Autoregressive sequence models can suffer hidden future-token leakage when auxiliary operations—such as pooling, routing, or normalization—broadcast full-sequence summaries outside the causal attention mask. This paper audits these violations through prefix-invariance checks, a LeakScore metric, and layer-wise FirstLeakLayer localization, demonstrating that such leakage can produce false architectural gains: a CDR-AttnRes checkpoint drops teacher-forced validation loss from roughly 3.6 to 1.43 while its free-running generation perplexity explodes from under 20 to about 700. The core lesson is that teacher-forced validation loss must be paired with whole-model causal audits and open-ended generation diagnostics before claiming progress.

This paper adapts the Feature Selection Layer (FSL) into a post-hoc interpretability tool for pre-trained neural networks on tabular data. Instead of training an embedded feature weighting layer jointly with the network, the authors freeze the pre-trained model and insert a lightweight diagonal scaling layer that learns per-feature weights via fine-tuning. The approach aims to let practitioners retrofit already deployed models to obtain feature importance scores without retraining from scratch, bridging embedded feature selection and post-hoc attribution methods.

Feature flags become technical debt once fully rolled out, yet manually cleaning them at scale is prohibitively expensive. This paper describes a multi-agent LLM system deployed at DoorDash that automates stale flag removal by querying live rollout data via MCP, coordinating parallel cleanup agents in isolated git worktrees, and producing merge-ready pull requests. Evaluated on 50 stale flags across hundreds of repositories, the system produced usable PRs for 90% of cases at an average cost of $4.79 and 13.8 minutes per cleanup, offering one of the first quantitative industrial benchmarks for LLM-driven feature flag cleanup.

GUI agents struggle in long-horizon mobile tasks because they must repeatedly re-discover intermediate workflows from scratch. This paper proposes a subgoal-aware memory graph that segments past trajectories into contiguous action sequences sharing a single intent, then merges visually similar boundary screenshots into shared graph nodes and attaches annotated subgoal transitions as directed edges. The core claim is that this intermediate granularity is more reusable than either raw actions or full trajectories, which the authors test on a challenging mobile benchmark.

CausalCoT addresses the problem of unfaithful chain-of-thought reasoning in LLMs, where dense attention allows models to bypass their own generated rationales and exploit global context instead. It enforces procedural causal dependence by decomposing queries into a directed acyclic graph of sub-questions and evaluating each node in an isolated context containing only its direct parents. The framework promises more faithful explanations and precise test-time steerability via local interventions.

Tree-based speculative decoding wastes target-model compute when large retrieved candidate trees contain many low-quality branches. ValuSpec addresses this by introducing a plug-and-play verifier-side filter that scores branches with a lightweight model, calibrates a threshold $\tau_{\mathrm{tr}}$ from the target model's greedy trajectories without any parameter training, and prunes low-value candidates before expensive verification. This preserves exact greedy output identity and yields up to $2.38\times$ speedup, showing that candidate supply and verifier-side valuation are complementary optimization stages.

Peer review is buckling under submission volume, and LLM assistance is gaining traction. This paper asks how four current LLMs treat 200 already-published articles from top-tier journals such as Nature, Science, and Cell. The authors find that model recommendations are overwhelmingly harsh and starkly inconsistent: $605$ of $800$ reviews ($75.6\%$) recommended major revision or rejection, yet the rate for any single model ranged from $17.5\%$ to $100\%$ ($p < 0.001$). The takeaway is that LLM editorial signals are too variable to replace human judgment, though convergent criticisms may still be useful.

DIAL tackles the challenge of fine-grained control over instruction adherence—how strongly an LLM should follow instructions like "Keep it short." The core idea is to distill classifier-free guidance (CFG) into a lightweight, text-conditioned steering adapter that edits hidden states at a single layer. This matters because it promises scalable, efficient, adjustable adherence across diverse instructions without requiring labeled responses or per-instruction fitting.

Instrumental convergence predicts that capable agents develop power-seeking subgoals, but it is unclear whether reinforcement learning with verifiable rewards (RLVR) monotonically increases such behavior in language models. This paper evaluates OLMo-3-7B-Think across sixteen post-training checkpoints using InstrumentalEval and finds a nonmonotonic trajectory: instrumental-like outputs rise during early RLVR, peak near step 200, and then fall below the SFT baseline by the final checkpoint. The result challenges simple "more RL equals more convergence" narratives and shows that checkpoint-level monitoring can reveal dynamics hidden by endpoint comparisons.

AgenticEvolve replaces the one-shot LLM variation operator in evolutionary program search with a persistent agent that edits code, runs self-designed experiments, and commits candidates for formal evaluation. By letting the agent test hypotheses locally between expensive evaluations, the framework aims to improve sample efficiency and final program quality on tasks ranging from online bin packing to symbolic regression. The authors report gains over a matched No-analysis ablation and over OpenEvolve, while noting that cost trade-offs depend heavily on model pricing.

Retrieval scores for clinical RAG are often computed against a single annotated chunk per query, conflating retriever quality with the completeness of the ground-truth annotations. This paper introduces RAG-MAMA, a breast-lesion question-answering system, and proposes mapping canonical evidence spans to all containing chunks so that overlapping segments with equivalent facts are not scored as competing candidates. The authors report nearly perfect retrieval metrics ($\text{Recall}@3 = 0.94$) on an 18-query development set and cautiously frame the work as a small methodological baseline, though the narrow scale and lack of independent validation severely limit its conclusions.

Multi-agent LLM systems are becoming common, but it is hard to predict which model steers the group when agents disagree. This paper designs a minimalist naming game with meaningless labels to isolate consensus dynamics from pretrained knowledge or correctness. Across seven model families, the authors find that larger, more capable models dominate the final convention even when no answer is right, and they explain this with a simple energy landscape in which bigger models carry larger assertion weights.

The paper defines superintelligent software engineering as the discipline of building software with agents that exceed human reviewers, arguing that raw capability is necessary but insufficient without governance that scales mechanically. It proposes that correctness depends on decisions and contracts outside the code, that such knowledge decays, and that once agent throughput $r$ outstrips fixed human review capacity $c$, the share of stale-built changes caught by reading falls toward zero. An exploratory snapshot of 452 popular GitHub repositories finds widespread stale agent context files, and the paper introduces OSTEON, a released discipline of governed knowledge repositories, mechanical gates, and decay control. The author commendably marks OSTEON's unmeasured claims as opinion and frames a falsifiable research agenda.

This paper develops a decision-theoretic framework for valuing memory in language agents, treating memory not as raw information but as a resource whose worth depends on future decisions, access costs, and revision capabilities. The core idea is to track value from original history through retained state to realized output, using avoided task loss as a common unit and introducing risk-cost frontiers for recovery and revision under uncertain demand. This matters because current memory architectures optimize proxy metrics like relevance or compression ratio without guaranteeing that retained distinctions actually improve downstream decisions.

Control evaluations test whether safeguards can stop a deliberately harmful AI agent, but they typically run in separate settings that may not resemble real deployment. This paper introduces simulated control evaluations (SCEs), which use a language model to generate tool responses in place of live execution, allowing evaluators to test attacks on a deployment's own tasks and codebases without causing harm. The central finding on LinuxArena is that SCEs predict safety metrics far more accurately than executing attacks in the mismatched BashArena control setting, suggesting that deployment-specific task coverage can outweigh the loss of execution fidelity.

Most automated agent-design research evaluates candidates on static, immediate-reward benchmarks, ignoring the non-stationarity, delayed feedback, and switching costs of real deployment. Cmas formalizes meta-agent architecture search as sequential incumbent replacement with an explicit false-promotion target, using paired shadow replay, interventional credit, and a conservative gate followed by simulated canary and rollback. The paper matters because it treats the architecture itself as a time-varying decision variable and asks whether we can safely replace a running compound agentic system without backtested-return overfitting. The answer offered is guarded: standard statistical ingredients are integrated sensibly, but the calibration is empirical and the guarantees are deliberately scoped.

This paper studies fine-grained intent classification on BANKING77 and CLINC150 by adapting a flexible decision-model interface (Laya) into fixed-label classification heads. For BANKING77 it applies sharpness-aware minimization (SAM) to a single encoder, while for CLINC150 it fuses probabilities from one Laya specialist and three DeBERTa-v3-large models. Rather than claiming a universal advance, the authors foreground statistical uncertainty, selection effects, and execution-path drift, making the work a transparent---if narrow---benchmark audit.

This paper addresses multi-horizon photovoltaic power forecasting from 15 to 240 minutes by asking whether residual error from a strong gradient-boosted tree can be reduced through selective, MAE-aligned combination with neural temporal predictors. The core idea is a constrained simplex mixture where the tree anchors the ensemble and weaker candidates can be assigned zero weight, with shared coefficients keeping the combiner parsimonious. The contribution matters because it isolates genuine aggregation value against a competitive baseline rather than inflating model count or comparing across mismatched tasks.

Language model rankings on question-answering benchmarks are known to shift with prompt wording, raising doubts about single-prompt evaluation. This paper investigates whether augmenting models with deterministic tools stabilizes pairwise comparisons, proposing a matched protocol that isolates answer readout, reasoning, tool invocation, and report access to attribute changes in wording sensitivity. Applying the protocol to six open models across five domains and five prompt wordings, the authors find that tools improve accuracy where a program directly answers the question yet frequently widen ranking instability because call frequency itself becomes wording-dependent. The work matters because it shows that deterministic tool output is not enough to guarantee stable evaluation, and it recommends reporting call frequencies and pairwise stability alongside headline accuracy.

PhysTrace evaluates physical consistency in videos by temporally aligning MLLM-generated, history-conditioned expectations with actual observations in a shared embedding space. Unlike prior work that probes internal representations or assigns video-level scores, it pinpoints when and why physical violations occur. The authors also introduce IntPhys2-Timing, adding temporal annotations to benchmark violation localization. Applied to diffusion and flow-based video models, PhysTrace reveals how physical inconsistencies emerge across generative timesteps—a crucial diagnostic as pretrained video backbones become policy priors for robot control.