#01Jul 30, 2026
cs.CL
Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost, from 1.5B to 7B
Iliya Mirzaei
Methods that make a language model plan, criticise and rewrite its own answer, reflect on mistakes, pick the best of several attempts, or debate with copies of itself nearly all make it generate far more text than a single chain of thought. Because generating more text raises accuracy by itself, a gain over one chain of thought does not show the method's idea is what helped. Wang et al. (2024) reported that a simple baseline, sampling the same question repeatedly and keeping the most common answer, often wins once budgets are comparable, but gave point estimates with no confidence intervals or significance tests. We rerun that comparison as a designed experiment: seven methods, open models of 1.5B, 3B and 7B parameters, two mathematics benchmarks, 150 questions each. We count every generated token, including those spent on critiques, reflections, debate turns and checking, and compare each method against repeated sampling at its own measured cost. All 36 comparisons are paired by question, with bootstrap intervals and multiplicity correction. No method is reliably better than repeated sampling at equal cost anywhere. Ten are reliably worse, all of them methods where the model inspects its own output, and all 18 self-inspection comparisons are negative. The two kinds of self-inspection part company as models grow. Choosing stops hurting: taking Best-of-N's eight samples and just counting the most common answer beats letting the model pick by 8.0 and 11.3 points at 1.5B, but only 2.0 and 1.3 at 7B, no longer distinguishable from zero. Rewriting does not recover: Self-Refine and a forced Reflexion stay 3.6 to 10.1 points below baseline at 7B. Reflexion as published never triggered its own retry on the smallest model. It judged itself correct every time and silently became a single chain of thought. We release code, prompts, all generations, and our verification scripts.
#02Jul 30, 2026
cs.CV
ScaFE: Data-Efficient Scar Classification with LLM-Generated Clinical Feature Programs
Ruman Wang, Hangting Ye
Classifying pathological scars from clinical photographs requires distinguishing keloids from hypertrophic scars despite limited expert-labeled data and substantial acquisition variation across hospitals. End-to-end image models remain data-dependent, whereas sending photographs to a hosted vision-language model (VLM) may conflict with local data-governance requirements and yields decisions that are difficult to reproduce and audit. We introduce ScaFE (Scar Feature Engineering), which transfers clinical knowledge from a large language model (LLM) into deterministic, executable feature programs instead of asking the model to diagnose images. A web-enabled LLM retrieves clinical evidence and synthesizes programs that measure visually assessable scar attributes. Candidate programs execute in a restricted local environment, and only aggregate validation statistics and feature-level SHAP summaries are returned for iterative repair and refinement; raw images and patient-level outputs remain local. A lightweight Random Forest then operates on the resulting structured representation. On 600 photographs from three hospitals under leave-one-site-out evaluation, ScaFE achieves 81.0% site-macro balanced accuracy, exceeding the strongest baseline, BiomedCLIP, by 10.0 percentage points. With only 10% of the development data, ScaFE retains 72.0% balanced accuracy and an 11.8-point lead. Iterative refinement also raises the executable-program rate from 66.7% to 95.0%, with verified evidence for 91.7% of the final features. These results show that LLM knowledge can support data-efficient, cross-site medical image classification through local and auditable feature programs rather than direct VLM decisions.
#03Jul 30, 2026
cs.GT
Agents That Certify Their Own Exploits: Confidence-Scheduled Restricted Responses for Safe Opponent Exploitation
Boning Li, Longbo Huang
An agent playing a Nash-equilibrium strategy in a two-player zero-sum imperfect-information game secures the game value but forfeits the additional value offered by a flawed opponent. Diffuse deviations pose a particular challenge: binary release rules may gather too little evidence to act, while a full best response to an incomplete opponent model can be highly exploitable. We introduce \emph{budget-constrained confidence-scheduled restricted responses} (CS-RNR), the first opponent-exploitation method whose safety guarantee is a certificate the agent computes on the strategy it actually deploys, so that every exploit it commits to is one it has audited itself. The method tracks pooled action frequencies with anytime-valid confidence sequences and treats a frequency as exploitable only once its interval separates from an equilibrium reference. The confirmed deviations define a conservative opponent model, which a restricted-response solve turns into candidate counter-strategies over a grid of pin levels. Before deployment, each complete candidate is evaluated by a full-tree best response. The resulting certificate is compared with a user-specified budget and committed atomically with the strategy. Because this check is performed on the played strategy, model quality determines the exploitation achieved while the certificate controls reference-relative expected loss. In Leduc hold'em, CS-RNR obtains $6.2\times$ the steady-state gain of a money-verified binary gate while keeping every deployed strategy within budget. A trajectory mixture using the same estimator reaches $13.6\times$ the budget. Across Leduc, Liar's Dice, and 5-rank Leduc, all $36{,}000$ audited hands satisfy the reported certificate tolerance.
#04Jul 30, 2026
cs.LG
Same Graph Cross-Task Transfer in GNNs: Protocols and Predictors
Neelam Akula, Surbhi Kumar, Murat Kantarcioglu and 1 more
Many real-world graphs support multiple predictive tasks over the same underlying structure, creating an opportunity to reuse supervision across node classification (NC) and link prediction (LP). However, existing evaluations often rely on incompatible splits, observed-graph assumptions, and negative sampling rules, making conclusions about same-graph cross-task transfer unreliable. We formalize same-graph NC-LP transfer and propose a leakage-free protocol that fixes node and edge splits, uses a shared message-passing graph that excludes evaluated edges, and employs fixed negatives for LP. Across three backbones (GCN, GraphSAGE, GPS), we find that transfer is strongly directional and predictable: NC $\to$ LP is consistently beneficial on homophilic graphs, while LP $\to$ NC is fragile and can even degrade accuracy under naive representation reuse. LP $\to$ NC becomes reliably positive mainly in a structure-dominant regime where LP is easy but NC is unsaturated, suggesting that LP acts as structural pretraining. Finally, we introduce the CoTask Score (CTS) to summarize joint NC+LP utility when a shared encoder must serve both tasks, and show that simple dataset statistics, especially homophily, can guide mechanism choice and help avoid negative transfer.
#05Jul 30, 2026
cs.CL
ORCA-bench: How Ready Are Language Model Agents for Oncall?
Albert Gong, Kyuseong Choi, Abhineet Agarwal and 5 more
Large language models can write, patch, and search code, but oncall root cause analysis (RCA) demands something different: reasoning over noisy metrics, logs, traces, and source code, starting from ambiguous user-facing reports, often hours after the incident began. We introduce ORCA-bench, a benchmark that puts general-purpose coding agents in a production-fidelity oncall setting. ORCA-bench pairs a live OpenTelemetry-instrumented microservice system--exposing six days of metrics, logs, and traces through real telemetry interfaces (Prometheus, Jaeger, and OpenSearch via Grafana) and full source-code access--with 1,079 RCA tasks that systematically vary report specificity, time-to-detection, and co-occurring fault scenarios. Ground-truth symptoms are curated and signed off by expert SREs, and our LLM-as-judge is independently re-scored by humans (Cohen's $κ_w=0.90$). Across five frontier agents, the best RCA Accuracy is 25.3% on Medium-difficulty tasks (the realistic-input setting) and 10.0% on Hard--a gap that remains even with Claude Fable 5. The weakest model hallucinates an implausible root cause in 40% of incident reports, and removing source-code access degrades every metric. Crucially, these are performances on a curated 50 GB / six-day testbed with tasks investigated in isolation on a system whose code and instrumentation are public. Since real production systems are order of magnitudes larger, more dynamic, and more idiosyncratic, the gap we report is a lower bound on the engineering investment required before frontier coding agents can be safely entrusted with production reliability. We release the public set at https://hub.harborframework.com/datasets/orca-bench/ORCA-bench.