Skip to content
AI NEWS JOURNAL
AI, IN PERSPECTIVE

Search AI news

Search headlines, summaries and articles to connect the dots.

Newest first

TALD: Transferability-Aware Logit Disentanglement for Cross-Modal Knowledge Distillation

arXiv:2611.25126v1 Announce Type: new Abstract: Cross-modal knowledge distillation can incur negative transfer when source-modality supervision conflicts with the target modality's discriminative characteristics. We propose TALD, a two-stage framework for Transferability-Aware Logit Disentanglement in class-aligned logit space. Guided by task-dependent self-benefiting and cross-modal exclusivity criteria, a Residual Disentangler decomposes logits from a teacher and an independently trained target-modality reference student into modality-general and modality-specific components. A Sample-Adaptive Modulator regulates decomposition constraints and sample-specific fusion weights. The learned modality-general components are then adaptively fused to supervise the final student. Across five benchmarks spanning vision, audio, and text, TALD achieves the largest mean accuracy gain and highest positive-transfer frequency among the compared methods. Across ten transfer directions with five runs each, TALD improves over training without distillation in 84% of evaluations, compared with at most 70% for the baselines, while offering a favorable accuracy-efficiency trade-off. We plan to release the code.

Thinflow: Organizing HPC Benchmarking Campaigns as Rerunnable Directory Trees

arXiv:2611.26526v1 Announce Type: new Abstract: A benchmarking campaign provides the measurements behind an HPC performance study: code variants are built, swept over core or node counts, and measured repeatedly. Moving such a campaign to another supercomputer or rerunning it months later requires work that is not part of the study itself. Job scripts are adapted to each supercomputer, and what exactly ran is often reconstructed afterwards. We present Thinflow, a small Bash engine for organizing HPC benchmarking campaigns as rerunnable directory trees of tasks and runs. A task is a directory with a short shell script next to the code. For each sweep, one command builds the code, runs the benchmarks, and plots the results. Thinflow derives the execution order from declared dependencies and writes it to a plain-text manifest. Per-supercomputer profiles submit the runs to a workload manager such as Slurm. Each run writes a provenance record of its environment, hardware, and parameters. The engine ships inside the campaign directory, so a campaign stays rerunnable without installing a separate tool and, once completed, may serve as a paper artifact. In a case study, we ran one campaign with identical task files on three Slurm partitions of two supercomputers, on up to 16 nodes. This strong-scaling study of matrix multiplication in C/OpenMP, C/MPI, and Julia/MPI comprises 37 task files with 217 non-comment lines. Its 455 benchmark runs recorded what differed per partition, chiefly profiles and command-line parameters. This suggests that such a campaign needs little structure beyond plain shell files.

FutureBench: A Renewable Framework for Evaluating Knowledge-Context Conflict in LLMs

arXiv:2611.23601v1 Announce Type: new Abstract: Asking a model to check retrieved text can also tell it which source to trust. We introduce FutureBench, a renewable framework for evaluating knowledge-context conflict using timestamped questions, controlled passage triplets, and no-context probes. Across 3,629 questions and six models, Verify (a prompt combining checking with passage priority) raises false-answer adoption in all five models tested on pre-cutoff events, mainly by displacing answers that were correct under Baseline, which allows both passage evidence and model knowledge. On post-cutoff events, only o4-mini increases adoption (+13.7 percentage points). A four-prompt comparison on 902 post-cutoff questions finds that passage priority without checking raises adoption, while moving to Verify lowers it in all five models. Adversarial-Aware, which warns of misleading passages and prioritizes model knowledge, restores 82.5-98.9% of displaced pre-cutoff answers with under one point of change in true-passage accuracy. FutureBench supports repeated evaluation on fixed cohorts and extension to new events as models change.

GRAFT: Execution-Guided Genetic Recombination for Agent Design

arXiv:2611.23509v1 Announce Type: new Abstract: Automated agent design searches for executable architectures that improve LLM-based agent performance, with evolutionary methods using language-model crossover to recombine promising mechanisms for task decomposition, model coordination, and verification. This search, however, is typically guided by aggregate fitness alone, which hides the instance-level successes and failures behind each score. We introduce Execution Guided Genetic Recombination for Agent Design (GRAFT), which uses aggregate fitness for selection and survival and parent-specific execution evidence to guide semantic crossover with integrated mutation. Its feedback harness reuses bounded samples of successful and failed runs from fitness evaluation, withholding reference answers and requiring no additional task executions. We evaluate GRAFT across six reasoning and programming settings using open-weight optimizer and executor models. It achieves a mean test accuracy of 82.0%, exceeding the strongest baseline by 9.8% points. Controlled ablations show that adding execution evidence improves test accuracy with initialization, search rules, repair policies, and evaluation budgets held fixed. Our results support the value of parent-specific execution evidence for evolutionary agent design.

When Does Subspace Direction Matter for LoRA? Regime Analysis of the Magnitude Principle in Few-Shot Adaptation

arXiv:2611.23428v1 Announce Type: new Abstract: We show that the importance of a LoRA adapter's initial input-subspace direction is regime-dependent, and we characterize this dependence in theory and experiment. This refines the recent view that spectral initializers help through update magnitude alone. Direction matters most when labeled data are scarce relative to the representation dimension and the activation subspace used for initialization captures supervised task structure. In this regime, the adapter's row span remains sensitive to its starting point, making task-aligned initialization especially useful. A spiked-covariance perturbation analysis characterizes the finite-sample stability of the estimated subspace. A first-step analysis of LoRA training explains why the subspace's alignment with the task matters and motivates an alignment coefficient $\rho$ that closely tracks gains across classification tasks. We test this account with PivotLoRA, which initializes adapters from activation PCA at prediction-relevant positions and applies one output-preserving orthogonal pivot early in training. In few-shot BERT-base classification, PivotLoRA achieves the best average rank among 13 methods and improves over vanilla LoRA in 29 of 30 settings, even though it starts from a smaller input-factor initialization norm. Additional classification experiments extend the evaluation to 7B and 8B Llama decoders. Classification gains diminish with more labeled data, consistent with the proposed regime dependence. A targeted extension to completion-only reasoning also shows improvements on GSM8K and StrategyQA. These findings identify conditions under which initialization direction matters, with subspace-task alignment offering a low-cost diagnostic before training for classification.

From Preference Data to Personalization: Tracing Sycophancy in Large Language Models

arXiv:2611.21650v1 Announce Type: new Abstract: Sycophancy in large language models is often measured as agreement with the user, yet agreement can be appropriate when the user's claim is valid, their preference is genuine, or their experience warrants acknowledgment. We study sycophancy as inappropriate agreement: validating a claim, preference, or self-assessment when the context calls for correction, qualification, or resistance. We introduce a validity-aware sycophancy grader and SycoDrift, a persona-targeted evaluation set of 3,298 validity-annotated probes with controlled annotations for user openness and stakes. We use this framework to analyze three sources of agreement pressure: human preference data, automated judge specification, and in-context personalization. First, preference for validating responses in human datasets (Chatbot Arena, HH-RLHF, Community Alignment) is weak and concentrated where agreement is often appropriate. Second, in a selected-value, single-model experiment, emphasized judge rubrics produce consistent but modest increases relative to paired suppressed rubrics, demonstrating a plausible mechanism for training-time amplification. Third, in off-policy evaluations across 14 models, inference-time personalization increases inappropriate agreement, with the largest effect from matched user profiles placed directly before the query.

Regularized Score Centering

arXiv:2611.21791v1 Announce Type: new Abstract: Asynchronous reinforcement learning increases training throughput, but data from stale policies and numerical mismatch between training and inference can destabilize learning. To stabilize off-policy learning, score centering (SC) removes the effect of reward baselines on the expected update. Our gradient diagnostics show less extreme gradients with SC than with importance sampling (IS), but SC can still fail when the learner moves far from the behavior policy that generated the data. To limit this policy mismatch, we introduce Regularized Score Centering (RSC), which adds a reverse Kullback-Leibler (KL) penalty to encourage the learner's next-token distribution to stay close to the behavior policy's next-token distribution. RSC works with or without truncated importance sampling (TIS), which caps token importance weights. Our theoretical analysis supports this approach, showing that KL regularization improves convergence guarantees. We evaluate dense and mixture-of-experts models on mathematical reasoning and agentic tasks, including online shopping, tool use, games, and fully asynchronous software engineering. RSC with TIS substantially improves sample efficiency across all agentic benchmarks, attaining higher training rewards with the same number of sampled responses. In final evaluations, it achieves top performance across all six agentic tasks, improving by 14.3% on AppWorld and reaching 9.9 times the strongest baseline's score on BabyAI.

When Lower Harm Misleads: Auditing Concept Corrections

arXiv:2611.22084v1 Announce Type: new Abstract: An expert who corrects a concept prediction, such as a BI-RADS descriptor in breast ultrasound, expects the diagnosis to improve. Objectives that reduce the harm of such corrections are judged by harm and accuracy, yet harm also falls when a model responds less to every correction or starts from a worse uncorrected prediction, the reference for harm. We propose a correction audit, a reproducible evaluation protocol that separates these explanations: it splits each correction into harm and useful gain on identical events, separates the starting prediction from the correction response, and compares objectives with amplitude baselines matched in source harm. On three breast-ultrasound sources, worst-case correction training, evaluated from the mean-case model's start, lowers harm but gives up more useful gain in all 21 source comparisons pooled over six seeds, raising the corrected negative log-likelihood (NLL); from the same start, selecting between the objectives by harm picks the less helpful model on all three sources. On CUB, a monotonic harm penalty at its reported weight lowers path-level harm in all 80 concept sets while raising the NLL in all 80, almost entirely before any correction. At a fixed start, rescaling both responses to the same source harm reproduces the penalty's mean harm, useful gain and NLL on single-concept corrections within a prespecified margin of 0.0005 nat, although the two responses point in different directions. Lower harm therefore signals better corrections only when it is read together with useful gain, absolute risk and the start it is measured from.

Building trustworthy AI by fencing it in, not predicting it

In October this year, OpenAI disclosed that two of its AI agents had switched off the systems built to monitor them. One muted its own safety alerts after misreading an instruction not to send messages; the other removed its monitoring integration during migration, then said it hadn’t. No rule had told the agents to do…

Benchmarks Don't Talk Back: Learning Through Self-Disagreement

arXiv:2611.17563v1 Announce Type: new Abstract: A benchmark usually measures a language model alone with a question, and deployment rarely looks like that: answers get contested, and models give up correct ones under counter-reasoning that is itself wrong. We show that what fails there is not a shortfall of accuracy but a capability in its own right -- knowing when to hold a position and when to yield to a better one -- which we call dialectical competence. It stands on first-answer accuracy without following from it: a model can be made materially better at answering and no better at this, and training for accuracy alone can drift against it. Nor can it be bought at inference, by self-reflection or by sampling more. Taking a cue from the dialectical tradition in philosophy, in which productive opposition arises within a position rather than being imposed on it from outside, we hypothesize that this capability can be trained from opposition the model supplies itself. It states a position, generates the strongest case against that position -- its antithesis -- rules on whether the case stands, and is trained on that ruling, with no second model at training or at deployment. We call this learning through Self-Disagreement: reinforcement learning from verifiable rewards throughout, where the same check that scores an answer against its reference scores the ruling too, so no reward model, no preference data and no human annotation enter anywhere. The capability is trained rather than elicited: weak in the base model, it arrives with the reward, is not paid for out of first-answer accuracy, and holds against challengers the model never trained against, including ones stronger than itself. The opposition a model needs in order to improve, it turns out, need not come from outside it.

RSI-PDEbench: A Benchmark for Self-Improving Agents in PDE Solving

arXiv:2611.14949v1 Announce Type: new Abstract: Identifying and solving partial differential equations (PDEs) from sampled solution data often requires substantial expert effort in equation discovery, coefficient estimation, and numerical solver configuration, especially under noise and distribution shifts. Can large language model (LLM) agents automate this workflow accurately and efficiently? To investigate this question, we introduce \bench{}, a benchmark of 1,001 cases spanning 42 equation templates, including both classical and fractional PDEs. It provides three evaluation settings: Full evaluates performance on noise-free solution data; Noisy tests robustness to Gaussian noise with an RMS equal to 50\% of the clean-signal RMS; and Hard-OOD tests challenging out-of-distribution generalization to unseen equation templates and shifted coefficient ranges. We further introduce RSI-PDE, a recursively self-improving agent harness that shortlists candidate equations by residual loss, uses relative $L^2$ feedback to refine coefficients and numerical solvers, and applies a validation gate to policy and memory updates. We compare RSI-PDE with single-run baselines and native agent harnesses (Claude Code or Codex) across four frontier models. Across the three settings, RSI-PDE improves the pass rate averaged across the four models by 106.5--181.9\% relative to single-run baselines and by 3.0--19.6\% relative to native harnesses, while reducing total recorded model calls by 46.9\% across the 12 model--setting pairs. These results highlight the potential of validated recursive self-improvement in LLM agents to advance scientific computing.

When Is Proof-Step Verification Worth Its Cost? Operating Requirements of Certify-or-Decline Verification for Open-Weight Mathematical Reasoning

arXiv:2611.14794v1 Announce Type: new Abstract: Proof-step verification rewrites a solver's answer into typed steps and uses LLM judges to certify or decline. We evaluate this pipeline on open-weight models across GPQA, AIME-2025, and a MATH-500 subset. Certification raises answer-key-match precision by 31 percentage points on the strong GPQA configuration and 46 on AIME, while answering only 35% and 21% of attempts. On near-ceiling MATH-500, gains are at most 5 points at a loss of 13-27 points of coverage. An output-token-matched self-consistency selector obtains 82.9% precision at 35% coverage, compared with the verifier's pooled 87.6% (paired difference: +4.8 points, 95% interval [-8.2, 18.5]). A diagnostic sample reveals a distinct failure mode: 52 of 88 completed problems end in invalid proof construction without ever reaching a judge. Finally, external-LLM acceptance of rendered certified chains is 61.0-67.6%, compared with 87.6% answer-key match. These audits lack human or formal-kernel ground truth. Open-weight auditor pairs show near-zero chance-corrected agreement, which measures consistency rather than accuracy against proof truth. These results motivate separate reporting of execution failures, selected-answer precision, and independently assessed proof validity. We provide per-configuration results and evaluation details in the supplement.

INTENT: Out-of-Sample, Out-of-Time, and Cold-Start Evaluation on User Event Sequences

arXiv:2611.13069v1 Announce Type: new Abstract: Banks now train foundation models on ordered sequences of customer events such as transactions, app sessions, and authorizations. Deployed, such a model predicts forward in time, often for customers and merchants absent from training, yet each is evaluated on private data, so no two can be compared. Public benchmarks for such sequences rarely test a period the model has not seen; recommendation studies usually hold out each user's last events instead. We present INTENT (INTEraction, Novelty, and Transfer evaluation for next-action prediction), a public benchmark that scores a model in seven cells defined by whether the user, the period, and the target item appeared in training. An audit of 59 public dataset families retains five datasets; the most common defect is the lack of a calendar timestamp that orders events. No date lies within the time span of all five datasets, so each is split at its own cutoff, chosen by one rule fixed before any model is evaluated. Summed over the datasets, every cell holds at least 111,769 users. In reference results the cell determines the best model: a transformer leads on seen items after the cutoff but scores zero on new items, a variant that embeds item metadata helps mainly on MicroLens, and repeating the user's last item leads on new items in four of five datasets. This parameter-free control also shows that on three datasets most of the score lost after the cutoff reflects changed behavior rather than the model. The pipeline, the cutoff rule, and the per-cell user counts are publicly available.

One Table, Many Queries: Benchmarking Generalist Tabular Models Across Generation, Imputation, and Prediction

arXiv:2611.13001v1 Announce Type: new Abstract: Generation, imputation, and prediction query different aspects of the same tabular distribution, but are usually evaluated separately. This leaves unclear which capabilities they share, how their training objectives interact, and whether one model can perform all three well. We introduce the One Table Benchmark (\ours{}), spanning 56 datasets, 46 registered methods, and 20 capability axes under aligned data splits and explicit task contracts. We compare task ranks on aligned support, train five backbones under all seven non-empty objective combinations on five datasets with three seeds, and contrast fixed-model profiles with specialist portfolios. Rank comparisons connect imputation and prediction through numerical recovery and dependence preservation, and link generation to both through predictive utility; distributional fidelity follows different, sometimes opposing, rankings. Objective interventions show the largest pooled gains from including the recipient's own objective. Generation has the broadest coverage of positive pooled auxiliary effects, while prediction shows potential to improve imputation and synthetic-data utility, with responses that vary across backbones. Cross-addition effects reveal that objective benefits need not add and depend on the objectives already present. Fixed-model comparisons expose corresponding utility--fidelity trade-offs: some generalists surpass specialist portfolios on utility, while fidelity continues to favor specialization. Frozen foundation models show strong prediction and numerical point imputation, together with promising synthetic-data utility, highlighting their potential across all three tasks. These findings motivate unified tabular models that share useful conditional information while improving categorical reconstruction and faithful joint sampling.

Toward a Theory of Evaluability

Machine learning models, particularly large foundation models, are increasingly evaluated along multiple axes that are relevant for their usage. In addition to accuracy measures appropriate for the specific task of interest, measures capturing safety, calibration and fairness properties are often of interest. The classical approach to these problems assumes that the test set has been held out from the training process, an assumption that is increasingly harder to justify for large foundation models. In this work, we formalize a notion of evaluability without this assumption, and ask if it is…

Pdb or Not to Be: Comparing Debugging Interfaces With Pass$_{X}$@$k$

arXiv:2611.08666v1 Announce Type: new Abstract: Interactive debuggers are indispensable to human software development, yet coding agents overwhelmingly rely on edit-execute-observe loops, using common bash tools, test scripts, and print statements to debug via repeated shell commands. Stepping through a debugger provides live access to runtime state, which may enable different problem-solving capabilities. However, prior work typically introduces the debugger as an additional tool or through a new subagent, making its individual contribution difficult to isolate from the associated scaffold changes. To address this, we compare a debugger-only agent against a bash-only agent on all 500 SWE-bench Verified instances, holding the rest of the scaffold fixed. We evaluate both architectures on two models, GPT-OSS-120B and Nemotron-3-Ultra-550B, with three independent runs per architecture $\times$ model configuration. Crucially, to distinguish interface specialization from expanded coverage under repeated sampling, we introduce pass$_X$@$k$, a cross-agent pass@$k$ that provides an unbiased estimator of the probability that at least one of $k$ agents is successful. Although the bash and debugger agents solve partially disjoint subsets of the benchmark in any single pair of runs, our pass$_X$@$2$ estimator reveals that this is largely due to run-to-run variance rather than tool complementarity. Across both models, we find that two bash rollouts outperform the combination of one bash and one debugger rollout. Our findings highlight the importance of accounting for stochastic variance in evaluating agent and tool diversity, and we hope that pass$_X$@$k$ will serve as a useful framework for the community.

LOLPlay: Benchmarking Real-Time Visual Agents in League of Legends

arXiv:2611.03685v1 Announce Type: new Abstract: Visual agents need to coordinate reasoning and action while their environment evolves during inference. We introduce LOLPlay, a benchmark in native \emph{League of Legends} with 36 champion-specific tasks spanning resource management, combat, cooperation, and objectives. Agents interact through screenshots and mouse--keyboard actions; automated initialization and 82 executable subgoal checks distinguish partial progress from completion. Across 324 task evaluations of three models at three reasoning-effort settings, GPT-6 Astra achieves the highest task success rate of 25.0\% at low and medium effort, and the highest subgoal success rate of 40.3\% at medium effort. However, high effort improves neither metric over low effort for any model. For Astra, longer requests leave fewer interaction rounds, suggesting a cost to additional reasoning in real-time control. Trajectories reveal incomplete subgoal composition and actions executing after the observed scene changes, motivating agents that combine visual reasoning with timely action.