Personal agents draw on past conversations to tailor their responses to each user. Their memory systems store and update what each user says, and coding agents now also rewrite the harness that turns this memory into the model's context. Such patches risk overfitting, because one tested only on the failures it targets can fix those cases and help nowhere else. We present Reglos, a training-free personal memory agent that tests whether harness changes help beyond the failures they were written to fix. Its self-evolving memory adds to each user's append-only record with every conversation, as new preferences supersede old ones and facts can be forgotten. Its harness improves through recursive self-improvement, in which a coding agent proposes patches from failure traces and a critic from a second model family screens each one for benchmark gaming. After the harness is frozen, a fixed evaluator decides which further patches to accept. With the same backbone behind every system, Reglos outperforms state-of-the-art memory systems such as Mem0, MemOS and EverOS, reaching 61.3% and 84.0% on PersonaMem-v2 and LongMemEval-S with Qwen3.8-27B, and 57.5% and 87.2% with GPT-5.6 Luna.
In recent years, AIGC technologies are capable of generating hyper-realistic forged facial images, posing severe threats to facial security safety. To tackle these challenges, early Deepfake research primarily centered on real-fake classification. Recently, growing efforts have been devoted to the Deepfake Attribution (DFA) of generated content. Nevertheless, existing research on Deepfake detection and attribution faces saturated performance in binary classification, limited diversity in datasets and algorithms, and imperfect evaluation protocols, which severely impede practical application. To address these limitations, we propose a comprehensive deepfake detection and attribution benchmark named DeepfakeGenome (DFG). It contains 100 facial forgery algorithms and 2M images in total, achieving 4x to 100x larger than prior DFA benchmarks. We further design 4 protocols for practical evaluation, including a novel retrieval-based attribution paradigm. Unlike previous open-set evaluation metrics, the proposed retrieval metrics are more aligned with the real-world active defense situation of blacklist registration mechanisms. Based on these elaborate designs, we investigate the performance ceiling of deepfake attribution task. Over 2k experimental evaluations are conducted, and 10 insightful findings are derived. We believe this work can provide new insights into the DFA research field. The Code and Benchmark of DFG are avaliable on https://huggingface.co/spaces/DeepfakeGenome/DFG_Page.
We view a $k\times n$ Latin rectangle as $n$ points in the grid $[n]^k$. Each symbol gives one point, whose coordinates are the columns it occupies. The rectangle is balanced when these points are pairwise equidistant in the $\ell_1$ metric. Its imbalance measures the distance from this ideal. We study the minimum imbalance from three sides. Constructively, we use a simulated annealing that moves symbols within rows and only penalises column clashes. Against the reference table of Voboril et al. (2025), it improves 33 of the 51 non-square values for $8\le n\le 13$, by up to 39.4%, and matches the rest. A cheap invariant, the row group, certifies that none of the 33 new rectangles is isotopic to rows of a group table, so the earlier column-swapping local search cannot reach any of them. Exactly, a computer-assisted branch and bound proves that the minimum imbalance of a $3\times 9$ rectangle is $56$. D\'iaz et al. found this value, but its optimality was open. A sharper pruning bound, derived from the identity $I=2\sum(d-t)^+$, reduces the search tree by a factor of $2.4$ to $3.3$ on the instances with $k\ge3$ that we solve. Asymptotically, the minimum imbalance is $\Theta(n^3)$ for every fixed $k$. A pigeonhole argument gives a lower constant of order $k\,4^{-k}$, and random rectangles give an upper constant of order $\sqrt k$. For $k=2$, the limit of $\min I/n^3$ exists, a permuton attains it, and it lies between $4/375$ and $0.1196$. Perfect balance is an $\ell_1$-equilateral problem, so it forces $k=\Omega(n/\log n)$ rows. For $2\times n$ rectangles, single two-minute annealing runs improve five of the seven six-hour MIP upper bounds for $14\le n\le 20$. Finally, the Lean 4 kernel checks every rectangle reported in this paper against a 74-line specification that imports only Lean's core library. We also prove the pigeonhole lower bound in Lean with Mathlib.
Helping users discover new interests is vital to the long-term health of industrial recommender systems, yet the usual remedy, boosting exposure of under-explored (discovery) content, often backfires in production. We report URED (Unexposed Recycling for Efficient Discovery), a training-free recall-stage system serving Taobao feed recommendation at Alibaba for over six months to tens of millions of daily active users. It rests on a production-validated insight: naive boosting fails because it injects cold, information-poor candidates that rankers systematically over-estimate, buying no discovery gain while risking engagement. Items scored highly but never exposed (signals industrial pipelines routinely discard) form a pool that is both relevant and reliably predictable. Over a 30-day horizon URED aggregates them and re-injects at recall time through a lightweight pipeline spanning offline storage, distributed aggregation, and low-latency online serving, adding under 1% latency via a single O(1) lookup per request with no model retraining. Online A/B tests show +3.47% discovered categories and +8.47% content richness with engagement preserved: completion video views rise significantly (+2.62%) and total views move positive (+0.36%, not significant). Ablations confirm both design pillars are necessary: a one-day window collapses the discovery gain to +0.55% (not significant), and random injection without score filtering yields no gain while trending negative on engagement (-0.17%). Across production recall channels, URED reaches 13.9x the platform-average discovery rate (6.3x the strongest nearline baseline) at engagement quality on par with the platform average. We share deployment details and lessons for practitioners facing the exploration-exploitation dilemma under tight resource constraints.
Channel sounding in Wi-Fi multiple-input multiple-output (MIMO) systems must balance channel state information (CSI) accuracy against CSI acquisition overhead. Device mobility causes channel aging, while autoregressive CSI prediction introduces errors that accumulate over sounding intervals. However, standardized statistical channel models cannot fully capture the environment-specific channel dynamics encountered in physical artificial intelligence (AI) applications, motivating the use of Sionna-based ray tracing to evaluate CSI prediction under device mobility and evolving propagation geometry. This paper develops a dynamic sounding framework that integrates autoregressive CSI prediction with signal-to-interference-plus-noise ratio (SINR)-constrained interval selection. A Transformer with rotary position embedding that converts device positions into CSI sequences predicts CSI across multiple subcarriers for beamforming updates between sounding exchanges. We formulate sounding-interval selection as a predicted average-throughput maximization problem under per-stream SINR lower-bound constraints, including CSI acquisition and processing overhead. A prediction-error perturbation analysis provides the SINR bounds, and cumulative-sum enumeration globally solves the finite single-variable problem. Actual SINR feasibility holds whenever the prediction errors satisfy the bounds. Ray-tracing experiments assess channel prediction, trajectory transfer, and autoregressive prediction errors. In single-STA Wi-Fi 7 link-level simulations, predicted-CSI beamforming improves mean goodput by 67.3%, 112.7%, and 119.1% over stale CSI for linear, circular, and cosine trajectories, respectively.
Graphs play an increasingly important role in large language model (LLM) systems, supporting structural reasoning, knowledge grounding, retrieval, and memory. As these systems evolve toward the agentic era, graphs also organize execution dependencies and multi-agent communication, linking intermediate observations and retained information to subsequent decisions. However, understanding their contribution requires distinguishing the function of each graph operation, the component responsible for its computation, and the subsequent use of its output. This survey introduces an operation-level role--locus taxonomy for reasoning, knowledge and evidence, retrieval and memory, execution, and coordination. We compare how graph observations and external computations guide decisions, how memory representations and updates support later tasks, and how dependencies constrain execution and information delivery. These comparisons distinguish improvements in local metrics from downstream task performance while retaining the conditions under which results were reported. We then organize benchmarks, comparison designs, process diagnostics, and system costs into an evaluation framework. Finally, we identify open challenges in adaptive computation and stopping, memory maintenance, and dependency-aware recovery. The survey provides a basis for assessing when graph mechanisms are useful and for evaluating their effects across components and stages.
Diffusion drafters accelerate speculative decoding by denoising multiple draft tokens in parallel. Their training uses anchor sampling to construct compact draft queries and performs cross-attention with a block-sparse mask over a much longer context. We empirically show that acceptance length remains stable as context scales with a fixed query budget, while increasing the query budget yields diminishing quality gains, jointly establishing a practical regime with queries far shorter than the context. No existing context parallelism (CP) method jointly exploits this query--context asymmetry and meets the nonlinear, sample-dependent load-balancing requirement created by anchor sampling. We present Split-KV, an exact CP method that keeps key/value shards local and limits per-layer attention communication to query-sized tensors, making this communication independent of context length. Its global query view further enables a cost-aware longest-processing-time-first (LPT) strategy that balances context key/value chunks by their realized attention cost. At a CP size of 64, Split-KV retains 85--91\% weak-scaling efficiency and achieves up to $3.67\times$ speedup over the fastest available CP baseline. LPT achieves $<$1\% imbalance and plans three orders of magnitude faster than the best load-balancing baseline. Code is available at https://github.com/zxgx/verl.
Reasoning models allocate test-time compute across distinct behaviors such as planning, deriving, and verifying. However, models frequently "overthink," continuing to verify and re-derive long after finding a correct answer. This wastes expensive compute and can even degrade performance. Existing interventions, such as token budgets or steering, primarily regulate reasoning duration rather than targeting specific behaviors. In this work, we demonstrate that Mixture-of-Experts (MoE) routers provide a natural, training-free interface for behavioral control. We first split reasoning traces into sentence-level segments and label each with one of six reasoning behaviors (planning, execution, restatement, fact, reflection, and conclusion), so that every token carries the behavior of its segment. Linear probes on router logits predict this label nearly as well as probes on the residual stream, the model's full hidden state and the standard target for probing, while using about half as many features. These signals generalize: experts associated with a reasoning behavior on one dataset predict the same behavior on the others, across six datasets spanning five domains (math, code, science, knowledge, and commonsense). Building on this, we selectively edit expert router scores during inference to curb overthinking. Across four models and four benchmarks, this intervention reduces the number of generated tokens by a median of 30.4% (and up to 63.3%). Crucially, it encourages the model to stop once the problem is solved, rather than skipping essential reasoning steps. Accuracy never drops by more than 0.5 percentage points, and frequently improves. Ultimately, MoE routing offers both an interpretable lens into reasoning states and a lightweight mechanism for optimizing test-time compute.
Merging companies that each run Salesforce must consolidate their customer data into a single organization while both source systems remain in use. This requires reconciling different schemas and picklists, resolving duplicate customers across and within organizations, preserving relationships between records, and keeping the target current until cutover. This article proposes a reference framework for an autonomous extract-transform-load pipeline for this task. It combines partitioned bulk extraction, declarative mapping templates, rule-based and statistical quality gates, master data management with deterministic and blocked fuzzy matching, constrained loading with per-record error capture, and change data capture guarded by a data-health circuit breaker. Each stage is mapped to Jitterbit Integration Studio constructs. The framework was implemented as an open reference prototype and evaluated on synthetic two-organization scenarios with known ground truth at three scales, from 32,040 to 383,973 source records, over five random seeds. At a preset threshold, hybrid matching reached pairwise F1 scores of 0.918 for accounts and 0.885 for contacts; calibrating the threshold per object would raise these to 0.938 and 0.963. The quality gates prevented all load failures and all silent loss of unmapped picklist values without wrongly flagging any clean record, whereas removing them produced about 1,000 load failures per run. Over five simulated days of hourly incremental synchronization, replayed events were ignored, an injected upstream fault was contained in the batch in which it occurred, and the target agreed with a full rebuild on more than 99.3 percent of field values. Validation on a production deployment remains future work.
Planning electricity systems requires repeated decisions that connect regional resources, investment, operation and network constraints. Here we present an autonomous planning framework that carries prepared regional scenarios through proposal, numerical evaluation, feedback interpretation, redesign and final plan selection. Agents propose investment bounds, storage operating conditions and feeder interventions, interpret numerical feedback, and decide how to revise designs and when to stop. The framework automatically executes annual investment-and-dispatch optimization and representative-feeder verification, applies fixed acceptance criteria and retains the lowest-cost accepted candidate. We evaluate forty carbon-constrained trajectories across ten system-year scenarios in five electricity systems and four countries, using two model deployments and single- and multi-agent architectures. All forty retain candidates satisfying the implemented physical and fixed annual operational carbon-budget checks. Twenty-eight of forty matched reference designs exceed the new scenario ceilings. A separate archive of ninety-six trajectories under a common 1,000 Mt annual ceiling provides matched references and broader regional coverage. The results demonstrate autonomous planning under explicit emissions constraints using public regional data and published institutional demand outlooks.
Patient world models are increasingly proposed as simulators of clinical trials, and recent surveys rank their capabilities on a ladder from temporal prediction to counterfactual rollouts. Position on that ladder is currently assigned by reading papers rather than by measurement. We turn the counterfactual rung into a measured quantity. From 30,119 completed phase 2/3 trials with results posted to ClinicalTrials.gov we construct 9,996 audited within-trial contrasts from 7,207 trials, and we score a simulator on how much of its accuracy depends on knowing which intervention was tested. The instrument is an intervention-blind reference: the same model, trained on the same trials, with every intervention-identifying field withheld. Interventional skill is the error the full model removes relative to that reference. On a temporal split (878 test contrasts from 655 trials) a registry-trained gradient-boosting simulator removes a fifth of the error of a trivial baseline, but its interventional skill is -0.12 (95% CI -0.21 to -0.05): withholding the intervention makes it better, and no alternative featurisation, stratum, temporal cut or same-molecule retrieval yields an interval above zero. A frontier language model reading only the registered protocol has interventional skill 0.20 [0.08, 0.31]; smaller and open-weight models sit at the limit of detection. Perturbation tests show the large model uses the intervention as such: swapping the arms reverses the sign of its prediction in 86% of contrasts. The blind reference is itself not drug-blind (a language model recovers the withheld drug class 45% of the time from the blind prompt), so interventional skill is a conservative lower bound. A reliability layer shows that ranking by predicted error lowers it but that calibrated error control does not transfer under temporal shift. Benchmark, code, prompts and per-call model outputs are released.
As sequence lengths continue to grow, distributed attention based on sequence parallel (e.g., Ring-Attention) has become a mainstream solution for large language models (LLMs). Nevertheless, inherent communication dependencies limit computation-communication overlap, reducing GPU utilization. Moreover, in cross-hardware (NUMA or node) architectures, due to imbalanced communication bandwidth, frequent communication synchronization will limit overall performance. Based on these insights, we propose CHARM+, Cross-Hardware Attention with Re-Merge Multistream Mechanism. We reorganize sequential communications into several levels and achieve intra-level parallelism and computation overlap. Furthermore, we decouple cross-hardware communication from intra-hardware communication, eliminating synchronization overhead caused by bandwidth imbalance. Forward and backward algorithms were evaluated on multiple hardware architectures and consistently outperformed state-of-the-art methods. Compared to Ring-Attention, our forward and backward algorithms achieve average speedups of 1.3x and 1.7x on a dual-NUMA PCIe architecture, and 2.8x and 2.9x on a dual-node NVLink architecture, respectively.
KV cache eviction is not merely a token-ranking problem: the workspace retained for generating eviction queries also determines the quality of those queries. Consequently, an early workspace error can distort later query observations and reinforce an incorrect final eviction decision, particularly when evidence is dispersed across a long context. We present EchoKV, which discovers and addresses this failure mode with iterative correction. It initializes a feedback-compatible bootstrap workspace using proposals at future decode positions, then uses early decode-generated queries to recover the workspace before the remaining lookahead steps. It uses the same number of workspace forward passes as LAQ while adding a tensor re-scoring operation for workspace repair. On LongBench (16 tasks, Mistral-7B-Instruct-v0.2 and Qwen2.5-7B-Instruct), EchoKV shows notable improvements over LAQ on several tasks requiring dispersed evidence, including multi-hop cross-document QA and long single-document comprehension. Across budgets and backbones, EchoKV selectively improves evidence-intensive tasks while maintaining comparable performance on the remaining LongBench subsets. These results identify query--workspace coupling as a practical failure mode of query-driven KV eviction.
Recycling a target model's own predictions offers a training-free source of proposals for speculative decoding. In vision-language models, however, existing recycling indexes predictions by the last token alone, conflating predictions computed under different images and prefixes; we find that such reuse fails most often at image-sensitive decoding states. Longer keys reduce this aliasing but match fewer states. We introduce \method, which stores target predictions under one-, two-, and three-token keys: contextual memories live only for the current request and are thus implicitly conditioned on its image, while token-level memory persists across requests to retain coverage. At each lookup, entropy-confidence weighting fuses the retrieved predictions according to their key specificity and concentration. Best-first search composes the fused transitions into a candidate tree, supplemented by suffix-matched continuations, and one target pass verifies the tree under greedy decoding. Without training or an auxiliary draft model, \method\ achieves mean speedups over autoregressive decoding of $1.66\speed$ and $1.56\speed$ across seven single-turn benchmarks on Qwen2.5-VL-7B and LLaVA-1.6-7B, exceeding the strongest training-free baseline, Token Recycling ($1.51\speed$ and $1.37\speed$). Our implementation is available in https://github.com/Senmo996/RECAP-official-code.
Biological discovery requires using molecular knowledge and new measurements to infer responses beyond the conditions already observed. We introduce CellCanon, the first agentic cellular world model to unify single-cell expression prediction, evidence-grounded explanation, and observation-driven updating. Its predictive core reuses condition-level biological evidence across gene-specific numerical readouts to guide single-cell expression generation. A shared world state links predictions, acquired measurements, and source-linked claims, supporting explanation and revision. On 339 held-out SciPlex3 conditions across three cell lines, CellCanon improves differential-expression AUPRC and Macro-F1 over the best baseline for each metric by 63.3% and 17.3%, respectively, and direction Macro-F1 by 7.2%. Relative to gene-wise reasoning on matched drug-cell cohorts, measured language-model token use falls by 99.77% for Predict and 98.41% for Explain. On the source-grounded benchmark, Explain correctly answers all answerable questions with supporting proofs and identifies all insufficient-evidence cases. In simulated experimental feedback, Update uses one measured source context per drug to exceed the best fine-tuned baselines in differential-expression AUPRC and Macro-F1 without changing model weights. Together, these capabilities support a cumulative account of cellular responses, with predictions and biological judgments that can be revised as evidence accumulates.
When does an AI assistant improve human decision-making, and when does it hurt? We argue that the answer depends not just on AI accuracy but also on how much its evidence overlaps with the human's and how the human combines the two. A worker makes a binary decision by observing her own signal and receiving a recommendation from an AI. We develop a human behavioral model with two primitives: her awareness of the overlap and her trust in the AI. Our analysis gives a closed-form phase diagram over AI capability and overlap, with regions of complementarity, impairment, and automation. A worker who neglects the overlap (correlation neglect) can be made less accurate by a more capable AI, double-counting the shared evidence. Trust reshapes the regions - undertrust may compensate for correlation neglect, so that the trust level maximizing accuracy depends on the deployment parameters. Empirically, we identify the two model parameters by pairing the same workers with different AIs, and break the accuracy loss into a correlation-neglect cost and an undertrust cost, one tied to reducing overlap on the AI side, the other to calibrating trust on the human side. On ChatBench, AI assistance raises accuracy by 21 points, yet workers apply only about a quarter of the Bayes-rational weight to AI information; GPT-4o falls in the automation region and Llama-3.1-8b stays at the complementarity boundary; and calibrating trust recovers about 2.6 points for GPT-4o but almost nothing for Llama-3.1-8b.
Machine unlearning aims to remove a user's data from a trained model after a deletion request. Existing methods are majorly evaluated on three axes - how well they forget, how much capability they retain, and how robust they are to membership inference attacks, but not on whether the provider can offer the user any evidence that the deletion occurred. To fill this gap, we introduce auditability as a fourth axis, defined as the separation between requested (forget set) and retained (retain set) data under a statistic the provider computes. We show that across eleven unlearning methods on TOFU, MUSE and RWKU no method satisfies all four axes, with different methods falling short on different axes. To address this, we propose a two stage method that can be used on top of any unlearning method, in the the first stage we propose ${GapWiden}$ which uses a hinge loss to increase the difference between audit statistics of forget set and retain set. In the second stage we propose ${GapVeil}$, an inference time rejection-sampler which uses the base model as the drafter and re-samples based on the audit statistics of the drafted response. We observe that using our method, auditability and leakage reach oracle parity for every unlearning method, and forget performance converges to the oracle; retain performance improves for nine of eleven methods but remains short of parity. Our results show that evidence of deletion or auditability can be constructed by design, and that treating it as an explicit requirement changes the viability of existing unlearning methods.
Prenatal ultrasound biometry must return a head circumference (HC) and refuse when the image cannot support one. Selective prediction trades coverage against error, but its finite-sample guarantees bound the rate of gross errors, not their magnitude. We audit a tool-grounded biometry pipeline on two separately pre-registered four-way splits of the 999 labelled HC18 images, retraining the segmenter three times on each. Calibrating the contour-completeness threshold by exact Clopper-Pearson risk control returns a certificate indistinguishable from a hand-set floor, yet the pipeline still emits a 271.3 mm error: a small, geometrically complete ring fitted to an image whose true HC is 324.6 mm. The accepted candidate scores 0.94 completeness, so the score cannot see the failure, and one outlier among $\approx 150$ answers leaves the certified risk bound ($\alpha = 0.10$) intact. A second, untouched split reproduces the mechanism on a different image (233.3 mm). Constraining the admissibility region---the anatomical range in which the measurand is defined---removes the family instead: raising the lower bound from 50 to 55 mm costs 0.4 coverage points and cuts the worst-case error on both splits to 11.4 and 11.1 mm, whereas suppressing it through the completeness threshold costs 19.0 and 12.1 points. At an anatomically defined 70 mm the frozen tests give 1.91 and 1.81 mm MAE.
This vision paper argues that distributed learning, including federated and decentralized, should become a first-class paradigm in LLM post-training, not merely a privacy add-on. Current methods like RLHF and DPO rest on an invisible assumption: preference data should be centralized. We contend this assumption is scientifically unjustified, practically unsustainable, and ethically problematic. As LLMs deploy into healthcare, law, finance, and multilingual communities, preference data originates from populations that cannot or will not centralize. We develop a three-pillar argument grounded in (1) privacy and data sovereignty, (2) pluralism and representational equity, and (3) robustness and scalability. We formalize key technical challenges unique to this intersection, reference drift, preference heterogeneity, and the DP-utility trade-off for contrastive losses, and propose concrete algorithmic directions, and sketch a research agenda. Our analysis shows distributed post-training is not only necessary but tractable, and ignoring it risks AI that is legally noncompliant, representationally narrow, and brittle at scale.
Assessing the deployment performance of learned robot policies during training is challenging, as directly measuring task success requires costly real-world trials for each candidate checkpoint. Averaged validation loss offers a rollout-free alternative, but its checkpoint rankings can be inconsistent with deployment performance. We analyze two limitations of this criterion: it penalizes action deviations without considering whether the policy can recover from them, and gives later reference predictions equal weight even when earlier deviations may lead execution away from the reference. We propose RACE, a novel rollout-free metric for ranking checkpoints using a fixed set of held-out successful trajectories. RACE calibrates action deviations against state-conditional thresholds derived from action differences at similar states in these trajectories. We aggregate these scores through a running product, so that a substantial deviation reduces the contribution of subsequent reference predictions. Averaging the resulting cumulative scores over time and across trajectories yields the checkpoint score. In simulation, RACE outperforms baselines across validation sources, while several baselines degrade sharply when the validation set changes. On real robots, its rankings align with deployment success across the evaluated policy backbones and tasks of varying difficulty. Perturbation analyses further show that RACE distinguishes deviations with markedly different effects on task success, even when baselines respond similarly.