{"id":"ee850b3c-6b8e-49a8-b590-0a0f9aab8079","arxiv_id":"2607.21461","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A recursively self-improving research agent using constraint-wise verification, a learned context-update tool, and key-step-focused RL reaches frontier-level accuracy with only 10B active parameters.","lead":"AREX trains a 10-billion-parameter AI research agent that alternates between gathering evidence and auditing its own draft answer, compressing its search history into a compact 'improvement state' so it can refine the same problem across multiple rounds. On six deep-research benchmarks it reports scores competitive with much larger frontier agents, with a 4B variant also beating some 35B models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3's outer-loop ablation is confounded: the w/ outer-loop condition gets additional research rounds and the model was trained with that loop, so the 22.9-point gain may not isolate the recursive double-loop architecture.","rationale":"The reader's weakest assumption concerns confidence-score reliability and the unreported threshold in Eq. 10. That is a valid issue, but I see a more fundamental problem: even with a perfectly calibrated confidence score, Table 3's outer-loop ablation would not isolate the architecture. The on/off outer-loop comparison differs in two ways beyond the decision rule: the amount of inference-time compute actually consumed, and the match between training and inference conditions. The ACU ablation (59.6 vs 71.4) is explicitly described as a 'matched single-round' setting, so it is relatively clean. The outer-loop comparison is not described as matched, and the w/ outer-loop condition can use additional recursive rounds under the same episode-level cap. Thus the 11.1-point gain attributed to the outer loop could simply be a test-time-compute effect. Additionally, the model's training trajectories include verification-driven transitions and outer-loop refinement; removing the outer loop at inference puts the model in an off-distribution setting, so the single-round baseline may be unfairly depressed. A matched-budget generic-continuation control would resolve this. The paper's absolute benchmark results and the cleaner ACU ablation remain useful, so the appropriate verdict is still CONDITIONAL; the concern is addressable and does not by itself require rejection.","tokens_in":18253,"tokens_out":7711,"duration_ms":81352,"concrete_test":"Run a matched-budget control on BrowseComp: keep the w/ ACU condition, but replace the outer-loop diagnosis (Eq. 9) with a generic continuation that restarts or continues the inner loop for the same total number of additional turns used by the full system (e.g., issue a fixed 'continue researching' objective). If this control reaches approximately 82.5, the outer-loop gain is from extra compute, not from the targeted self-improvement mechanism. If it falls short, the targeted diagnosis is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central mechanistic claim is that the recursive double-loop architecture drives the 22.9-point BrowseComp gain (Sec 5.3, Table 3). But the comparison is not controlled for two confounds. First, compute: 'w/ outer loop' can run up to 5 recursive rounds, while 'w/o outer loop' accepts the first finish output; the stated 300-turn cap applies per episode, not per condition, so the outer-loop condition can consume more of the budget. The 11.1-point gain (71.4 to 82.5) may only reflect extra research turns, not the targeted diagnosis in Eq. 9-10. Second, training distribution: the model was trained on trajectories that include outer-loop refinement (Sec 3.2, Sec 4.1), so disabling the loop at inference creates a train/inference mismatch; the 'w/o outer loop' row may understate what a single-round model could do. To support the claim that recursive verification-driven refinement, and not simply additional search, is responsible, the authors need a matched-budget control and/or a model trained without the outer loop.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"AREX introduces a bi-level recursive research agent for deep research. The inner loop gathers evidence through search/browsing tools and periodically invokes a learned update_context tool that compresses the interaction history into a structured research state; it terminates with a provisional answer, supporting evidence, and a self-reported confidence score. The outer loop then either accepts the answer (confidence ≥ τ), refines it by preserving verified evidence and issuing a targeted next-round objective, or restarts from the original query. The paper trains dense 4B and 122B-A10B MoE variants using synthetic constraint-based tasks, filtered teacher trajectories, multi-stage agentic mid-training with key-step supervision, and step-aware reinforcement learning. It reports strong results on BrowseComp, WideSearch, DeepSearchQA, GAIA, xbench-2510, and HLE, and attributes the main gains to autonomous context updating (ACU) and the outer self-improvement loop in Table 3.","tokens_in":18543,"tokens_out":7902,"duration_ms":78005,"significance":"If the mechanism is validated, this is a meaningful contribution: it offers a concrete instantiation of recursive verification-driven refinement and shows that a 10B-active MoE can be competitive with much larger research agents on external benchmarks. The paper has real strengths: it evaluates on external benchmarks rather than only on in-house tasks, the ACU ablation in Table 3 is internally framed as a controlled comparison, Appendix B is honestly caveated, and the models are released. However, the central outer-loop ablation is confounded by compute and training-distribution differences, and the confidence-based control loop is underspecified because τ and τ_conf are never reported. Those issues must be addressed before the central mechanistic claim is established.","major_comments":[{"comment":"The 11.1-point gain attributed to the outer loop (71.4→82.5 with ACU) is confounded. The w/ outer-loop condition may use up to 5 outer-loop operations (Sec. 5.1), while w/o outer loop accepts the first finish output; no matched turn/token budget is reported for the two rows. Moreover, the model was trained on trajectories that include outer-loop refinement (Sec. 3.2, 4.1), so the w/o row is a train/inference mismatch and may understate what a single-round model could achieve. The 22.9-point headline in Table 3 therefore cannot be assigned specifically to the recursive double-loop architecture. Please add a matched-budget single-round control and/or train a variant without outer-loop trajectories.","section":"§5.3, Table 3; §3.2, §4.1"},{"comment":"The accept/refine/restart rule is governed by a threshold τ that is never reported, and the teacher-trajectory filter threshold τ_conf in Eq. (17) is also undisclosed. This makes the control loop non-reproducible and leaves open the possibility that gains were selected over a favorable threshold. The calibration evidence is weak: Fig. 3 shows that 33.0% of incorrect outputs with ACU have confidence in the 90–100 bin, so high confidence does not reliably indicate correctness. Please report τ and τ_conf, and provide a threshold sweep or calibration analysis.","section":"§2.3, Eq. (10); Eq. (17); §5.3, Fig. 3"},{"comment":"The abstract's claim that AREX 'substantially outperforms comparable-scale baselines' is not supported by the table: AREX-Base trails Kimi-K2.6 on BrowseComp (82.5 vs 83.2), trails MiroThinker-H1 on GAIA and xbench-2510, and many margins are within a few points. No error bars, significance tests, or multiple runs are reported, so 'substantially' is unjustified. In addition, the HLE block mixes full-HLE and text-only scores (per the footnote), making direct ranking within that block unreliable. Please report uncertainty estimates and separate the HLE protocols clearly.","section":"§5.2, Table 1; Abstract"},{"comment":"The parameter-efficiency claim ('competitive with models using substantially more activated parameters') is not matched for inference-time compute. Deep-research performance scales with test-time budget; AREX is given a 300-turn/5-round ceiling, but no token or turn budgets are reported for the comparison systems. Without matched compute or at least reported budgets, the capability-to-parameter conclusion is not established. Please report average/median turns and tokens for AREX and the baselines, or add compute-matched comparisons.","section":"§5.1, §5.2"}],"minor_comments":[{"comment":"G_θ, the trajectory assessor, is never defined. Clarify whether it is the same policy, a separate model, and how it is trained.","section":"§2.3, Eq. (9)"},{"comment":"The text says '5 outer self-improvement loop operations' but elsewhere implies '5 recursive rounds.' State the exact maximum number of finish calls per episode.","section":"§5.1"},{"comment":"Report the number of outputs in each confidence bin and add a reliability diagram/calibration curve; histograms alone overstate the separation between correct and incorrect outputs.","section":"§5.3, Fig. 3"},{"comment":"The row labeled 'Upper bound 128,000' is ambiguous; label it as the active context-window size.","section":"Table 2"},{"comment":"The teacher models used to collect trajectories are not named. State which models generated the trajectories and how many trajectories were retained.","section":"§3.2"},{"comment":"Define the acceptance criteria for 'accepted generated trajectories' in the self-distillation experiment.","section":"Appendix B"},{"comment":"The rule-based key-step detectors are described only qualitatively; report the actual detector thresholds or at least their precision/recall on a validation set.","section":"§4.1"},{"comment":"The citation 'Team et al. (2026)' is used ambiguously; disambiguate which team/paper is meant.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The central idea is promising and the external benchmark results are interesting, but the main mechanistic claim in Table 3 needs matched-budget and training-distribution controls before it can be accepted. Please also require the missing hyperparameters (τ, τ_conf, and key-step detector thresholds) and error bars on the headline numbers. If the authors can provide those, the paper could be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick read on arXiv:2607.21461.\n\nThe first thing to know: AREX is a genuine systems contribution. A 10B-active MoE that hits 82.5 on BrowseComp and 82.0 on WideSearch, competitive with models using far more parameters, is practically useful. The two-loop design — inner research loop externalizes a structured provisional answer, outer loop audits it constraint-wise and issues refine/restart — is a real combination not present in the cited prior work. The learned context-update tool is also new in this form, and the training recipe (key-step focused supervision, step-aware RL) has internal ablations that look consistent, with matched budgets.\n\nThe second thing to know: the headline mechanistic claim is softer than the abstract implies. Table 3's 22.9-point gain (59.6 to 82.5) is the load-bearing number, but that comparison is confounded. The 'w/ outer loop' condition can run up to 5 recursive rounds while 'w/o outer loop' accepts the first finish; the 300-turn cap is per episode, not per condition, so the gain could partly reflect extra compute. And the model was trained with the outer loop in its trajectory distribution, so disabling it at inference creates a train/inference mismatch — the 'w/o outer loop' row understates what a single-round model could do. The ACU gain in the single-round setting (59.6 to 71.4) is clean and supports the inner-loop claim. But the 22.9-point figure shouldn't be presented as the effect of the recursive double-loop architecture.\n\nOther soft spots are minor but real. The confidence threshold τ in Eq. 10 is never disclosed, and Fig. 3 shows 33% of incorrect outputs with ACU sit in the 90–100 confidence bin, so the accept/refine control is less reliable than the paper's language implies. No error bars or significance tests anywhere. Table 1 shows AREX-Base trailing several models on multiple tasks, so 'substantially outperforms comparable-scale baselines' is overstated. And no code or environment release is mentioned, which limits reproducibility.\n\nWho this is for: anyone building deep-research agents or doing long-horizon tool-use training. The ACU usage analysis (80% of cases, mostly on strategy revision) is genuinely informative. It deserves a serious referee, but the referee should push for a matched-budget ablation of the outer loop and disclosure of τ before the central claim is accepted.","headline":"Strong systems paper with a genuinely useful architecture, but the headline ablation is confounded and the confidence threshold is undisclosed.","tokens_in":19145,"tokens_out":2880,"would_cite":true,"duration_ms":28592,"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 deep-research agent that recursively verifies its provisional answers and uses unresolved constraints to target the next round of search outperforms far larger systems on deep-search benchmarks.","keywords":["recursive self-improvement","deep research agents","discovery-verification asymmetry","constraint-wise verification","autonomous context updating","long-horizon reinforcement learning","mixture-of-experts","key-step supervision"],"falsifier":"Run AREX on BrowseComp with the outer loop's refinement objective replaced by a randomly selected unresolved constraint from the task template (keeping the budget, tools, and training fixed); if accuracy stays near 82.5 rather than dropping toward the no-outer-loop level, the claim that targeted follow-up from unresolved constraints drives the gain would be falsified.","tokens_in":18112,"feed_emoji":"🔁","tokens_out":3427,"duration_ms":34246,"temperature":0.7,"pith_summary":"AREX is a family of recursive self-improving deep-research agents built on a discovery–verification asymmetry: finding an answer that satisfies many constraints is hard, but checking a candidate constraint-wise is comparatively cheap. The paper's core claim is that an agent should convert a provisional answer into a partially verified research state, then launch a new round of targeted research on whatever remains unresolved — rather than simply searching longer. Across BrowseComp, WideSearch, DeepSearchQA, GAIA, xbench-2510, and Humanity's Last Exam, the 122B-total/10B-active AREX-Base substantially outperforms comparable-scale baselines and remains competitive with far larger models. The load-bearing mechanistic result is that combining an inner research loop with an outer self-improvement loop and an autonomous context-update tool raises BrowseComp accuracy from 59.6 to 82.5.","feed_headline":"Recursive verification adds 23 points to deep-research accuracy","feed_subtitle":"A 10B-active-parameter agent with a self-improvement loop beats far larger systems on BrowseComp and WideSearch.","key_machinery":"The central mechanism is the bi-level recursive loop: an inner research policy (search, browse, integrate, finish) and an outer self-improvement policy that audits the structured result constraint-wise and emits accept/refine/restart decisions (Equation 10). The second load-bearing component is the autonomous context-update tool, which the model invokes itself to condense its trajectory into a compact improvement state; the paper shows it is used proactively (80.3% of BrowseComp cases, usually well below the 128K-token limit) and that replacing the full history with this state alone yields 11.8 accuracy points. A third component is step-aware training: key-step detectors identify evidence di","core_discovery":"The central claim is that verification can serve as the transition operator between research rounds, not merely a final filter. AREX alternates an inner loop that gathers evidence and produces a provisional answer with supporting evidence and a confidence score, and an outer loop that adopts a simple decision rule: accept if confidence exceeds a threshold, refine by preserving useful findings and targeting the unresolved constraints, or restart if the trajectory is judged too noisy. Sustaining this recursion over long horizons requires a learned context-update tool that compresses the interaction history into an improvement state preserving verified findings, unresolved constraints, rejected","pith_inferences":["The paper leaves implicit that the confidence score's reliability is the fulcrum for the entire accept/refine/restart rule; the authors' own Figure 3 shows 33% of incorrect outputs with the context-update tool landing in the 90–100 confidence bin, so a robust deployment would need calibrated confidence or a richer stopping signal than a scalar threshold.","A natural extension the authors do not explore is using the update_context tool as a domain-agnostic pretrained module, decoupling state maintenance from the research policy; a testable prediction is that such a module would improve transfer to new long-horizon tasks.","Because the outer loop's refinement objective is derived from unresolved constraints, one can test whether the gain comes from the targeting itself or merely from having a second research round: replacing the targeted objective with a random unsolved constraint should, under the paper's claim, degrade accuracy.","The paper's framing suggests a broader recipe: any verification-aware agent that preserves 'what is established' and tracks 'what remains to check' can turn a limited inference budget into compounding progress, which may matter more as models become cheaper per token."],"forward_implications":["If the recursive verification loop is the real driver, then deep-research agents should treat verification as an active control signal that decides when to continue, what to look for next, and when to stop — not just as a scoring step at the end.","Learned, self-invoked context management organized around the research state (verified findings, unresolved constraints, rejected paths) can outperform fixed truncation or external summarization for long-horizon tasks.","Concentrating supervision and RL credit on a small number of decision-critical steps (evidence discovery, redirection, context updates) is more effective than uniform token-level training when budgets are matched.","A 10B-active-parameter agent using this architecture can be competitive with models using substantially more activated parameters on several deep-research benchmarks.","The same double-loop structure could generalize to any task where a candidate solution can be decomposed into independently checkable constraints, including code and mathematical verification."],"fun_headline_variants":["Verification-driven loop lifts deep-research accuracy","Self-auditing agent beats larger models on research tasks","Recursive verification powers small research agent","Verify, refine, repeat: how AREX wins deep research","Inner loop gathers, outer loop audits, accuracy soars"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the model's self-reported confidence score (produced by the finish tool) genuinely reflects how likely the provisional answer is to be correct — yet the paper's own Figure 3 shows 33% of incorrect outputs with ACU falling in the 90–100 confidence band.","fun_headline_variants_meta":{"raw":{"variants":["Verification-driven loop lifts deep-research accuracy","Self-auditing agent beats larger models on research tasks","Recursive verification powers small research agent","Verify, refine, repeat: how AREX wins deep research","Inner loop gathers, outer loop audits, accuracy soars"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000143,"raw_usage":{"total_tokens":1028,"prompt_tokens":785,"completion_tokens":243,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":181}},"tokens_in":529,"tokens_out":243,"duration_ms":2962,"temperature":1.0,"reasoning_tokens":181,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T07:21:03.254274+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run AREX on BrowseComp with the outer loop's refinement objective replaced by a randomly selected unresolved constraint from the task template (keeping the budget, tools, and training fixed); if accuracy stays near 82.5 rather than dropping toward the no-outer-loop level, the claim that targeted follow-up from unresolved constraints drives the gain would be falsified.","supporting_citations":[],"review_version":1}