{"id":"86180c0d-47cf-431d-8775-758751315750","arxiv_id":"2502.07942","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"AgentSymbiotic couples a large LLM's trajectory generation with a small LLM's exploratory distillation to reach 52.1% and 48.5% success rates on WebArena for Claude-3.5 and a distilled 8B LLaMA model, respectively.","lead":"This paper proposes a training loop where a large and a small language model take turns generating web-navigation data and learning from each other, and it reports state-of-the-art scores on the WebArena benchmark for both model sizes. It matters because it suggests that small, cheap models can be iteratively improved to nearly match large proprietary models on complex web tasks, and that the two model types can reinforce each other.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'symbiotic' claim lacks an ablation isolating small-LLM trajectories in the large model's RAG; the reported SOTA numbers may hold, but the mutual-improvement mechanism is not established.","rationale":"The reader's weakest-assumption identification is correct and load-bearing. The framework is clearly described, and the Table 2 ablation does support the two distillation innovations for the small model, which is genuine evidence in favor of part of the contribution. No reproducible code or error bars are provided, which limits verification but does not by itself falsify the results. The decisive gap is on the large-LLM side: no experiment separates the contribution of small-LLM trajectories from the contribution of RAG data in general. Because the claimed novelty is precisely the bidirectional feedback loop, this missing ablation is the single most load-bearing concern. The reported numbers are plausible and the shortfall is missing evidence rather than demonstrated error, so the reader's CONDITIONAL verdict should stand unchanged rather than moving to reject.","tokens_in":33750,"tokens_out":5042,"duration_ms":45089,"concrete_test":"Run the full 812-task WebArena evaluation for the large LLM in three configurations: (a) AgentOccam without RAG, (b) AgentOccam + RAG where the knowledge base contains only large-LLM-generated trajectories, and (c) full AgentSymbiotic with small-LLM trajectories included. Repeat each with three independent runs at temperature 0.6 and report mean and standard error. If (b) and (c) match within noise, the symbiotic mechanism is not supported; if (c) significantly exceeds (b), the central claim survives. Also recompute Δ in Eq. 3 on the full task set with SR(ML) defined as configuration (b), not as a no-RAG baseline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that small-LLM exploration trajectories from Step 3, not merely additional data volume, drive the large LLM's improvement in Step 4. The only evidence is the synergy metric Δ in Eq. 3 and Figure 5, computed over three iterations on an unspecified subset of tasks. The definition compares the iterative system against max(SR(ML), SR(MS)), but the paper never states whether SR(ML) is the large LLM without RAG or with RAG using only large-LLM trajectories. If SR(ML) is AgentOccam without RAG, then any RAG gain from any trajectory source counts as 'synergy.' No ablation removes small-LLM trajectories from the retrieval library, and no error bars are reported despite the acknowledged temperature 0.6 and lack of multiple trials (Limitations). Thus the paper's central contribution—mutual enhancement through small-LLM exploration—is not isolated from a simpler 'accumulate more RAG data' explanation. The 52.1% large-LLM result and the distillation ablations in Table 2 may still be correct; the missing experiment is what distinguishes the symbiotic claim from a data-scaling claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AgentSymbiotic, an iterative framework for web-browsing agents in which a large LLM and a small LLM alternately improve each other: the large LLM generates high-quality trajectories for distillation and RAG, the distilled small LLM explores the environment and produces diverse trajectories that are fed back into the large LLM's retrieval knowledge base, and the loop repeats. Two distillation innovations are introduced: Speculative Data Synthesis, in which the teacher's top-K action candidates are used to accept or replace the student's proposed actions to mitigate off-policy bias, and multi-task learning that trains the student to predict both actions and reasoning rationales. A hybrid privacy-preserving mode delegates sensitive steps to a local small LLM. Experiments on WebArena report state-of-the-art success rates for both model classes: 52.1% for Claude-3.5-based agents and 48.5% for an 8B LLaMA-3 agent, compared with prior published results of 45.7% and 28% respectively. The paper also reports a synergy metric that increases over iterations, and ablations showing that each distillation component contributes to the small-model gains.","tokens_in":33965,"tokens_out":3764,"duration_ms":33957,"significance":"If the reported results are confirmed, AgentSymbiotic would be a practically useful contribution to web-agent research: it couples data synthesis and task performance in a way that improves both large and small models, and it demonstrates that an 8B open-weight model can approach the performance of much larger closed models on WebArena. The two distillation techniques (speculative data synthesis and multi-task reasoning training) are clearly described and individually ablated in Table 2, which is a strength. The paper is also transparent about its limitations, explicitly acknowledging budget constraints, the lack of multiple trials, and the absence of reproducibility of all baselines. However, the central 'symbiotic' mechanism—the claim that small-LLM exploration is what drives large-LLM improvement—is not isolated by any ablation, and all headline numbers are single-run point estimates at temperature 0.6. These two issues materially affect the strength of the central claims, though they do not invalidate the possibility that the reported SOTA numbers are correct.","major_comments":[{"comment":"The synergy metric Δ is defined as SR(iter)(ML, MS) − max(SR(ML), SR(MS)), but the paper never specifies what SR(ML) refers to: is it the large LLM without RAG, or the large LLM with RAG using only large-LLM-generated trajectories? Without this baseline, a positive Δ could arise merely from adding more RAG data rather than from the small LLM's exploratory trajectories. This is load-bearing for the paper's central 'symbiotic improvement' claim. Please add an ablation that removes small-LLM trajectories from the knowledge base (or equivalently, compare a RAG library built only from large-LLM trajectories with one that also includes small-LLM trajectories), and state explicitly which baseline is used for SR(ML) in Eq. (3).","section":"§3.1, Eq. (3), and Figure 5"},{"comment":"All success rates are reported as single point estimates without variance or confidence intervals, and the paper acknowledges that at temperature 0.6 results may vary across attempts. Several headline comparisons are small—52.1% vs. 48.5% for the large model, and 48.5% vs. 43.6% between the two small models—so the reported differences may not be statistically meaningful. Please run multiple trials (or at least multiple seeds for the distillation and evaluation) and report standard errors or confidence intervals, and state whether the reported numbers are the mean or the best run.","section":"Limitations (Temperature Setting), and Tables 1–2"},{"comment":"The abstract and §1 claim that the large LLM 'surpasses the previous best of 45%,' but Table 1 lists AgentOccam Claude-3.5 at 48.5% (shown without an asterisk, which appears to indicate the authors' own measurement rather than a cited score). If 48.5% is a prior result, the claim should be updated to reflect a margin of 3.6 points over 48.5%, not over 45%. If 48.5% is the authors' replication of AgentOccam Claude-3.5, the provenance should be stated, and the comparison should be framed accordingly. Please clarify this discrepancy.","section":"Abstract, §1, and Table 1"},{"comment":"Figure 5 reports the synergy metric Δ over three iterations on 'a subset of WEBARENA tasks,' but the size and composition of this subset are not given, and per-iteration success rates are not provided. As written, the figure cannot be used to assess whether the increase in Δ is meaningful or generalizes to the full benchmark. Please specify the subset size, the task domains, and the underlying SR values for each iteration, and if possible compute Δ on the full 812-task benchmark.","section":"§5.1, Figure 5"}],"minor_comments":[{"comment":"The inequality E(ML, T) > E(MS, T) is described as an empirical observation, but no experimental evidence or citation is given for this specific claim; consider adding a reference or softening the wording.","section":"§3.1, Eq. (1)"},{"comment":"When the student's action is rejected and replaced by 'Best(ak)' from the teacher's K candidates, the criterion for selecting the best candidate is not defined; please specify the ranking procedure (e.g., teacher log-probabilities, or a separate judge).","section":"Algorithm 2, line 16"},{"comment":"The privacy-detection prompt contains a typo: 'not a PII,not a sensitive information or anythong else' should be corrected to 'not PII, not sensitive information, or anything else.'","section":"Appendix G"},{"comment":"The benchmark name is written inconsistently as 'WEBARENA,' 'WebArena,' and 'Webarena'; please standardize to a single convention throughout the text and tables.","section":"Throughout"},{"comment":"The transition between the Speculative Data Synthesis flow and the Multi-task Learning flow is visually dense and not clearly separated; adding labels for the data flow direction and the training loop would improve readability.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a plausible and useful framework, and the two distillation ablations are a genuine strength. The main weakness is that the 'symbiotic' mechanism is asserted on the basis of the synergy metric without a controlled ablation, and the headline numbers lack variance estimates. These are fixable with additional experiments. I would also double-check the AgentOccam Claude-3.5 48.5 entry in Table 1 against the abstract's claim of surpassing 45%, as this discrepancy will likely be noticed by reviewers and readers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this about arXiv:2502.07942: the paper reports strong WebArena numbers for both a Claude-3.5-based large agent (52.1%) and an 8B Llama-3 distill (48.5%), and the small-model distillation is genuinely well ablated. The weak point is the paper's headline mechanism: \"symbiotic\" improvement is not actually isolated. The authors claim small-LLM exploration enriches the large model's RAG, but they never ablate away the small-LLM trajectories. So the mutual-benefit story could just be \"more RAG data helps.\" That doesn't make the numbers wrong, but it makes the central contribution under-supported.\n\nWhat's new: the closed-loop coupling of large-LLM RAG, small-LLM exploration, speculative action-candidate filtering during distillation, and multi-task CoT training. The individual pieces are known—especially speculative knowledge distillation (Xu et al. 2024b, which they cite)—but the combination and the results are new. The ablation in Table 2 is the best part: plain SFT gets 40.8%, adding speculative gets 46.8%, adding multi-task gets 43.2%, both together 48.5%, and the per-domain breakdown is informative. That segment supports the distillation claims.\n\nWhere it's soft: First, the symbiotic mechanism. The delta metric in Eq. 3 is computed over only three iterations on an unspecified subset, and they don't define SR(ML) unambiguously (without RAG? with RAG but only large trajectories?). A single ablation removing small-LLM trajectories from the knowledge base would settle it. Second, no variance: they set temperature to 0.6 and state in Limitations they couldn't run multiple trials, so the headline scores are point estimates. Third, the baseline comparison: the abstract says \"surpassing the previous best of 45%\" but Table 1 lists an AgentOccam Claude-3.5 score of 48.5; the 45.7 comparison is to the best open-source method (AgentOccam-Judge), which is stated in the Limitations but not in the abstract. That's sloppy but minor. Fourth, single benchmark and no code release.\n\nWho it's for: people working on web-agent distillation or RAG for agents. It deserves a serious referee, but the referee should require the missing ablation and variance estimates. My recommendation: engage with it, send it to review with the expectation of heavy revision.","headline":"Strong, well-ablated distillation results on WebArena, but the paper's headline 'symbiotic' mechanism is not isolated from a simpler data-scaling explanation.","tokens_in":34536,"tokens_out":2147,"would_cite":true,"duration_ms":19607,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A symbiotic loop coupling a large LLM's trajectory generation with a small LLM's exploratory divergence lifts both to state-of-the-art WebArena success rates of 52% and 49%.","keywords":["web agents","knowledge distillation","retrieval-augmented generation","speculative data synthesis","multi-task learning","WebArena","symbiotic improvement","privacy preservation"],"falsifier":"Run the agent loop on the full 812 WebArena tasks with a control knowledge base that replaces every small-LLM trajectory with an equal number of large-LLM trajectories matched by task and length; if the large model's success rate does not drop when the small trajectories are excluded, the small model's exploration is not the causal driver and the 'symbiotic' attribution fails.","tokens_in":33510,"feed_emoji":"🤝","tokens_out":12865,"duration_ms":89958,"temperature":0.7,"pith_summary":"This paper claims that a large and a small language model can be coupled in an iterative loop so that each improves the other, and that the pair reaches state-of-the-art results on the WebArena benchmark for both model classes. The large model (Claude-3.5) generates high-quality navigation trajectories and retrieves past experiences through a retrieval-augmented generation (RAG) store; the distilled small model (Llama-3.1-8B) explores the environment more cheaply and, because its action choices diverge from the large model's, discovers new trajectories that are fed back into the large model's retrieval base. Two distillation innovations—speculative data synthesis that filters student actions against teacher candidates, and multi-task learning that trains the student to predict both actions and reasoning—are introduced to keep the small model from becoming a bottleneck. The reported success rates are 52% for the large agent and 49% for the 8B agent, surpassing prior bests of 45% and 28% respectively. If the mechanism holds, it would mean that small models can contribute more than they receive in agent training pipelines.","feed_headline":"Symbiotic agent loop hits 52% on WebArena","feed_subtitle":"An 8B distilled model reaches 49% by feeding its divergent explorations back to the large model.","key_machinery":"The load-bearing mechanism is the iterative 'symbiotic' loop itself, plus the two distillation techniques that keep the small model viable inside the loop. Speculative data synthesis works as a teacher–student action filter at each step: the small model proposes an action-reason pair, the large model generates $K$ candidate actions, and the student's proposal is accepted only if it falls inside the teacher's candidate set, otherwise the teacher's best candidate is executed; this shifts from near-supervised distillation early on to more speculative acceptance as the student improves. Multi-task learning augments the distillation objective by training the small model to predict both the next action and a chain-of-thought (step-by-step) rationale, preserving reasoning capacity that plain action cloning would discard. The RAG knowledge base, built from subsequences of validated trajectories with instruction and summary annotations and retrieved by three complementary strategies (task-guided summary retrieval, direct observation–instruction matching, and trajectory similarity search), is the conduit through which small-model discoveries reach the large model.","core_discovery":"The central discovery is that large and small LLMs form a complementary exploration–exploitation pair for web navigation: large LLMs exploit well-understood situations with high action accuracy, while distilled small LLMs, being faster and more variable in their action selection, cover more of the state–action space and stumble onto unorthodox or edge-case solutions. AgentSymbiotic institutionalizes this complementarity in a four-step cycle—large-LLM trajectory generation with RAG, multi-LLM-debate-filtered distillation with speculative data synthesis and multi-task learning, small-LLM exploration, and merging of the new trajectories back into the knowledge base. The authors define a synergy metric $\\Delta$ as the excess success rate of the iterative scheme over the best single-LLM baseline, and report that $\\Delta$ grows over three iterations on a subset of WebArena tasks. They further report that the full pipeline reaches 52% success rate for Claude-3.5 and 49% for a distilled 8B LLaMA-3 model on all 812 WebArena tasks, exceeding the previous open-source bests.","pith_inferences":["The paper's numbers leave open whether the symbiosis is causal or a data-volume effect: swapping small-LLM trajectories for an equal number of large-LLM ones in the retrieval base would distinguish 'divergent exploration helps' from 'more examples help.'","If divergence itself is the driver, deliberately increasing the student's action stochasticity during exploration (higher temperature, action-level dropout) could amplify the effect; the paper fixes temperature at 0.6 and does not sweep it.","The framework's design transfers to any environment with a teacher, a student, and a retrieval store, but the paper evaluates only WebArena, so transferability is a design inference rather than a demonstrated result.","The hybrid mode's modest 1.6-point drop relative to cloud-only inference suggests privacy protection is cheap here; comparing selective delegation against a fully local small agent would show how much of the gap is recovered."],"forward_implications":["If the reported gains replicate, an 8B-parameter web agent lands within about three points of a Claude-3.5-level agent, making small local agents a realistic option for web automation on cost or privacy grounds.","The iterative loop implies web-agent performance can keep improving with added rounds of coupling rather than saturating after a single distillation pass; the paper reports the synergy metric $\\Delta$ rising across three iterations on a subset of tasks.","The ablation isolates each distillation innovation: speculative data synthesis alone lifts a plain-SFT 8B model from 40.8% to 46.8% on WebArena, and combining it with multi-task learning reaches 48.5%.","The hybrid privacy mode routes steps flagged as private to a local small model and carries a measured cost of about 1.6 points (50.5% combined vs 52.1% for the non-hybrid large agent) while keeping sensitive data on-device."],"supporting_citations":[{"why":"Provides the WebArena benchmark and its 812-task evaluation protocol used for all success-rate comparisons.","marker":"Zhou et al., 2023b"},{"why":"AgentOccam is the base framework the large-agent pipeline is built on and the previous open-source best (45.7%) the paper surpasses.","marker":"Yang et al., 2024b"},{"why":"Learn-by-Interact supplies the data-centric trajectory-synthesis paradigm and is the prior small-model baseline (28% for a 22B model) the distillation results exceed.","marker":"Su et al., 2025"},{"why":"Speculative knowledge distillation is the direct antecedent of the paper's speculative data synthesis for mitigating off-policy bias.","marker":"Xu et al., 2024b"},{"why":"Multi-agent debate is the mechanism used to filter and select high-quality trajectories for the knowledge base and distillation set.","marker":"Liang et al., 2024"},{"why":"Identifies off-policy bias in fine-tuning web agents, the root cause the speculative synthesis strategy is designed to correct.","marker":"Caccia et al."},{"why":"DeepSeek-R1 is the local model used as the privacy detector in the hybrid mode.","marker":"Guo et al., 2025"},{"why":"Chain-of-thought reasoning prompting provides the rationale-generation target used by the multi-task learning objective.","marker":"Wang and Zhou, 2024"}],"fun_headline_variants":["Symbiotic loop lifts WebArena to 52%","Small LLMs explore, big LLMs exploit: 52% on WebArena","AgentSymbiotic: Divergent paths, dual gains (52%/49%)","Iterative symbiosis: 52% large, 49% small on WebArena"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that the loop is genuinely symbiotic rests on the assumption that the small model's exploratory trajectories, rather than the sheer increase in retrieval data, are what drive the large model's improvement—an assumption the paper supports only with a subset-scale synergy metric and no ablation that removes small-model trajectories from the knowledge base.","fun_headline_variants_meta":{"raw":{"variants":["Symbiotic loop lifts WebArena to 52%","Small LLMs explore, big LLMs exploit: 52% on WebArena","AgentSymbiotic: Divergent paths, dual gains (52%/49%)","Iterative symbiosis: 52% large, 49% small on WebArena"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000295,"raw_usage":{"total_tokens":1786,"prompt_tokens":1090,"completion_tokens":696,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":706,"completion_tokens_details":{"reasoning_tokens":610}},"tokens_in":706,"tokens_out":696,"duration_ms":5960,"temperature":1.0,"reasoning_tokens":610,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T11:20:40.347368+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the agent loop on the full 812 WebArena tasks with a control knowledge base that replaces every small-LLM trajectory with an equal number of large-LLM trajectories matched by task and length; if the large model's success rate does not drop when the small trajectories are excluded, the small model's exploration is not the causal driver and the 'symbiotic' attribution fails.","supporting_citations":[],"review_version":1}