Artificial Intelligence · arXiv:2609.26532 v1 · submitted 22 Sep 2026 · 12 references
REFLEX with Jev for Efficient Selective Control in LLM Agents
The paper says
“LLM agents often use generative models for bounded decisions, raising the question of when these decisions can be handled more efficiently without reducing task success.”
more of the abstract
We study REFLEX, an agent architecture that uses Jev as a fast, typed decision layer and calls a strong LLM when confidence is low, or generation is required. On a frozen 100-task benchmark, REFLEX achieves 95% success with 72.7% fewer strong-model calls than a strong-only agent, with reductions persisting across three fallback families. Controlled interventions show that reliability depends on action-set size and near-valid alternatives near authorization boundaries. External BFCL and τ-style evaluations reveal limited advantages over a cheap generative cascade when ordinary routing is already highly accurate. These findings identify when selective control with Jev can reduce computation and where its benefits are limited.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 24 Sep 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. The efficiency half of the headline is real; the success half is one roll of a temperature-zero die, and the authors know it — send to review, but on condition the artifacts actually ship and the abstract gets reconciled with the body. This line is wrong
3 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 12 references say.
The paper says, at Abstract (b2/b20), §5.1, Table 1, Table S8, Table S13
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.77):
“Two runs do not provide a general estimate of run-to-run variance or justify a fixed uncertainty band.”
Pith says: Objection 1 of 3 · major
The headline operating result (95% success with GMR 0.727) rests on a single temperature-zero execution per condition. The paired success comparison is significant only by a knife-edge: exact McNemar p=0.0391 on exactly 9 discordant pairs (8 REFLEX wins, 1 loss, Table S8/b130).
more of objection 1
Flipping a single REFLEX win raises the two-sided p to 0.070. The paper's own Table S13 documents B0 flipping 15/100 tasks between the M1 and E6 executions (0.88 vs 0.91) and states (§6, b208, b480) that this run-to-run band is as large as the E6 effects. Because the 8 paired wins are drawn from B0's 12-task failure set, which is itself unstable across runs, the observed 95% vs 88% gain is plausibly concentrated on noise-driven B0 failures. The abstract's '95% success with 72.7% fewer strong-model calls' therefore pairs a noise-level success estimate with a robust efficiency gain as if both had the same evidentiary status. The defensible central claim is 'comparable success (within the documented run-to-run band) at roughly one-third of the strong-model calls.' The efficiency half survives; the success-gain framing must be corrected in the abstract and headline.
The paper says, at Abstract (b21), §5.5, Table S12
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.85):
“It was not a formal non-inferiority test.”
Pith says: Objection 2 of 3 · major
The abstract states that across three fallback families REFLEX 'keeps success within a 2-point non-inferiority margin.' This is stated without qualification, but the body is careful to distinguish the preregistered point-estimate rule from a formal non-inferiority test (b198: 'It was not a formal non-inferiority test').
more of objection 2
Under the confidence-bound criterion (lower 95% bound above -0.02), only Qwen and DeepSeek pass; Kimi's lower bound is -0.050 and fails (Table S12). Per-condition success intervals are absent from the frozen output (b492). A strict non-inferiority reading of the abstract therefore contradicts the body. Recommend either qualifying the abstract margin claim or reporting the per-condition intervals so the reader can see which families meet which criterion.
The paper says, at §5.1, Table 1, b113
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The operating point τ=0.5 is the best point on a threshold sweep over the same frozen benchmark, and the '95% success' headline is that selected point. Disclosure of the full sweep (Table 1) partially mitigates the selection concern, but the headline simultaneously (a) selects the best τ on the evaluation data and (b) reports a point estimate within the documented run-to-run band (Table S13).
more of objection 3
Similarly, the claim that R1 'matches or exceeds B0's success rate at every threshold' (b113) is asserted from within-run point estimates; given B0's ±3-point run-to-run band, the at-every-threshold ordering cannot be asserted against run-to-run variation. Recommend presenting the τ sweep with success bands reflecting the documented variability, and leading the abstract with the efficiency result rather than with the selected success point.
5 minor objections
The paper says, at Appendix B.2, Table S2 note
Pith says: minor objection 1 of 5
The table note contains a typo: 'Itsp is not reported' should read 'Its p is not reported.'
The paper says, at Abstract, §5.4
Pith says: minor objection 2 of 5
The notation for the external evaluation is inconsistent: the abstract writes 'τ-style' while §5.4 and Figure 5(b) use 'τ 2-bench.' Define the benchmark name once and use it consistently.
show 3 more
The paper says, at Table S11, §5.4
Pith says: minor objection 3 of 5
Zero control-selection and zero argument-generation errors across all 300 episodes is a striking figure that anchors the negative τ2-bench result. A sentence on the granularity and coverage of the audit categories would help the reader calibrate how much confidence to place in that zero, since it is the basis for concluding the benchmark cannot show routing value.
The paper says, at Appendix B.1 / Table S1 note (b492)
Pith says: minor objection 4 of 5
Per-condition success intervals are absent and the paper says they are deliberately left unspecified. Since the headline success figure is a single-run point estimate, recommend adding an explicit sentence in the main text (not only in the supplement) stating that the 95% figure is a point estimate subject to the documented run-to-run band of Table S13.
The paper says, at §5.1 (b122)
Pith says: minor objection 5 of 5
The phrasing 'approximately 7.6× lower using the displayed values' is unnecessarily hedged; the AURC comparison (0.019 vs 0.145) can be reported plainly as a ratio computed from the reported values.
What would settle it
Run the frozen REFLEX-SIMv1.2.0 100-task test set under B0 (Max-only) and R1 (tau=0.5) for K independent executions each (K >= 5), with identical prompts, model versions, and temperature-zero settings. For each execution, record per-condition success and the paired discordant table (REFLEX-win, REFLEX-loss, both-succeed, both-fail). Report (1) per-condition success intervals across the K runs and (2) the distribution across executions of the paired gain (REFLEX wins minus REFLEX losses) and of the exact McNemar p. If the median paired gain falls below about 3 tasks, or if in most executions the McNemar p exceeds 0.05, then the no-loss-with-gain reading of the headline is not established at the claimed precision, and the central claim should be restated as 'comparable success with a 66–72% reduction in strong-model calls,' anchored by the robust GMR measured across families.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 24 Sep 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Main thing to know: this is the rare agent-cost paper that does its homework. The compute-reduction result is robust — 72.7% fewer strong calls in the main run, 71.9% in E6, 66–71% across three fallback families — and it survives the stress test. The success side does not. The 95%-vs-88% headline rests on a single temperature-zero execution: nine discordant pairs, McNemar p=0.039, and one flipped REFLEX win takes p to 0.070. Table S13 documents 15 of 100 tasks flipping between two identical B0 runs, and the authors state flat out that the run-to-run band is as large as the E6 effects. The honest claim, which the body mostly supports, is comparable success at roughly a third of the strong-model calls. That is still a real contribution.
What's actually new: applying a typed, confidence-gated decision model to individual steps inside a trajectory, not per-query routing. The RF-5C factorial and the matched read-vs-write competitor intervention are genuinely good — shifting the error type from irreversible commit to deferral while holding overall accuracy fixed, with both shift intervals excluding zero, is a non-obvious causal result. And the negative external results are reported straight: on tau-2-bench a cheap cascade is competitive, and on BFCL the hard decision is whether-to-act (52%) rather than which-tool (98.4%).
Soft spots, in proportion. The benchmark is private, self-authored, self-frozen, with tau=0.5 selected from a sweep on the same data. Nothing is released, so the self-certification burden is real. The abstract's "2-point non-inferiority margin" overstates what the body concedes: the formal confidence-bound test fails for Kimi (Table S12, sentence b198). The authors are unusually honest — they retire four of twelve claims including superiority over the cheap cascade, and admit in the appendix there are no per-condition success intervals (b492) and that two runs give no variance estimate (b480). But the appendix carries the honesty while the abstract carries the stronger reading. That mismatch matters for how the result gets cited.
Bottom line: the efficiency result and the intervention deserve a serious referee. The fragility around the success claim is real but correctable — repeating the execution and releasing the frozen artifacts would settle it. Conditional accept, not reject. Anyone doing cost-aware agent design or agent benchmark methodology gets value here, and the claim register plus audit history are worth studying as a template on their own.
The strongest objection
The central claim has two independent parts: the strong-call reduction and the success comparison. The reduction is robust — GMR is 0.727 in M1, 0.719 in E6, and 0.661–0.714 across the three fallback families — so that half of the headline survives scrutiny. The fragile half is the success side, where REFLEX is claimed to reach 95% against B0's 88% with no loss and an observed gain. The gain rests entirely on a single temperature-zero execution. Its significance is McNemar p=0.0391 computed on exactly 9 discordant pairs (8 REFLEX wins, 1 REFLEX loss). Recomputing the exact two-sided test: if a single REFLEX win flips to a concordant or loss pair, the outcome becomes (7,1), n=8, whose p is 2×(1+8)/256 = 0.070, above 0.05; a single flip destroys the headline significance. The paper itself documents that B0 flips 15/100 tasks between the M1 and E6 executions (0.88 vs 0.91) and REFLEX flips 8 tasks, and states that this run-to-run band is as large as the E6 effects themselves (Table S13). Because the 8 paired wins are drawn from B0's 12-task failure set, and that failure set is largely unstable across runs, the wins are plausibly concentrated on noise-driven B0 failures. If so, the defensible claim is 'comparable success at roughly one third of the strong calls,' which preserves the efficiency contribution, but the headline '95% success with 72.7% fewer strong-model calls' (abstract sentences b2, b20) overstates by implying a simultaneous success gain. The paper is unusually honest here: the authors concede (b480) that two runs give no variance estimate and (b492) that per-condition success intervals are absent from the frozen output. Those concessions are in the appendix, while the abstract carries the stronger reading. This is a correctness-risk concern about the central empirical claim, not an internal inconsistency, and it is directly testable by repeated execution.
The reader's weakest_assumption is exactly the single-execution stability of the temperature-zero point estimate, with Table S13's 15-flip B0 band cited as comparable to the gains. My analysis agrees and sharpens the concern quantitatively: the entire statistical significance of the success gain sits on 9 discordant pairs, and flipping just one of them — well inside the documented 15-task provider-noise band — pushes the McNemar p from 0.039 to 0.070 and removes significance. This makes the fragility concrete rather than impressionistic. I do not find a separate load-bearing concern the reader missed: the cross-family 2-point non-inferiority overstatement for Kimi is acknowledged in the body (b198, Table S12), and the RF-5C and tau-bench results are honestly bounded with confidence intervals spanning zero. The verdict should remain CONDITIONAL rather than moving to REJECT, because the efficiency claim is robust (GMR 0.66–0.73 across families and both executions), the paper is unusually transparent with its claim register and audit history, and the defect is correctable by releasing multi-run data. It should also not move to ACCEPT because the headline success gain is not established at the claimed precision until repeated-execution evidence is supplied. The paper deserves credit for the honesty of its limitations section, but that honesty relocates the burden of proof to the abstract, which still carries the stronger claim.
The referee's report
If the efficiency result holds, the contribution is genuinely useful: robust evidence that a typed decision layer can remove roughly two-thirds of strong-model calls from agent trajectories on workloads dominated by bounded decisions, with unusually honest methodology — a frozen v1.2.0 benchmark with an audit history, preregistered contrasts, paired cluster bootstrap over scenario families, a disclosed threshold sweep, and a claim register that retires unsupported claims. The controlled interventions on decision structure (action-set size, procedural ambiguity, competitor shape) are clean and generalize beyond the specific architecture. However, the headline success claim (95% vs 88%) is likely within the run-to-run noise band that the authors themselves document (Table S13), so the significance of the paper rests on the efficiency result and on the honest negative external results, not on the success gain as advertised in the abstract. The abstract's unqualified '95% success with 72.7% fewer strong-model calls' overstates what is established; the defensible claim is comparable success within noise at roughly one-third of the strong calls.
Pith's reading of the paper (a machine's summary, not the authors')
This paper asks whether the many small bounded decisions inside an LLM agent's trajectory—which tool to call, whether to continue, whether to ask for clarification—can be handled by a fast typed decision layer instead of a strong generative model. It proposes REFLEX, which uses a decision model named Jev to return a bounded action and a confidence score, and escalates to a strong LLM only when confidence is low or free-form generation is needed. On a frozen 100-task benchmark, REFLEX reaches 95% task success while cutting strong-model calls by 72.7% relative to a strong-only agent, and the 66–72% call reductions persist across three fallback families. The same experiments show the limits: when ordinary tool selection is already 98.4% accurate and multi-turn failures contain no routing errors, a cheap generative cascade stays competitive. The paper's conclusion is that selective control can reduce computation, but its value depends on whether the workload contains control decisions that actually need to be made.
What carries the argument. The load-bearing mechanism is the confidence gate: a typed decision model Jev returns $(\hat{a}_t,c_t)=J(s_t,A_t)$, and the policy executes $\hat{a}_t$ when $c_t\ge\tau$ and the action is executable, otherwise the strong LLM handles the step. This is what separates bounded control from open-ended reasoning inside a trajectory, and the reported gain is measured as $\mathrm{GMR}=1-N^{\mathrm{REFLEX}}_{\mathrm{strong}}/N^{\mathrm{strong\text{-}only}}_{\mathrm{strong}}$. The reliability analysis is carried by a procedural distance $d(a\mid s)$ that counts the state or policy facts that would have to change for an action to become correct; it defines the A0–A3 ambiguity levels without relying on embedding similarity, which the paper shows does not order procedural validity. These two objects—the gate and the distance ladder—supply the mechanism and the controlled test of when that mechanism fails.
Core claim. REFLEX formalizes agent control as a selective decision problem rather than a generation problem. At each step a typed "System One" model Jev returns a bounded action and confidence; the policy executes that action when confidence passes a threshold and the action is executable, and otherwise passes the state to a strong LLM. On the frozen 100-task REFLEX-Sim benchmark at threshold 0.5, the REFLEX agent succeeds on 95 tasks versus 88 for the strong-only agent while using 1.12 strong-model calls per task instead of 4.10, a 72.7% reduction; the paired outcome table shows it recovers eight tasks the strong-only agent fails and loses one. With Jev and the threshold fixed, strong-call reductions across Qwen3.8-Max, Kimi K3, and DeepSeek-V4-Pro fallbacks are 66% to 72%, the same 66-task autonomous subset succeeds in all three families, and no success decrease is statistically significant, although a formal confidence-bound non-inferiority check passes for Qwen and DeepSeek but not Kimi. External evaluation narrows the claim: on BFCL function selection is 98.4% accurate while deciding whether any function should be called is only 52.0% accurate, and on τ-style multi-turn tasks REFLEX is 3.7× cheaper than a strong-only agent but a cheap Flash cascade with self-escalation shows numerically higher success. The paper's central claim is that separating bounded control from generation can reduce computation, but only when the task's difficulty actually lies in control decisions.
The premise it rests on. The load-bearing premise is that a single temperature-zero execution on the frozen 100-task benchmark is a stable measurement of task success, even though the paper itself reports 15 of 100 tasks changing outcome between two identical strong-only runs.
If the paper is right (Pith's inferences, not the authors')
On agent workloads dominated by bounded, repeated decisions, deployment can remove roughly two-thirds of strong-model calls; on REFLEX-Sim the strong calls per successful task fall from 4.66 to 1.18.
The savings do not depend on imitating the fallback: the same frozen gate selects the same 66-task autonomous subset, all 66 succeed under Qwen, Kimi, and DeepSeek, and every observed failure occurs in the escalated 34-task subset.
The gate is selective but not proven perfect: at the primary threshold 250 of 251 autonomous decisions are valid, the single invalid one is recoverable, and no unsafe autonomous decision is observed at any threshold.
On benchmarks where routing is not the bottleneck, the layer can cut cost but shows no demonstrated quality advantage; τ2-bench records zero audited control-selection and argument-generation errors across all 300 episodes, so a cheap cascade remains competitive.
Because BFCL relevance accuracy is 52% while incorrect calls carry mean confidence 0.778, the fix is not just a higher threshold but richer features for authorization and evidence; a testable extension is to feed those features into the typed decision layer and re-measure the 52% figure.
The read-versus-write intervention implies that deployments should optimize an expected-consequence objective: two configurations with statistically unresolved accuracy differences can differ by 8.3 percentage points in irreversible-commit errors, so the gate's operating point should be chosen by the cost of the errors it can cause.
The same 66-task autonomous subset across three unrelated fallback families suggests a learnable class of already-decidable states; one could identify that subset offline and run those steps without even invoking the gate.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
LLM agents often use generative models for bounded decisions, raising the question of when these decisions can be handled more efficiently without reducing task success. We study REFLEX, an agent architecture that uses Jev as a fast, typed decision layer and calls a strong LLM when confidence is low, or generation is required. On a frozen 100-task benchmark, REFLEX achieves 95% success with 72.7% fewer strong-model calls than a strong-only agent, with reductions persisting across three fallback families. Controlled interventions show that reliability depends on action-set size and near-valid alternatives near authorization boundaries. External BFCL and τ-style evaluations reveal limited advantages over a cheap generative cascade when ordinary routing is already highly accurate. These findings identify when selective control with Jev can reduce computation and where its benefits are limited.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-24
The authors' main claim, as the classifier picked it out 1.00
“These findings identify when selective control with the classifier can reduce computation and where its benefits are limited.”
How the main claim is stated: asserted 0.99
What the abstract itself states, as probabilities: claims first 0.03 · states limitation 0.96 · human participants 0.05 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.94 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| LLM agents often use generative models for bounded decisions, raising the question of when these decisions can be handled more efficiently without reducing task success. | background 0.93 | 0.06 | 0.03 | asserted 0.90 |
| We study REFLEX, an agent architecture that uses Jev as a fast, typed decision layer and calls a strong LLM when confidence is low, or generation is required. | method 0.62 | 0.75 | 0.09 | asserted 0.98 |
| On a frozen 100-task benchmark, REFLEX achieves 95% success with 72.7% fewer strong-model calls than a strong-only agent, with reductions persisting across three fallback families. | result 1.00 | 0.96 | 0.41 | asserted 1.00 |
| Controlled interventions show that reliability depends on action-set size and near-valid alternatives near authorization boundaries. | result 0.88 | 0.88 | 0.07 | asserted 1.00 |
| External BFCL and $\tau$-style evaluations reveal limited advantages over a cheap generative cascade when ordinary routing is already highly accurate. | result 0.60 | 0.86 | 0.13 | asserted 0.99 |
| These findings identify when selective control with Jev can reduce computation and where its benefits are limited. | interpretation 0.98 | 0.90 | 0.83 | asserted 0.99 |
What kind of paper the classifier reports this is: method or tool 0.85 · evidence from observation or experiment 0.93
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-24: 1 request, 4,899 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (6)
Figures are the authors', reproduced from the paper.
References (12)
- [1] Almeida, Diogo , title =. 2026 , month = sep, howpublished = on Pith
- [2] Lingjiao Chen and Matei Zaharia and James Zou , journal=. Frugal. 2024 , url= on Pith
- [3] Isaac Ong and Amjad Almahairi and Vincent Wu and Wei-Lin Chiang and Tianhao Wu and Joseph E. Gonzalez and M Waleed Kadous and Ion Stoica , booktitle=. Route. 2025 , url= on Pith
- [4] Qitian Jason Hu and Jacob Bieker and Xiuyu Li and Nan Jiang and Benjamin Keigwin and Gaurav Ranganath and Kurt Keutzer and Shriyash Kaustubh Upadhyay , booktitle=. RouterBench: A Benchmark for Multi-. 2024 , url= on Pith
- [5] LLMRouterBench: A Massive Benchmark and Unified Framework for LLM Routing , author=. 2026 , eprint= on Pith
- [6] Shishir G Patil and Huanzhi Mao and Fanjia Yan and Charlie Cheng-Jie Ji and Vishnu Suresh and Ion Stoica and Joseph E. Gonzalez , booktitle=. The Berkeley Function Calling Leaderboard (. 2025 , url= on Pith
- [7] Wang, Hongru and Huang, Wenyu and Wang, Yufei and Xi, Yuanhao and Lu, Jianqiao and Zhang, Huan and Hu, Nan and Liu, Zeming and Pan, Jeff Z. and Wong, Kam-Fai. Rethinking Stateful Tool Use in Multi-Turn Dialogues: Benchmarks and Challenges. Findings of the Association for Compu... on Pithdoi:10.18653/v1/2025.findings-acl.284
- [8] Shunyu Yao and Noah Shinn and Pedram Razavi and Karthik R Narasimhan , booktitle=. 2025 , url= on Pith
- [9] 2 -Bench: Evaluating Conversational Agents in a Dual-Control Environment , author=. Forty-third International Conference on Machine Learning , year= on Pith
- [10] Subramani, Nishant and Eisner, Jason and Svegliato, Justin and Van Durme, Benjamin and Su, Yu and Thomson, Sam. MICE for CAT s: Model-Internal Confidence Estimation for Calibrating Agents with Tools. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of ... on Pithdoi:10.18653/v1/2025.naacl-long.615
- [11] Xuan, Weihao and Zeng, Qingcheng and Qi, Heli and Xiao, Yunze and Wang, Junjue and Yokoya, Naoto. The Confidence Dichotomy: Analyzing and Mitigating Miscalibration in Tool-Use Agents. Proceedings of the 64th Annual Meeting of the A ssociation for C omputational L inguistics (V... on Pithdoi:10.18653/v1/2026.acl-long.520
- [12] Causal evidence that language models use confidence to drive behaviour , author=. Nature Machine Intelligence , pages=. 2026 , publisher= on Pith
Cited by 21 Pith papers
Reviewed papers on Pith that cite this one. OpenAlex reports about 24 citations worldwide. The full citation record
- Operon-S1: a fast decision controller for long-horizon biomedical agents cs.AI · Oct 2026
- Thinking Fast, Anywhere: One System-1 for the Decision Points of Any Agent Workflow cs.AI · Nov 2026
- Browser Immune System (BIS): Innate and Adaptive Threat Detection in the Browser with a Cybersecurity Transformer cs.CR · Nov 2026
- NIOCIS: Typed-Decision Next-Hop Routing with Node-Conditioned Option Spaces for Tool-Use Agents cs.CR · Oct 2026
- Suggested Is Not Superseded: A Pre-Registered Evaluation of a System-One Model as a Memory-Retirement Gate cs.AI · Oct 2026
- Jev with Foresight: Safe Language-Guided Autonomous Driving via a Language World Model cs.RO · Nov 2026
- FlipJev: Fast Black-Box Red Teaming Finds Universal Adversarial Suffixes for Typed Decision Models cs.CR · Oct 2026
- CASEL: Consequence-Aware Selective Execution for Reliable Long-Horizon Tool-Using Agents cs.AI · Oct 2026
- Decision-Native Recommendation with Large Language Models cs.IR · Oct 2026
- Contestable Agents for Data Governance: Recording Human Stewardship of AI-Generated Metadata Across Structured and Unstructured Systems cs.AI · Oct 2026
- Truncated Reasoning Under a Budget: Evaluating System-1 Models as Decision-Makers cs.CL · Oct 2026
- Jev-LiteMAS: Separating Decisions from Generation in Multi-Agent Workflows cs.AI · Oct 2026
- CT0: Cybersecurity Transformer cs.CR · Oct 2026
- OmniMed-Jev: Calibrating LVLM Confidence for Trustworthy Medical Multimodal Decisions via System One cs.LG · Sep 2026
- VisualJevBench: A Benchmark and Black-Box Attacks for Visual Typed Decision Models cs.CR · Nov 2026
- DisasterRoute: Evidence-Guided Dual-System Routing for Visual Disaster Question Answering cs.CV · Oct 2026
- JevCascade: Readout Stability and Inference-Time Compute Allocation for Prefill-Only Decision Models cs.CL · Oct 2026
- AnyJev Technical Report cs.LG · Sep 2026
- Calibrated Decision Models for Autonomous Penetration-Testing Harnesses: JEV and Laya as System One Decision Layers for LLM-Driven Pentest Agents cs.CR · Sep 2026
- Component Replacement in LLM Agents: A Critical Scoping Review of Task Benefit and Attribution cs.AI · Sep 2026
- Typed Decision Models: An Early Evidence Audit and Evaluation Checklist cs.CL · Sep 2026
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment