verification.json

{
  "round": 11,
  "date": "2026-08-01",
  "papers": {
    "2606.04703": {
      "title": "Rethinking Continual Experience Internalization for Self-Evolving LLM Agents",
      "version": "v1",
      "md5_pdf": "be70e2148334a4eefa0106f3fc84bc01"
    }
  },
  "protocol": "全文 PDF→pdftotext；水印校验；4 主张 × 3 票 = 12 票",
  "verdict_summary": "CE1 坍缩现象 3-0 确认（3 轮掉到 base 下）；CE2 1C/2CORR（teacher=自教师+rejection sampling，非外部强模型）；CE3 0C/3CORR（稳定配方含 rejection sampling 正确性门；坍缩归因非无差别写入）；CE4 3-0（无 per-user 隔离）",
  "identities": [
    "PASSED — watermark \"arXiv:2606.04703v1 [cs.CL] 3 Jun 2026\" present in the text; paper is \"Rethinking Continual Experience Internalization for Self-Evolving LLM Agents\" (Chen, Yang, Fan et al., RUC/Beihang/Meituan).",
    "PASSED — line 27 of the pdftotext contains the watermark \"arXiv:2606.04703v1 [cs.CL] 3 Jun 2026\"; title \"Rethinking Continual Experience Internalization for Self-Evolving LLM Agents\" (RUC/Beihang/Meituan).",
    "PASSED — line 27 of the pdftotext output contains the watermark \"arXiv:2606.04703v1 [cs.CL] 3 Jun 2026\", matching the requested paper. Title on p.1: \"Rethinking Continual Experience Internalization for Self-Evolving LLM Agents\"."
  ],
  "tally": {
    "CE1": {
      "CONFIRMED": 3,
      "REFUTED": 0,
      "CORRECTED": 0,
      "UNSUPPORTED": 0,
      "evidence": [
        "Abstract: \"Experience internalization converts contextual experience from past interactions into reusable parametric capability\" and \"we discover that under multi-iteration experience learning, existing methods suffer from a progressive capability collapse rather than compounding improvement.\" Figure 1 caption: \"Performance degradation under iterative on-policy context-distillation.\" Formulation section names the framework \"Continual Experience Internalization\" (iterative process k = 0..K; Eq. 3). Appendix B: \"We run self-evolution for three internalization iterations.\" Benchmarks: WebWalkerQA (in-domain), GAIA-Text-103, BrowseComp-ZH (out-of-domain); models Qwen3-4B-Instruct-2507 and Qwen3-8B; 15K training corpus from five web-QA datasets; ReAct agent with 5 tools. Magnitude (Table 4): e.g. Qwen-generated experience + global injection + off-policy: WebWalkerQA 21.0 → 18.7 → 8.5 (base 16.6), GAIA 15.9 → 11.3 → 6.5 (base 13.6), BrowseComp-ZH 3.5 → 2.1 → 0.7 (base 4.5) — collapse well below the base model by iteration 3. DeepSeek + global + off-policy: WebWalkerQA 25.9 → 31.0 → 12.8. On-policy even under the best experience config degrades: 4B step-wise DeepSeek 35.0 → 32.4 → 31.5 (GAIA 28.8 → 27.2 → 25.6); 8B on-policy BrowseComp-ZH 4.5 → 1.4.",
        "Abstract: \"Experience internalization converts contextual experience from past interactions into reusable parametric capability\"; \"we discover that under multi-iteration experience learning, existing methods suffer from a progressive capability collapse rather than compounding improvement.\" Section 3 names the iterated setting \"Continual Experience Internalization\" (\"an iterative process indexed by k = 0, 1, . . . , K\"; θ(k+1) = Internalize(θ(k), E(k))). Evidence: Figure 1 \"Performance degradation under iterative on-policy context-distillation\" on WebWalkerQA, GAIA (GAIA-Text-103), and BrowseComp-ZH; Appendix B: \"We run self-evolution for three internalization iterations.\" Magnitudes (Table 4, Qwen3-4B off-policy w/ global injection): WebWalkerQA 21.0→18.7→8.5 vs base 16.6; GAIA 15.9→11.3→6.5 vs base 13.6; BrowseComp-ZH 3.5→2.1→0.7 vs base 4.5; DeepSeek-experience global: WebWalkerQA 25.9→31.0→12.8. Collapse = falling below the untrained base model by iteration 3.",
        "Abstract: \"Experience internalization converts contextual experience from past interactions into reusable parametric capability, offering a promising path toward continual learning\" and \"under multi-iteration experience learning, existing methods suffer from a progressive capability collapse rather than compounding improvement.\" Section 3 formalizes \"Continual Experience Internalization\" as an iterative process k=0..K where trajectories are summarized into experience pool E(k) and internalized into the next student. Figure 1 caption: \"Performance degradation under iterative on-policy context-distillation\" showing decline across Iteration 1/2/3 on WebWalkerQA, GAIA, BrowseComp-ZH. Setup: Qwen3-4B-Instruct-2507 and Qwen3-8B students, ReAct agent with 5 tools, 3 internalization iterations (Appendix B: \"We run self-evolution for three internalization iterations\"). Benchmarks: WebWalkerQA (in-domain), GAIA-Text-103, BrowseComp-ZH (out-of-domain). Magnitude (Table 4, internalized inference): e.g. Qwen-generated experience + global injection + off-policy: WebWalkerQA 21.0 -> 18.7 -> 8.5 (base 16.6); GAIA 15.9 -> 11.3 -> 6.5 (base 13.6); DeepSeek-generated + global: WebWalkerQA 25.9 -> 31.0 -> 12.8; on-policy DeepSeek+global: 29.0 -> 22.5 -> 19.9. Instance-level experience \"quickly diminish[es]... and fall[s] below the base model\" (Sec 5.1)."
      ],
      "notes": [
        "Exact phenomenon wording is \"progressive capability collapse\" under \"multi-iteration experience learning\"; the process is termed \"continual experience internalization\". The headline Figure 1 collapse is specifically iterative ON-POLICY context-distillation, but Table 4 shows collapse also under off-policy with global injection — the paper's point is that collapse afflicts existing configurations generally, fixed only by the combined recipe.",
        "The paper has no single coined term beyond \"progressive capability collapse\"; the studied process is called \"continual experience internalization\" / \"experience-based self-evolution\". Rounds = 3 iterations; models Qwen3-4B-Instruct-2507 and Qwen3-8B; ReAct agent with 5 tools (Search, Visit, Python, Scholar, File Parser).",
        "Exact phenomenon name for the failure is \"progressive capability collapse\"; \"continual experience internalization\" is the name of the studied process (Sec 3), not the failure. Collapse applies to EXISTING methods (esp. on-policy context-distillation and global injection); the paper's own recipe (principle-level + step-wise + off-policy) is shown stable (Fig 6, Table 4: e.g. 30.6 -> 30.7 -> 33.1 on WebWalkerQA)."
      ]
    },
    "CE2": {
      "CONFIRMED": 1,
      "REFUTED": 0,
      "CORRECTED": 2,
      "UNSUPPORTED": 0,
      "evidence": [
        "(a) Abstract: \"principle-level experience is more durable than instance-level experience, as it effectively abstracts transferable strategies away from trajectory-specific details\"; §5.1: instance-level \"yields only transient gains... quickly diminish as self-evolution proceeds and fall below the base model\" (74.4% of instance items contain specific URLs/domains, 93.9% query/entity-specific strings vs 84.0% of principle items containing reusable strategy-like statements). (b) Abstract: \"step-wise injection significantly outperforms global injection by aligning experience with intermediate decision states, a property that is critical for long-horizon tool use\"; Table 1: step-wise vs global 31.2 vs 23.2 (WebWalkerQA), 22.7 vs 16.8 (GAIA), 5.2 vs 4.5 (BC-ZH); Table 2 premature-answer rate 63.82% (global) vs 0% (step-wise). (c) Abstract: \"off-policy context-distillation on high-quality teacher trajectories provides a substantially more stable training signal than on-policy context-distillation, which is inherently limited by local corrections on student-induced flawed states.\" §5.3.1: \"Off-policy internalization instead samples trajectories directly from the experience-aware teacher (i.e., the student conditioned on step-wise experience) and applies rejection sampling to retain successful trajectories.\"",
        "(a) CONFIRMED — Abstract: \"principle-level experience is more durable than instance-level experience, as it effectively abstracts transferable strategies away from trajectory-specific details\"; §5.1: instance-level gains \"quickly diminish as self-evolution proceeds and fall below the base model\" (74.4% of instance items contain specific URLs/domains; 84.0% of principle items contain reusable strategy statements vs 3.7%). (b) CONFIRMED — Abstract: \"step-wise injection significantly outperforms global injection by aligning experience with intermediate decision states\"; Table 1: WebWalkerQA 23.2→31.2 (+8.0); Table 2: premature-answer rate 63.82% (global) vs 0% (step-wise). (c) Paper: \"off-policy context-distillation on high-quality teacher trajectories provides a substantially more stable training signal than on-policy context-distillation, which is inherently limited by local corrections on student-induced flawed states.\"",
        "(a) Abstract: \"principle-level experience is more durable than instance-level experience, as it effectively abstracts transferable strategies away from trajectory-specific details\"; Sec 5.1: instance-level \"yields only transient gains... fall below the base model\" (74.4% of instance items contain specific URLs/domains, 93.9% query/entity-specific strings vs 84.0% of principle items containing reusable strategy-like statements). (b) Abstract: \"step-wise injection significantly outperforms global injection by aligning experience with intermediate decision states, a property that is critical for long-horizon tool use\"; Table 1: single-iteration WebWalkerQA 23.2 -> 31.2 (+8.0); Table 2: premature-answer rate 63.82% (global) vs 0% (step-wise) at iteration 3. (c) Abstract: \"off-policy context-distillation on high-quality teacher trajectories provides a substantially more stable training signal than on-policy context-distillation, which is inherently limited by local corrections on student-induced flawed states\"; Sec 5.3.1: on-policy supervision \"is fundamentally reactive... the teacher can only provide corrections on states that may already be inefficient or off target\", while off-policy \"provides proactive experience-guided supervision\"."
      ],
      "notes": [
        "Legs (a) and (b) are exact. Leg (c) needs one correction: the \"teacher\" is NOT an external stronger model — it is the SAME policy conditioned on the experience pool (\"the experience-aware teacher (i.e., the student conditioned on step-wise experience)\"). \"High-quality\" refers to the trajectories (teacher-generated, then rejection-sampled to keep successful ones), not to a superior teacher model. Corrected statement: off-policy context-distillation on rejection-filtered successful trajectories generated by the experience-conditioned self-teacher is more stable than on-policy distillation on student-induced trajectories.",
        "Leg (c) needs correction on two points: (1) the \"teacher\" is NOT an external stronger model — it is the same student policy conditioned on experience (§5.3.1: \"the experience-aware teacher (i.e., the student conditioned on step-wise experience)\"); \"high-quality\" refers to teacher-generated TRAJECTORIES made high-quality via rejection sampling (\"applies rejection sampling to retain successful trajectories\"), not to a high-quality teacher model. (2) The paper's precise claim is stability across iterations: on-policy \"delivers strong gains in a single iteration but fails to sustain them\" (Table 4: on-policy step-wise WebWalkerQA 35.0→32.4→31.5 declining vs off-policy 30.6→30.7→33.1 rising). Also the paper's word is \"injection\", not \"insertion\", and \"step-wise\" refers to per-step experience selection by an LLM selector Rϕ during teacher supervision construction.",
        "All three legs confirmed as stated, with one precision on (c): the \"teacher\" is NOT an external stronger model — it is the same student policy conditioned on (step-wise) experience (Sec 5.3.1: \"the experience-aware teacher (i.e., the student conditioned on step-wise experience)\"). \"High-quality\" is achieved via rejection sampling of the teacher's trajectories, not via a stronger teacher model."
      ]
    },
    "CE3": {
      "CONFIRMED": 0,
      "REFUTED": 0,
      "CORRECTED": 3,
      "UNSUPPORTED": 0,
      "evidence": [
        "Sweep results: admission=0, admit=0, threshold=0, correctness=0, criterion=0, consensus=0, vote=0, discard=0, prune=0, screen=0, curat=0, score=0, confidence=0, trust=0, factual=0, majority=0; gate/gating raw hits are false positives (\"mitigates\", \"Navigating\"). EXPERIENCE-POOL write path has NO correctness gate: \"we summarize trajectories into natural-language experience with DeepSeek-V4\" and Appendix B \"each iteration refreshes the experience pool using trajectories generated by the current model\" — no success/quality filter on what enters E(k); principle-level experience even deliberately retains \"failure patterns\". Limitations confirms gating is unexplored: \"other factors such as experience-pool size, selector quality, and filtering criteria may also affect stability. We leave a more comprehensive exploration of these factors to future work.\" HOWEVER, the pipeline is not gate-free end-to-end: (i) correctness/quality filter on internalized training data — off-policy regime \"applies rejection sampling to retain successful trajectories\" / \"trains on rejection-filtered teacher-generated trajectories\" / \"applying rejection sampling to filter low-quality variants\" (this is part of their FINAL recipe, absent from the collapsing on-policy baselines); (ii) relevance selection — LLM-based selector Rϕ \"selects experience according to the current interaction history\" for step-wise injection (relevance/state-alignment, not correctness); (iii) evaluation-side: task-level reward r(HT) used for benchmark metrics. Attributed causes of collapse: instance-level trajectory-specific traces that \"transfer poorly\"; global injection's state misalignment causing the premature-answer failure mode (63.82%) and loss of experience-use ability that \"destabilize[s] the model–experience loop\"; on-policy supervision being \"fundamentally reactive... the teacher can only provide corrections on states that may already be inefficient or off target\" — reduced to \"local corrections on flawed states\".",
        "Exhaustive sweep run (admission 0, admit 0, gate/gating 0 real hits — only 'mitigates'/'Navigating', threshold 0, correctness 0, curat 0, score 0, etc.). Non-zero hits classified: (A) \"applies rejection sampling to retain successful trajectories\" / \"applying rejection sampling to filter low-quality variants\" / Appendix B: \"off-policy context-distillation trains on rejection-filtered teacher-generated trajectories\" → a SUCCESS/CORRECTNESS-based filter on the distillation training trajectories in the off-policy regime (class i, applied to training-data admission, not experience-pool writes). (B) \"an LLM-based selector Rϕ selects experience according to the current interaction history\" (step-wise injection) → relevance selection, class ii. (C) §5.1 \"filtering out such local artifacts\" → describes abstraction of principle-level experience, not a gate. (D) Limitations: \"other factors such as experience-pool size, selector quality, and filtering criteria may also affect stability. We leave...to future work\" → confirms no pool-level filtering was studied. (E) Broader Impact: \"Practical deployment should therefore include trajectory filtering, experience-pool auditing, human oversight\" → recommendation only, not part of their pipeline. Experience-pool construction itself is ungated: \"These trajectories are summarized into an experience pool E(k)\" / \"each iteration refreshes the experience pool using trajectories generated by the current model\" — no correctness check; principle-level experience explicitly includes \"failure patterns\", implying failed trajectories are ingested too.",
        "Exhaustive sweep run (word-boundary + substring) over all 29 requested keywords; zero hits for admission/admit/correctness/threshold/criterion/factual/trust/confidence/consensus/agreement/majority/vote/discard/prune/screen/curat/score; gate/gating hits are the substring false positive \"mitigates/mitigate\"; all filter/select/reject/quality/verify/criteria hits read in context. Counterevidence to the strict negative: the recommended off-policy regime DOES contain a success-based admission gate on distillation data — Sec 5.3.1: \"Off-policy internalization instead samples trajectories directly from the experience-aware teacher... and applies rejection sampling to retain successful trajectories\"; Sec 5.3.2: \"applying rejection sampling to filter low-quality variants\"; Appendix B: \"off-policy context-distillation trains on rejection-filtered teacher-generated trajectories.\" Classification of every selection step found: (i) correctness/quality write-admission gate: rejection sampling retaining only successful (reward-passing) teacher trajectories before training — present ONLY in the off-policy regime; (ii) relevance/format selection: the LLM-based step-wise selector R-phi \"selects experience according to the current interaction history\" (Sec 3) — relevance, not correctness; principle-level abstraction \"filtering out such local artifacts\" (Sec 5.1) — format/abstraction, not correctness; (iii) evaluation-side: none beyond standard Pass@1/accuracy metrics; (iv) unrelated: \"trajectory filtering, experience-pool auditing\" appear only as RECOMMENDED future safeguards in Broader Impact, and Limitations lists \"filtering criteria\" as unexplored future work."
      ],
      "notes": [
        "Corrected negative claim: TRUE that no correctness/quality-based admission gate exists on the EXPERIENCE POOL (writes are unfiltered summarization; filtering criteria explicitly deferred to future work), and the collapsing on-policy baselines internalize without any correctness gating. FALSE as a blanket statement about all pipelines studied: the paper's recommended off-policy regime rejection-samples for SUCCESSFUL trajectories — a category-(i) correctness gate on the distillation data (though not on the experience pool). Note collapse also occurs under off-policy + global injection (Table 4: 25.9→31.0→12.8), so the paper blames representation/injection/regime misalignment, not indiscriminate writing or data quality per se.",
        "Corrected statement: the EXPERIENCE POOL has no correctness/quality admission gate (all current-model trajectories, including failures, are summarized in), and the ON-POLICY pipelines — where the collapse of Figure 1 occurs — have no quality gating at all. But the blanket negative is false for the paper's recommended OFF-POLICY regime, which gates distillation data via rejection sampling on task success. Cause attribution: the paper does NOT blame missing quality gating or indiscriminate writing; it blames (1) instance-level experience's trajectory-specific artifacts that \"transfer poorly\", (2) global injection's misalignment with decision states, causing premature answering (63.82%) and loss of experience-use ability, destabilizing the model–experience loop, and (3) on-policy supervision being \"fundamentally reactive\" — \"the teacher is reduced to local corrections on flawed states, rather than coherent demonstrations of experience-guided behavior.\"",
        "Corrected statement: (1) The EXPERIENCE POOL admission has no correctness/quality gate — trajectories (including failures; principle-level experience explicitly abstracts \"failure patterns\") are summarized into the pool by DeepSeek-V4/Qwen with no verification, and Limitations concedes \"filtering criteria\" were not studied. (2) The on-policy pipelines that exhibit the collapse likewise have no correctness gating anywhere. (3) BUT the paper's stable off-policy recipe includes rejection sampling that retains only successful teacher trajectories — a correctness/quality-based gate on what gets distilled into parameters — so the blanket claim \"NO correctness/quality-based admission gating\" is false for the full set of pipelines studied. Cause attribution: collapse is blamed NOT on indiscriminate writing but on (a) instance-level experience's trajectory-specific artifacts that \"transfer poorly\" and risk \"reinforcing instance-specific behaviors across iterations\"; (b) global injection's state misalignment, which causes premature answering and \"weakens experience-use ability\", destabilizing \"the model-experience loop\"; (c) on-policy supervision being \"fundamentally reactive\" — the teacher \"reduced to local corrections on flawed states, rather than coherent demonstrations of experience-guided behavior.\""
      ]
    },
    "CE4": {
      "CONFIRMED": 3,
      "REFUTED": 0,
      "CORRECTED": 0,
      "UNSUPPORTED": 0,
      "evidence": [
        "Title: \"Rethinking Continual Experience Internalization for Self-Evolving LLM Agents\". Authors: Jingwen Chen*, Wenkai Yang*, Shengda Fan (1: Gaoling School of Artificial Intelligence, Renmin University of China); Wenbo Nie (2: School of Software, Beihang University); Chenxing Sun, Shaodong Zheng, Yangen Hu, Lu Pan, Ke Zeng (3: Meituan); Yankai Lin† (1, corresponding; * equal contribution). Contact: cjw259wen@outlook.com, yankailin@ruc.edu.cn. Code: github.com/RUCBM/ExpInternalization. Per-user isolation sweep: \"interlocutor\"=0 hits, \"personal\"=0 hits, \"user\"=2 hits, both non-isolation contexts (\"Given a user query x, at each step t...\" in the formulation, and \"for user goal Extract the first recorded year...\" inside the Figure 4 case-study transcript). No experiment involves per-user or per-interlocutor isolation of internalized knowledge — the experience pool is a single global pool per iteration shared across all queries.",
        "Title: \"Rethinking Continual Experience Internalization for Self-Evolving LLM Agents\". Authors: Jingwen Chen*, Wenkai Yang* (equal contribution), Shengda Fan (1: Gaoling School of Artificial Intelligence, Renmin University of China); Wenbo Nie (2: School of Software, Beihang University); Chenxing Sun, Shaodong Zheng, Yangen Hu, Lu Pan, Ke Zeng (3: Meituan); Yankai Lin† (1, corresponding). Contact: cjw259wen@outlook.com, yankailin@ruc.edu.cn. Code: github.com/RUCBM/ExpInternalization. Per-user isolation sweep: \"interlocutor\" 0 hits; \"personal\" 0 hits; \"user\" appears only as \"Given a user query x\" (§3 formulation) and inside a Figure 4 tool-response snippet (\"for user goal Extract the first recorded year\"). No experiment involves per-user or per-interlocutor isolation of internalized knowledge — the experience pool E(k) is a single global pool per iteration, shared across all queries.",
        "Title: \"Rethinking Continual Experience Internalization for Self-Evolving LLM Agents\". Authors: Jingwen Chen*, Wenkai Yang* (equal contribution), Shengda Fan, Wenbo Nie, Chenxing Sun, Shaodong Zheng, Yangen Hu, Lu Pan, Ke Zeng, Yankai Lin (corresponding). Affiliations: (1) Gaoling School of Artificial Intelligence, Renmin University of China — Chen, Yang, Fan, Lin; (2) School of Software, Beihang University — Nie; (3) Meituan — Sun, Zheng, Hu, Pan, Zeng. arXiv:2606.04703v1 [cs.CL] 3 Jun 2026. Code: github.com/RUCBM/ExpInternalization. Per-user isolation sweep: \"interlocutor\" 0 hits; \"personal\" 0 hits; \"user\" appears only in Sec 3 (\"Given a user query x\") and inside a scraped URL summary in Figure 4 (\"for user goal Extract...\") — no experiment isolates internalized knowledge per user or per interlocutor; all experience pools are global per model iteration."
      ],
      "notes": [
        "As expected: no per-user/per-interlocutor isolation anywhere; all internalization operates on one shared experience pool E(k) per iteration.",
        "arXiv:2606.04703v1 [cs.CL], 3 Jun 2026.",
        "As expected: no per-user/per-interlocutor isolation of internalized knowledge anywhere in the paper — the experience pool E(k) is a single shared pool refreshed per self-evolution iteration."
      ]
    }
  }
}