ < 0.25) additionally carry a conditional-seeding flag (allow_seed) that permits evidence-consistent hinting during RL rollouts. The set comprises 4,099 grounded finalize-decision states (evidence for the answer already present in the prefix) plus a 296-state control complement (ungrounded / out-of-distribution first-turn / next-final / long-context states) that regularizes against degenerate always-stop behavior.
Questions originate from an in-house web-search QA corpus (English and Korean); trajectories were generated by the SFT policy itself against a live browsing tool surface. No benchmark data was used at any stage (see Contamination below).
Fields
| Field | Type | Description |
|---|---|---|
id |
str | unique pivot id (stoppivot_<traj>_<turn>) |
prompt |
str (JSON) | serialized message list: browsing system prompt, user question, then alternating assistant / tool turns up to the pivot |
tools |
str (JSON) | function specs of the browsing tool surface: search, open, find |
label |
str (JSON) | {"keys": [[...], ...]} — gold answer keysets; each inner list is a set of acceptable surface forms, all groups must be matched for a finalize to be verified correct |
metadata |
str (JSON) | per-state annotations (below) |
metadata keys: evidence_state (grounded/ungrounded), a_star (reference action: finalize/toolcall), a_star_text (reference finalize text with evidence-citation markers), sft_action_name (action the SFT policy took at this state in the source trajectory), trajectory_id, assistant_depth (prefix depth in assistant turns), mu_profiled (μ̂ from K=8 rollout profiling), allow_seed (conditional-seeding flag), key_cursors (cursors of evidence pages containing the gold keys; grounded rows), track (A = grounded finalize states; C_* = control complement), plus auxiliary source/next_action on control rows.
from datasets import load_dataset
import json
ds = load_dataset("Minbyul/PivotRL-Asis-collection", split="train")
row = ds[0]
messages = json.loads(row["prompt"]); label = json.loads(row["label"]); meta = json.loads(row["metadata"])
Contamination screening
Both this file and its Correct-arm counterpart were screened row-by-row against three browsing benchmarks — LiveBrowseComp (335 questions), K-BrowseComp (300), and BrowseComp-EN (1,266) — by matching each row's first user message with three matchers: exact normalized equality, 120-char prefix containment, and 8-gram Jaccard ≥ 0.5. Result: 0 hits on all benchmarks and all matchers (rescan dated 2026-08-20; consistent with the corpus-wide screen of 2026-08-13). A field-level scrub scan additionally verified that no internal corpus identifiers, infrastructure paths, or generator-model references appear in system/user/instruction or metadata fields; model names occurring inside retrieved web text are genuine web content.
Associated models
- Trained on this data: Qwen3.5-35B-A3B-Asis-GRPO (turn-level GRPO, K=8, deterministic stop-verifier reward, KL 0.01 to the SFT reference)
- SFT policy / reference: Qwen3.5-35B-A3B-Asis
- Correct-arm counterpart dataset: PivotRL-Correct-collection
Intended use & limitations
Research artifact for studying stop-decision RL and over-search reduction in tool-using web agents. Prompts are tied to a specific browsing tool surface and citation format; reward semantics assume the deterministic stop-verifier described in the model cards. Tool-turn contents are snapshots of live web pages at generation time and may contain outdated or incorrect claims; they are included as environment observations, not as ground truth.
- Downloads last month
- 2