Skip to content
AI NEWS.ccAI正在发生 未来正在更新
Bookmarks
THE JOURNAL
AI NEWS JOURNAL
AI, IN PERSPECTIVE

All AI updates

Explore the complete news stream by source, topic and date.

Newest first

arXiv cs.CV

Confusion-Adaptive Margin Learning for Handwritten Letter Recognition: A Controlled Study on EMNIST Letters

arXiv:2611.09953v1 Announce Type: new Abstract: Handwritten character recognition systems can achieve high accuracy while retaining concentrated errors among visually similar classes. Motivated by residual letter confusions in earlier geometry-based recognition work, we investigate whether these errors can be addressed through the training objective. We introduce Confusion-Adaptive Margin Learning (CAML), which constructs a sparse weighted confusion graph from a discovery partition held out from pilot training and applies pair-specific cosine-separation constraints during final training. CAML adds no classifier, routing rule, or inference-time parameters. We evaluate CAML on EMNIST Letters against a strictly matched cross-entropy CNN baseline, using the same architecture, final training data, deterministic augmentation, optimizer, and checkpoint selection. Eight confusion pairs are targeted: I/L, G/Q, U/V, R/V, D/O, C/E, I/J, and H/N. On the official clean test set, accuracy changes from 95.6346% to 95.6731%, and macro-F1 from 0.956326 to 0.956724. Mean accuracy across ten predefined corrupted conditions increases from approximately 95.3812% to 95.4303%; symmetric errors within targeted pairs decrease from 614 to 605 (1.47%). The global differences are small and not statistically reliable. CAML fixes 124 baseline errors but introduces 116 new errors on clean data (exact McNemar p=0.6515), and no corruption-specific difference remains significant after Holm correction. Thus, CAML does not establish global superiority, but shows that a leakage-controlled, pair-specific margin objective can alter targeted error structure with limited overall benefit for an already strong classifier.

arXiv cs.CL

DramaGrow: A Regime-Aware Benchmark for Auditing LLM Judges of Multi-Episode Drama

arXiv:2611.09938v1 Announce Type: new Abstract: LLMs can now draft complete multi-episode dramas, and LLM judges are increasingly used to evaluate them. Assessing the reliability of these judges is difficult because drama quality encompasses specification-backed errors, reproducible preferences, and legitimate disagreement among human raters. Most existing evaluations reduce these cases to a single score from one judge call and leave tie calibration, call-to-call variation, and the vote margin underlying the human reference unreported. We introduce DramaGrow, a regime-aware benchmark and audit protocol for AI-generated Chinese five-episode drama. The benchmark assigns verifiable constraint violations, four-of-five preferences, and within-panel disagreements to separate denominators and evidence rules. It contains 800 episodes, 120 byte-preserving interventions, and 1,155 blinded five-rater records, including a frozen extension with 114 supermajority pairs and 21 disagreement pairs. On the frozen sets of 114 preference pairs and 35 constraint interventions, the repeated absolute-score rubrics yield overlapping performance, with mean accuracies of 53.8-55.8% and three-call agreement of 29.8-32.5%, and their accuracy falls from 71.6-71.9% to 34.8-37.1% as the tie band widens. DirectPair-GLM reaches 85.1% preference accuracy, 89.5% constraint F1, and 81.9% Detect+Locate on the same items, and exact subpanel enumeration identifies four-of-five as the smallest stable human vote margin. Evaluation protocol, tie calibration, judge-call variation, and panel margin therefore affect performance as much as rubric design. We will release the specification packages, scripts, interventions, anonymized human votes and rationales, frozen judge runs, and analysis code.

arXiv cs.DC

FirstSwap: Pre-Consensus Trade Approvals for Decentralized Market Design

arXiv:2611.10254v1 Announce Type: new Abstract: Decentralized exchanges typically give traders no execution guarantee until a block is finalized. In the meantime, external prices move, traders strategically delay submission, and block producers can influence inclusion and ordering. We introduce FirstSwap, a DEX built on a constant-function market maker pricing curve that gives traders an early pre-approval before consensus, guaranteeing inclusion and execution within their submitted price limit, with final ordering determined by limit prices rather than by a block producer. We demonstrate the benefits of trade pre-approvals in an economic model combining auction theory with CEX-DEX arbitrage. When the CEX price changes gradually, FirstSwap rewards earlier traders with better prices, discouraging strategic delay and yielding higher social welfare and gains to AMM liquidity providers than a batch auction. When the CEX price jumps abruptly, FirstSwap yields higher welfare and smaller AMM losses than fair ordering. The central challenge is delivering fast trade pre-approvals without consensus. We introduce a decentralized, scaled admissibility protocol under which each validator reserves only a fraction of the available liquidity at every price. A trader obtains approval in a single round of messages with a quorum of validators. Somewhat surprisingly, despite Byzantine validators who may approve trades arbitrarily, we show that all quorum-certified trades are jointly admissible and execute within their limits. Our protocol extends naturally to a setting where buy and sell orders can be pre-approved against each other before a block is finalized.

arXiv cs.DC

OneReel: Efficient Multi-Scene Video Generation with Scene-Structured Attention

arXiv:2611.10594v1 Announce Type: new Abstract: Multi-scene video generation must follow scene-specific prompts while preserving consistency across scenes. Dense joint generation enables cross-scene information exchange but incurs quadratic attention cost, whereas separate generation avoids this cost at the expense of cross-scene consistency. We present OneReel, a training-free system that exploits scene structure to enable efficient multi-scene generation. Our key insight is that multi-scene generation can preserve complete intra-scene attention while restricting cross-scene information exchange to a bounded shared memory. OneReel uses scene prompts and durations to establish scene-specific attention scopes and dynamically selects a bounded set of representative latent frames for cross-scene sharing. This reduces video attention complexity from quadratic to linear in the number of scenes. To translate these savings into execution speedups, OneReel packs scene-local and shared K/V into dense variable-length attention and employs work-balanced multi-GPU partitioning with communication-aware selection between K/V gathering and Ulysses-style parallelism. Our evaluations across four video generation models show that OneReel achieves 1.7-4.9x end-to-end speedup over dense attention and 1.2-2.2x over existing sparse-attention advances, while improving text-following accuracy by 34-37 percentage points and maintaining cross-scene consistency.

arXiv cs.DC

Performance Benchmarking of Federated Learning Frameworks for Medical Image Classification

arXiv:2611.10688v1 Announce Type: new Abstract: Selecting a Federated Learning (FL) framework requires understanding the resource demands of its training and evaluation workflow. This work compares Flower, NVIDIA FLARE, and FedBioMed using two-dimensional MedNIST classification with ResNet-10 in a containerized, single-host environment. Following a dated review of repository activity and documented medical applications, we evaluated nine configurations per framework, with three random seeds, for a total of 81 executions. The configurations varied federation size, communication rounds, and batch size. We monitored execution time, CPU and RAM consumption, host-level GPU utilization, server network traffic rates, and model accuracy at the end of each execution. The implementations shared nominal learning settings while retaining differences in data splitting and federated evaluation. Configuration-blocked non-parametric analysis identified differences in execution time, client and server memory, GPU utilization, and network traffic rates. NVIDIA FLARE exhibited the lowest server-memory consumption and the highest host-level GPU utilization, Flower exhibited the highest average server network traffic rate, and FedBioMed required more client memory and longer execution time. No statistically significant difference in final accuracy was detected within the evaluated short training schedules; this does not establish convergence or equivalent predictive performance. The results characterize the configured workflows under shared hardware resources. They provide a concrete case study of system behavior, with conclusions bounded by workflow asymmetries, local networking, and the lightweight MedNIST workload.

arXiv cs.CL

Rethinking Supervision Targets for Multimodal LLM Unlearning: Less Private Information, Reduced Distribution Shift

arXiv:2611.09939v1 Announce Type: new Abstract: Multimodal Large Language Models (MLLMs) can memorize and reproduce sensitive information from their training data, posing significant privacy risks. Their ability to encode knowledge across both textual and visual modalities further complicates the selective removal of such private information. Many existing methods guide model unlearning by training the model toward constructed labels or output distributions, which we refer to as supervision targets. However, these methods often suffer from a trade-off between forgetting effectiveness and model utility. We identify supervision target limitations as a key factor in this trade-off and empirically show that targets containing more private information lead to insufficient forgetting, while larger distribution shifts from the original output distribution are associated with greater utility degradation. Motivated by these findings, we propose FFN-Perturbed Self-Distillation (FPD), which starts from the teacher's original output distribution as a zero-shift reference and selectively perturbs FFN activations associated with private information across textual and visual modalities. By precisely reducing private information from the teacher's original output distribution, FPD constructs supervision targets with less private information and reduced distribution shift. Extensive experiments on S-MLLMUn Bench demonstrate that FPD achieves a better forgetting-utility trade-off than existing methods.

arXiv cs.CV

Video-DCI: How Far Are VLM Agents from Human Viewers in Direct Video Corpus Interaction?

arXiv:2611.09951v1 Announce Type: new Abstract: Humans answer complex visual questions by inspecting short moments, updating what they know, and redirecting attention toward missing evidence. Existing video benchmarks, however, often provide the relevant video, evaluate only final answers or isolated localization, or permit textual shortcuts, leaving multi-round visual evidence seeking underexplored. We introduce Video-DCI, a diagnostic benchmark and functionally matched human-agent protocol for Direct Video-Corpus Interaction. It comprises DCI-Loc for temporal evidence localization and DCI-QA for end-to-end discovery, retention, and composition of cross-video evidence. In DCI-QA, videos within each pool are topically related, yet every answer requires hidden temporal evidence from multiple videos. Given only the question and video pool, humans and agents iteratively inspect bounded local windows, retain supporting evidence, and answer with cited temporal spans. Human trajectories provide reference curves for evidence acquisition and use across rounds. Experiments reveal a consistent retrieval-to-grounding gap. On DCI-Loc, frontier VLMs often retrieve and expose the reference moment, yet convert far less of this exposure into precise final citations than the human reference. On DCI-QA, the loss begins earlier and compounds across videos. Ablation studies further confirm that iterative search outperforms one-shot retrieval. Multi-round search is therefore necessary but insufficient: current agents accumulate candidates and observations more readily than precise, jointly sufficient evidence.

arXiv cs.LG

Adaptive Depth Aggregation in Graph Transformers for Long-Range Graph Reasoning

arXiv:2611.08500v1 Announce Type: new Abstract: Graph Transformers improve long-range node-to-node communication through global or structure-aware interactions. However, improving cross-node communication does not determine how computations produced at different depths should be combined. Standard residual streams accumulate successive computations with fixed unit coefficients shared across nodes. This fixed aggregation is a residual-interface constraint: once intermediate contributions have been produced, the interface cannot explicitly reweight them by node. We hypothesize that this constraint can create a depth-wise aggregation mismatch when the relative utility of these intermediate contributions varies across nodes and structural contexts. To study this hypothesis, we formalize Graph Transformer residual computation as a causal node-wise history of graph-propagation and feed-forward network (FFN) contributions, making depth aggregation an explicit interface choice. We realize node-adaptive aggregation over this history with Attention Residuals (AttnRes) while preserving the backbone's graph-propagation and auxiliary-state machinery. Block AttnRes reduces the routing cost of long histories relative to Full AttnRes by pooling consecutive contributions. We instantiate the interface in GPS and GRIT and evaluate it on all five ECHO tasks. Both variants achieve lower mean absolute error (MAE) than their independently tuned backbones on every task, and at least one obtains the lowest mean among all compared methods in each task. Routing-only donor-swap interventions support the mismatch hypothesis, while additional ablations examine the key design choices. Together, these results support node-adaptive depth aggregation as a useful complement to cross-node communication in long-range graph reasoning.

arXiv cs.DC

Adaptive Scheduling Pipeline for Multi-Instance Asynchronous Reinforcement Learning

arXiv:2611.08864v1 Announce Type: new Abstract: We present the Adaptive Scheduling Pipeline (ASP), an offline rollout scheduler for multi-instance asynchronous reinforcement learning (RL) of large language models (LLMs). ASP jointly reduces three sources of rollout underutilization - padding bubbles from intra-minibatch length heterogeneity, skewness bubbles from long-tail outputs, and sorting bubbles from fluctuations in training-queue occupancy induced by length-sorted dispatch - through a single offline preprocessing pass. This enables efficient multi-instance rollout that overcomes the early scaling plateau of single-instance inference while smoothing the arrival of generated samples at the trainer. We integrate the scheduler into an asynchronous architecture that decouples rollout from training while keeping policy lag observable and bounded. Complete ASP improves end-to-end throughput by up to 3.22 times and reduces iteration time by up to 76.8% relative to MindSpeed-RL, while reducing iteration time by up to 62.5% relative to verl. Within ASP, scheduling reduces mean iteration time by 10.9%. ASP also scales more robustly with additional generation resources than these baselines and shows no visible convergence degradation in the measured bounded-lag regime. ASP has since been used in an internal telecom-domain RL post-training workflow, reducing the reported training period from more than three months to about two months and producing a domain-enhanced model for knowledge answering and tool calling.

arXiv cs.AI

Aligning, Fast and Slow: A Control-Theoretic Oversight Framework for Long-Horizon Trajectories and Maintained Deployments

arXiv:2611.08781v1 Announce Type: new Abstract: AI agents can pass checks on individual tasks while retaining memory or code changes that alter their later behavior. We introduce a control-theoretic framework that couples a fast loop controlling proposed actions with a slow loop reviewing persistent changes and reassessing configurations after model or tool updates. When measurement errors and changes between checks are bounded and controllers can enforce their decisions, we derive a deployment-level bound and a rule for dividing review effort. In simulation, dividing a fixed review budget between both loops reduces mean maximum departure from intended behavior roughly twelvefold compared with review confined to individual runs. In coding stress tests with authored memories and five matched seeds, combined controls reduce mean severity on a 0-3 scale from 0.325 without gates to 0.117, resolving 33 of 120 issues versus 35 without gates; in trading stress tests, combined controls reduce mean shortfall in low-emission purchases from 0.892 without gates to 0.107. The overall framework helps developers and auditors compare and enforce oversight designs, while the agent studies test component controls rather than the full deployment-level guarantee.

arXiv cs.CV

An Anatomy-guided 3D Segmentation Pipeline for Prostate Surgical Phantoms from CT Scans

arXiv:2611.08652v1 Announce Type: new Abstract: Full autonomy has been a long-term goal in robotic surgery and a near-to-medium-term goal for prostatectomy. Due to the relatively closed structure of the procedure, its standard-of-care robotic/laparoscopic solutions, and procedure economics, prostate procedures are an optimal early target for surgical automation. Prototyping and testing of possible solutions is done in constructed laboratory phantoms, either synthetic tissues or animal phantoms. To enable measurement of perceptual and robotic motion error, phantoms undergo volumetric imaging; these images are critical to both a potential deployed automatic system, used in planning motion and treatment (cutting, ablation, suturing, etc.), as well as post-hoc assessment of system performance. Rapid iteration on these phantom models yields small, inconsistent datasets, while the current workflow for human annotation is often anatomically unreliable and prohibitively time-consuming, rendering deep learning unsuitable at reasonable scales. To address these limitations, we present an automated 3D segmentation pipeline that uses classical medical image-processing techniques to encode the known anatomy of the phantoms as explicit geometric and intensity priors, which requires no training corpus and is transferred to each new phantom revision without re-annotation. Across 57 scans spanning multiple phantom revisions, our pipeline recovers the capsule interior at 0.83 median Dice on 11 contrast-suppressed scans and detects all 32 annotated focal lesions without false positives, completing a full scan in 2-8 minutes on commodity CPU hardware versus 1-2 hours by hand.

arXiv cs.AI

Aspect-wise Pairwise Tournaments for Test-Time Selection

arXiv:2611.08640v1 Announce Type: new Abstract: Scaling test-time compute has proven highly effective at improving LLM reasoning. A popular method is best-of-n sampling, where a model generates many candidate solutions and one is selected. This motivates the question: how can we select the best candidate? When humans make choices, we find it more natural to compare two options than judge each in isolation. Additionally, when the options become complex, we like to judge the options one aspect at a time, such as comparing two job candidates on experience, communication, etc. In the test-time setting, candidate solutions should be judged in the same way. Here we introduce Aspect-wise Pairwise Tournaments (APT), which verifies by comparison and compares by aspect. Each matchup of a single elimination tournament is decided by a vote over multiple off-the-shelf aspect verifiers, each expressing a pairwise preference along its own aspect. Across five domains and three generator LLMs, APT improves over binary multi-aspect approval, self-consistency and reward model verification, with the largest gains on open-ended tasks. We find that both components matter; neither binary approval over the same aspects nor a generic pairwise judge at a matched budget matches APT. Because aspects are specified in natural language, APT also lets users steer selection toward properties they care about, better aligning with user preferences without needing to regenerate solutions. No component is trained, so APT transfers to new domains by simply rewriting prompts.

arXiv cs.CV

AtariBench: Evaluating Spatiotemporal Reasoning of Vision-Language Models in Atari Gameplay

arXiv:2611.08755v1 Announce Type: new Abstract: Vision-Language Models (VLMs) have achieved strong performance on static visual understanding benchmarks, yet their ability to act in dynamic environments remains less well understood. We introduce AtariBench, a controlled benchmark that evaluates VLMs across 20 Atari 2600 games, requiring models to make sequential decisions from high-temporal-resolution gameplay streams. By decoupling fast in-game dynamics from wall-clock inference speed, AtariBench focuses on whether models can perceive rapid visual state changes, reason over spatiotemporal dynamics, and ground actions in the current observation. Our experiments reveal substantial limitations in state-of-the-art VLMs: models often fail to infer object motion and direction, exhibit brittle sequential decision-making, and over-rely on textual interaction history despite newly appended visual feedback. We further introduce a separability analysis to identify which in-game scores provide reliable signals of model competence, and use the benchmark to study how memory design and prompting strategies affect performance. Overall, AtariBench exposes a persistent gap between passive visual perception and active decision-making, providing a lightweight and rigorous testbed for developing VLMs that can reason and act in dynamic environments.

arXiv cs.AI

AttentionSwitch: Jailbreak Defense for LLMs via Joint Learning of Sparse Attention-Head Selection and Scaling

arXiv:2611.08752v1 Announce Type: new Abstract: Large language models have achieved remarkable performance across a wide range of tasks, yet they remain susceptible to jailbreak attacks that circumvent safety mechanisms and elicit harmful or unintended outputs. Although safety can be improved through model fine-tuning, such approaches can be computationally expensive and may degrade previously acquired capabilities through catastrophic forgetting. Recent jailbreak defenses have therefore explored targeted interventions on specific model components associated with unsafe behavior. However, existing approaches typically decouple two inherently related decisions: where to intervene and how to modify the selected components. Such sequential optimization may be suboptimal because the optimal intervention location can depend on the intervention applied there, and vice versa. In this work, we introduce AttentionSwitch, the first jailbreak defense to jointly optimize intervention sites and strengths within a frozen LLM. Specifically, AttentionSwitch learns a sparse set of attention heads together with query-conditioned multiplicative scaling factors, allowing the outputs of the selected heads to be either attenuated or amplified without altering the directions of their representations. Across popular language models and jailbreak attacks, AttentionSwitch outperforms existing defenses, with particularly strong results when training data is scarce.

arXiv cs.AI

Cancellation Is Not Enough: What Retries Leave Behind in an Overloaded LLM Engine

arXiv:2611.08684v1 Announce Type: new Abstract: Clients and proxies in front of an LLM engine retry calls that time out. After a transient overload, these retries can keep the engine from recovering even when calls arrive below its capacity. Since the engine may keep computing for clients that have left, propagating cancellation looks like the natural remedy. In this paper, we test it on vLLM with overload experiments on three open models, and with scheduler traces of one that follow each client attempt to the execution it creates in the engine. We find that with retries on, cancellation alone did not restore service in any run, and only 12-38% of the calls that arrived after the overload were delivered. The traces suggest why: retries that vLLM ran typically started prefill less than half a second before they timed out. Executions that vLLM later aborted took 79-85% of the prompt tokens it computed, and vLLM's prompt-token counter missed a third to half of all computed prompt tokens. A retry budget restored service by shedding retries but left many of the overload's calls undelivered. Request coalescing delivered every call from the overload onward, some late, unless cancellation had already aborted the running execution. Overload controls should thus be judged by the executions that retries add and the calls that clients receive.

arXiv cs.CL

Compression of Emotion in Language Model Representations

arXiv:2611.08683v1 Announce Type: new Abstract: Large language models (LLMs) have been evaluated on their ability to identify emotions expressed in text and produce fluent emotional language. However, output predictions alone do not reveal how fine-grained emotion distinctions are organized internally. We study seven LLMs on GoEmotions, an English Reddit dataset annotated with 27 fine-grained emotion categories plus neutral. We compare model predictions and emotion-prompted final-token representations with the 28 GoEmotions categories and with coarser emotion-family and sentiment groupings. We find a consistent pattern of $\textit{overcompression}$ relative to the fine-grained GoEmotions taxonomy. Under the tested KMeans protocol, when we cluster their hidden representations without using emotion labels, the clearest split is into two groups that largely separate positive from negative emotion. Coarser emotion groupings are also recovered with fewer clusters than the full 28-category taxonomy. Yet linear probes retain substantial fine-grained predictive information. These results suggest that fine-grained emotion information remains accessible even though broader affective structure more strongly organizes the unsupervised geometry.

arXiv cs.CL

Contextuality and Contradictions: Exploring Real-World Identifications of AI-Generated Writing

arXiv:2611.08704v1 Announce Type: new Abstract: As AI-generated content spreads into nearly every domain of human interaction, practices and discourses of AI detection, both formal and informal, have become an everyday part of how people interact with text and with each other. In this paper, we explore the real-world practices and discourses of AI identification that have emerged in response to a shifting discursive landscape where the human provenance of a text can no longer be taken for granted. We pair qualitative coding with computational analysis of lexical change, grounded in linguistic anthropological perspectives on language and communication, to investigate AI identifications among English Wikipedia editors and students. We find that these identifications draw on a diversity of evidentiary types, from discrete textual features and judgments about content to circular sense-making claims and presumptions about the author's moral character, and that they are contradictory, dynamic, and contextually dependent. In practice, what often motivates an AI identification is not the presence of any specific textual element, but pragmatic infelicity: a perceived mismatch between the elements present and those deemed appropriate within the text's particular context. Consequently, real-world identifications of AI-generated writing do not merely neutrally point out the perceived presence of "AI-like" features, but can be leveraged as a discursive resource to maintain and define sociolinguistic norms: what can be said in a given context, what is appropriate, and, ultimately, what is human. Further, the design and implementation of systems and tools that treat AI-likeness as a context-free property of text may therefore inherit, and enforce at scale, the norms of whichever contexts they were trained on.

arXiv cs.DC

DataHelix: A CPU-GPU Co-Scheduling Data Plane for Multimodal Model Training

arXiv:2611.09411v1 Announce Type: new Abstract: Multimodal model training breaks the loose coupling between data loading and model execution assumed by conventional training pipelines. Large per-sample variation in decoding cost, token/latent count, and tensor size causes both CPU preprocessing time and GPU compute time to vary substantially across data-parallel ranks (training processes). In synchronous training, these two sources of heterogeneity interact tightly: GPU execution determines how long the input pipeline has to prepare the next batch, while delayed preprocessing in any rank in turn stalls all GPUs at synchronization. As a result, optimizing CPU throughput or GPU load balance in isolation cannot eliminate pipeline bubbles. We present DataHelix, a metadata-first CPU-GPU co-scheduling data plane for multimodal training. DataHelix jointly reasons about preprocessing demand and GPU execution progress, treating each GPU compute window as the readiness budget for the next batch and coordinating preprocessing across ranks to minimize deadline misses rather than CPU makespan. To avoid expensive data movement, placement is decided while samples remain lightweight descriptors, before decoding and collation amplify them into large tensors. Descriptor-level token-latent-aware batching performs coarse pre-balancing, while cost-gated collective rebalancing corrects residual skew only when its expected benefit exceeds transfer cost. Across 11 video and multimodal workloads on 8 Ascend 910B accelerators, DataHelix reduces mean deadline-miss rate by 67.8% and mean p95 input-wait latency by 87.0%, while achieving 2.72x geometric-mean throughput over rank-local DataLoaders and 1.82x over a Ray actor pool under the same CPU budget.

arXiv cs.AI

Decision Context: Persistent Behavioral Learning for AI Agents

arXiv:2611.08644v1 Announce Type: new Abstract: LLM agents do not learn from corrections across sessions. A user fixes the same mistake at the start of every conversation. Existing agent memory systems primarily handle facts, such as preferences, entities, and conversational history, rather than procedural knowledge. We present Forge Decision Context (forge_dc), a system that captures behavioral signals from human corrections and self-assessed tool-call failures, compiles them into imperative RULE directives, and injects them at the decision points where they apply. We evaluate forge_dc alongside four predecessor approaches on DC-Bench, a new benchmark of 71 scenarios across nine behavioral dimensions, built from real executed agent traces in a deterministic mocked-tool environment. forge_dc reaches 95% compliance, up from 55% with naive retrieval, while injecting approximately 10 times fewer tokens per query. Across the five-stage progression, the strongest systems represent behavioral memory as explicit directives and deliver it at the decision points where it applies. The results identify representation, activation, and placement as central design choices for persistent behavioral learning.

arXiv cs.LG

Dual-Stream LLM-Guided Reward Design for Reject-Inference-Aware Reinforcement Learning in E-Commerce Payment Fraud Detection

arXiv:2611.08706v1 Announce Type: new Abstract: Production fraud-detection models learn from a censored slice of their own traffic: transactions a legacy system blocked never receive fraud labels, so models train on a biased, approved-only distribution. We extend LLM-guided reward design, the framework of Qu et al. (2025) in which an LLM evolves the reward of a two-step REINFORCE fraud-blocking policy, to attack this reject-inference problem inside the reward. A single LLM designs a dual-stream reward: stream 1 applies the standard dollar-weighted precision objective to labeled transactions; stream 2 scores legacy-blocked, unlabeled transactions by contrasting the newest gradient-boosted risk signal against the legacy scores that drove the original block. Rather than prescribe which signals matter, we expose a superset of evidence and let the LLM decide for itself which to use, guided by evaluation feedback alone; because it writes the reward as code, that choice is directly auditable. On a held-out four-month out-of-time test, the dual-stream agent beats the prior labeled-only agent at every target recall, by +4.7, +2.4 and +3.0 points of dollar-weighted precision at 80%, 85% and 90% recall, and the production baseline by +6.8 points on average. The resulting policy is architecturally identical to that of Qu et al. (2025) and adds no inference cost.