Research
Research, simplified.
Peer-reviewed papers on AI, security, and software engineering. Abstracts trimmed to the key insight.
01 — Research feed
Latest papers.
Recent AI agents can generate playable games, but these games are not necessarily engaging. Mastery is a key source of engagement, yet designing for it requires extensive human play and tuning. We propose PlayCraft, an automated framework that uses a vision-based agent to play games and draws on its detailed play experience to revise them for mastery. Starting from an initial game, PlayCraft alternates between playing and crafting. In the playing stage, a player agent watches the game screen, uses the controls, and pursues the game objective over repeated episodes while retaining memory. After each episode, it reports its progress, outcomes, decisions, and reflections. In the crafting stage, a crafter agent uses these reports to revise the game. It estimates player abilities, predicts performance at different proficiency levels, and compares those predictions with a target learning progression. It diagnoses the game and addresses the gaps by giving players meaningful opportunities to improve through better decisions and execution. We evaluate games in fresh runs with player agents across three intelligence settings, measuring learning gain and learning slope. Compared with the initial games, PlayCraft yields larger learning gains and steeper learning slopes in most games. It also outperforms the baselines on the macro-average of each metric at every intelligence setting.
Hybrid architectures have become mainstream among open-weight language models. Their global-attention and bounded-state layers maintain heterogeneous caches whose demands vary with context length and concurrency. We present TideCache, a serving system on CPU--GPU superchips with three components for allocating GPU memory among caches and managing their residency across GPU and host memory. The Transparent Cache Pool redistributes physical pages among cache components behind fixed virtual addresses, preserving cache layouts and backend interfaces. The Cache Streamer uses computation intervals provided by bounded-state layers to prefetch the next global-attention layer's prefix into GPU slots shared across global-attention layers, reducing GPU residency while overlapping layer-wise transfers with computation. Generated KV remains resident. The Request Scheduler selects resident or streamed execution and triggers capacity adjustments according to request size and queue state. Together, they let HBM released by streaming accommodate other cache components and additional requests. Implemented in SGLang and evaluated on GB300 and GH200 across seven model/platform configurations, TideCache completes up to 2.04x as many agent turns, reduces overall P50 per-token latency by 20--59%, and improves agent P50 TTFT by 1.40--3.05x compared with SGLang with HiCache.
Videos reveal how the world changes, but rarely the actions that cause those changes. Latent action learning recovers such supervision from visual transitions, yet often learns the underlying dynamics largely from scratch. We ask whether a pretrained video generator can instead provide the prior for both inverse and forward dynamics. We introduce O$^{2}$LaF, built on a simple principle: one prior, two dynamics. We initialize both models from the same pretrained video DiT and adapt them to their distinct roles: a feed-forward transition encoder with learnable action queries for inverse dynamics, and a causal, action-conditioned video generator for forward dynamics. By controlling information access and optimizing both dynamics end-to-end through a compact bottleneck, O$^{2}$LaF learns transferable actions from unlabeled videos. Across navigation and robotic manipulation, O$^{2}$LaF sets a new state of the art in few-shot action probing and enables stronger zero-shot action transfer across visual contexts and embodiments. Beyond generative video priors, we further extend our framework to predictive representation priors such as V-JEPA2, achieving competitive performance. Our findings suggest a broader paradigm for latent-action learning: repurposing pretrained video models to discover transferable actions and predict their consequences.
Frontier multimodal language agents such as GPT-6 Astra have demonstrated strong capabilities in controlling robots directly via low-level numerical actions without high-level primitives. However, existing benchmarks report only episode-level success rates, which show neither how these agents succeed nor where they fail. In this paper, we present a systematic action-level analysis of frontier agents on direct robot control across 12 manipulation benchmarks. We introduce AgentRoboCritic, an intent-grounded critique framework in which the policy agent states a local intent alongside each action, and a critic evaluates the action's outcome and failure mode against this intent. Our action-level analysis demonstrates the promising capabilities of frontier agents on direct robot control, including closed-loop recovery from local execution failures, reliable semantic grounding across diverse scenes, and strong zero-shot manipulation without task-specific adaptation. It also identifies their current limitations, including task-relevant physical execution bottlenecks such as grasping and alignment, and uneven recovery across failure modes. Finally, we validate our critic on a 400-action human-annotated benchmark and release a dataset of 12,844 annotated actions alongside a fine-tuned open-source critic, Muse-Glimmer-30B, whose failure-mode accuracy rises from 32.58% to 74.24%, enabling low-cost action-level analysis.
LLM-based agents are becoming increasingly capable at discovering, validating and patching vulnerabilities in real software, and vendors and researchers report that this is already changing how vulnerabilities are found and fixed. These claims, however, are scattered across papers, vendor reports and disclosures and are rarely comparable. We introduce a framework that rates these capabilities on an autonomy scale modeled on the levels of driving automation, and qualifies each claim by its correctness, its availability and the strength of its evidence. We use this framework to systematize the published evidence on LLM capability in vulnerability research. We find that agents discovered, validated, and patched vulnerabilities in real-world software without human intervention before the restricted frontier models of 2026, and that open-weight models subsequently attained comparable autonomy for discovery and validation. We then measure the consequences of evolving LLM security capabilities in a decade of public vulnerability records, advisories and exploitation data. In 2026 the number of disclosed vulnerabilities more than doubled, with an increasing share rated critical or high, and severe bugs that had survived a decade or more in widely used code were reported several times more often than before. Yet the time between the stages of a vulnerability's life, from report to fix and from disclosure to attack, has not shifted noticeably so far. We grade the mitigations proposed in the literature with an explicit rubric and draw guidance for each group of stakeholders from the best-supported ones.
Acquiring every feature of an instance is often costly: a measurement can carry a monetary price, a time delay, or a privacy cost. Active Feature Acquisition (AFA) addresses this by learning a policy that decides, per instance, which features to acquire under a budget. Existing AFA methods train this policy by optimizing a self-constructed utility target rather than distilling an explicit teacher. We construct this teacher explicitly. At training time the label and the unobserved feature values are available, and together they determine the teacher in closed form: an oracle that ranks every candidate by its effect on the downstream predictor's error. OracleImitator distills this ranking into a deployable policy by soft-target imitation. Across multimodal medical diagnosis and the AFABench benchmark, OracleImitator matches or outperforms existing AFA methods across the budget range, recovering 90% of the attainable accuracy gain at under a quarter of the acquisition cost. Because the teacher is an explicit algorithmic object rather than an implicit optimization target, we can also measure directly how much of its ranking is recoverable by a policy restricted to what is visible at deployment.
Robotic vision-language-action (VLA) models can execute diverse individual manipulation skills, yet they still struggle with long-horizon loco-manipulation tasks, where temporal structure, skills, and precision are tightly coupled within a monolithic model. We propose Physically Grounded Execution (PGX), a framework that offloads temporal and precision control to two classical training-free components, allowing the VLA to focus on stage-specific skill dynamics. PGX decomposes long-horizon loco-manipulation into three layers: (i) a predicate-based state machine governs task progression, using physical predicates to gate phase transitions, trigger recovery, and coordinate navigation and waiting. Shared predicate semantics connect offline demonstration segmentation, online execution, and stage-wise evaluation; (ii) VLA skills operate within active manipulation phases, with flat or stage-trained models concentrating on task-relevant actions; and (iii) world-coordinate locking (WCL) maintains spatial consistency through base station keeping, object-relative positioning, and world-anchored contact targets. We demonstrate the effectiveness of PGX with extensive experiments conducted on three simulated tasks and one real-world task using a Unitree G1 humanoid robot for long-horizon loco-manipulation, showing improvements in stage progression and task success rates. Project page is available at: https://kl-research.github.io/pgx
Self-consistency certification treats repeated sampling from a language model as sequential evidence accumulation and stops when an anytime-valid e-process crosses a threshold, guaranteeing finite-sample Type I error control by Ville's inequality. Prior constructions assume sampled rollouts are conditionally independent. We show this assumption fails in practice: reasoning streams are positively correlated through shared reasoning prefixes or across-prompt heterogeneity, changing not merely certificate efficiency but the hypothesis certified. Modeling dependence via a per-unit latent state (the random-effects / intraclass-correlation structure) induces exchangeable, equi-correlated increments with correlation $ρ$. Correlation splits the single null into a realized null about the drawn belief and a population null about average model behavior, which coincide only when $ρ=0$. We show: (i) the standard betting e-process remains anytime-valid for the realized null, not the population null; (ii) read as a population-null test, a suitable mixture has asymptotic crossing probability at least $π=P(μ(Z)>0)$, potentially exceeding nominal $ε$; (iii) the within-unit mean's effective sample size saturates at $1/ρ$, so extra within-unit rollouts cannot restore population validity; and (iv) a prompt-level e-process restores population validity, with optimal allocation at one rollout per prompt under a pure rollout budget. Simulations and Claude Haiku 4.5 rollouts confirm this regime dependence: on GSM8K ($\hatρ=0.435$, $\hatθ=+0.96$) the gap is dormant, whereas on AIME ($\hatρ=0.480$, $\hatθ=-0.15$) the single-unit rule crosses on 24% of prompts at nominal $ε=0.05$, while the prompt-level procedure declines (0.00). This correctness result identifies which reliability claim an anytime-valid certificate actually supports under dependence.
Vision-language models (VLMs) integrate visual and textual information, but their outputs do not reveal internal concept representations. Global workspace theory (GWT) proposes that information is selected from parallel brain processes, integrated in a shared workspace, and made available to specialised processes. We introduce the Multimodal Concept Lens (MC-Lens), a multimodal Jacobian lens, to study these functions in VLM concept processing. Its vision encoder and language decoder lenses map intermediate states to a shared semantic space defined by the decoder's output vocabulary to compare concept accessibility across modalities. MC-Lens supports internal concept interventions to examine the role of accessed concepts in downstream computation. (i) Accessibility and selective access. Multimodal concept information is readable in intermediate states and selectively accessible as tasks require. We extend "readout carriers" from one name token to a token set to better capture and express concept information and support counterfactual experiments. (ii) Concepts in downstream computation. Counterfactual experiments show that accessed concept information participates in downstream computation and reasoning. (iii) Development of cross-modal alignment and selective access. We develop two J-space metrics tracking the development of cross-modal concept alignment more clearly than hidden state cosine in controlled VLM training. The experiment also shows that reliable selective access develops later than concept accessibility. (iv) Shared functions across models. Across VLM families and sizes, we find shared patterns of concept accessibility, query-dependent selectivity and cross-task intervention effects despite differences in internal organisation. These findings offer an experimental basis to study GWT in multimodal artificial systems and how it can guide cross-modal alignment improvements.
A passing UVM regression shows the design agrees with the reference model on what the scoreboard checks, not that it meets the specification. Bugs that pass every test yet violate the specification, termed scoreboard escapes, survive mature testbenches. We present Verify the Verifier, to our knowledge the first framework whose LLM-based Anomaly Cross-Checker examines a passing regression's traces against the specification. Of 564 bugs inserted into four OpenTitan IPs, 68 are candidate scoreboard escapes: non-equivalent, exercised by the tests, yet passing. The framework confirmed 41 as scoreboard escapes with assertions failing on the buggy design and proven on the original.
Graphic designs are ordered compositions of RGBA layers whose separation enables selection, movement, and reuse. Existing layered models represent each layer with a full-canvas or region-sized spatial latent grid, so representation size grows with the cumulative spatial extent of the layers. We present ProtoDiffusion, a latent-diffusion framework built on ProtoVAE, a multi-layer variational autoencoder that encodes a design into one shared spatial prototype and one candidate vector per layer. Each candidate conditions a shared decoder to reconstruct its corresponding RGBA layer. For $N$ layers, the representation contains $HW+N$ tokens before backbone-specific packing, compared with $NHW$ for full-grid-per-layer representations. ProtoDiffusion jointly denoises the prototype and a fixed set of candidate slots, conditioned on an image or text, with bounding boxes as optional layout guidance. The model leaves unneeded slots empty and determines the number of layers and, in absence of boxes, their layout, requiring neither as input. The same architecture supports both image-to-design and text-to-design. ProtoVAE achieves high-fidelity reconstruction from its compact shared latent representation, reaching 40.43 dB PSNR on layer crops. On the LayerD evaluation split, image-to-design achieves RGB L1 scores of 0.040 with bounding boxes and 0.057 without, with box-guided alpha soft IoU (0.850) comparable to the strongest fine-tuned baseline (0.871). On merged images, text-to-design reaches 66.41 FID versus 86.84 for FLUX.2 [klein], achieves a 0.860 text-accuracy score, and runs $19.1\times$ faster than the same-backbone generation-then-decomposition cascade.
Does post-training change what a language model represents, or how this representation is read out into an answer? We compared linear probes with the models' own answers, analyzed with signal detection theory, in six open-weight base models and their post-trained versions. For judgments of whether German texts were written by a human or an AI, probes and word n-gram classifiers separated the classes well, but the answers did not and often ordered the texts inversely. Post-training lowered the sensitivity of the answers on average ($Δd' = -0.34$ across model families; -0.29 across three prompts). It also changed the use of a stated base rate: Base models shifted their log-odds by less than the stated change in prior odds, most post-trained models by more. In two models with stage-wise checkpoints, the change occurred during supervised fine-tuning, model stitching implicated the early layers, and from mid-depth on internal states at positions after the text, not those of the text, transferred the change between checkpoints. Our corpora cannot be shared, so we add a case reproducible from public data, the truth of negated statements. Here, post-training acted in the opposite direction: Probes decoded truth at or near ceiling in every model, the answers of several base models did not use it, and post-training repaired them; by mid-depth the repair resided at the positions after the statement; below it, the representation of the statement contributed, strongly in one of four models. In a training study, adding 2,000 correctly labeled authorship judgments of news texts to otherwise identical fine-tuning data improved the answers on our newspaper and textbook corpora and inverted them on the chats, where they followed a news-trained probe of the base model. Post-training thus changed how a judgment is read out of what a model represents, and a small training signal can set its direction.
Remark: This work is in submission at STOC'27 and thus the full manuscript has been publicly posted on the conference website since November 3rd, 2026. Inspired by the Codeword-Hash Structure problem recently introduced by Cakan (arXiV, 2026) to separate $\mathsf{QMA}$ from $\mathsf{QCIP}$, we introduce a new problem called Codeword-Hash pairs in Noise ($\mathsf{CHN}$) and prove that, in the query complexity model, it is in $\mathsf{QMA}$ but hard for the $\mathsf{QCMA}$ hierarchy $\mathsf{QCMAH} = \bigcup_{d\geq1}\mathsf{QCMA}^{\mathsf{QCMA}^{\cdots^{\mathsf{QCMA}}}}$, the classical-witness quantum-verifier analogue of $\mathsf{PH}=\bigcup_{d\ge1}\mathsf{NP}^{\mathsf{NP}^{\cdots^{\mathsf{NP}}}}$. Using this problem, we show that there exists a classical oracle ${O}$ such that $\mathsf{P}^{O}=\mathsf{BQP}^{O}=\mathsf{QCMA}^{O} =\mathsf{QCMAH}^{O} \subset \mathsf{QMA}^{O} \text{ but } \mathsf{QMA}^{O}\not\subseteq\mathsf{QCMAH}^{O},$ thus showing that quantum state or quantum proof advantage could exist even if classical computers can efficiently solve any problem that can be quantumly verified with a classical proof. Further, using our problem, we show that various quantum cryptographic primitives can exist even when $\mathsf{P} = \mathsf{QCMA}$ and thus neither quantum-computable one-way functions nor (verifiable) quantum computational advantage exist. In particular, relative to a classical oracle ${O}$ with $\mathsf{P}^{O}=\mathsf{BQP}^{O}=\mathsf{QCMA}^{O}=\mathsf{QCMAH}^{O}$, there exist pseudorandom states with classical keys and public-key encryption with copy-protected quantum decryption keys and classical public keys and ciphertexts, among other primitives, answering an open question of Kretschmer, Qian, and Tal (STOC 2025).
Multimodal image generation systems have made rapid progress and increasingly emphasize content safety. However, their safety mechanisms tend to rely on a single modality and lack joint consideration of both visual content and text prompts, particularly in instruction-guided image editing, where the uploaded image may itself carry risk. We thus introduce EditSafetyBench, a benchmark that measures system-level safety awareness of advanced image editing systems, containing 5,940 image-instruction pairs, covering safe and unsafe images across 11 risk categories and three edit intents. We then probe the benchmark dataset on four systems, including Gemini, GPT, Qwen API, and Qwen Local. Based on our evaluation, we observe a systematic text-vision safety mismatch: Gemini and GPT refuse 87.1\% and 81.6\% cases containing unsafe-intent editing prompts, while only refusing 24.2\% and 22.4\% cases with both unsafe images and neutral instructions. At the category level, most risks are considered inconsistently by different systems, e.g., GPT allows 70.4\% editing attempts of unsafe political content while Gemini only allows 34.8\%. Notably, sexual content shows the strongest visual sensitivity, likely benefiting from mature NSFW identification works. Our findings highlight the need for safety guardrails that jointly reason over both visual and textual modalities in image editing systems.
Generative engines synthesize retrieved web pages into answers, recommendations, and rankings. Generative engine optimization (GEO) services sell visibility in these answers by publishing promotional articles on user-generated-content (UGC) platforms for paying clients. Such promotion raises concerns in medical and financial domains, where misleading advice can cause harm. To study the severity of this problem, we measure one such paid-content operation across three major Chinese UGC platforms, uncovering 1,985 accounts associated with the network comprising 466,357 inferred promotional placements. We find that the network promotes products across diverse categories through a growing pool of accounts, conducting coordinated campaigns spanning months across multiple platforms while evolving its listicle content over time. Across nine generative engines from both the US and China, we find that 4.5% to 20% of retrievable cited pages are attributable to the network, depending on how strictly attribution is drawn. On eight of these engines, its clients are named in 52% of answers, against 68% for category leaders. Our work demonstrates the effectiveness of UGC GEO placement at influencing generative engine outcomes, and reveals content-level defenses as insufficient against coordinated UGC placements, where account-level and platform-level signals are necessary for proper defense.
Barrett's esophagus dysplasia grading suffers substantial inter-observer variability, particularly for low-grade dysplasia, complicating clinical decision-making. We conducted a retrospective multi-institutional diagnostic accuracy study of a foundation model-based deep learning system for dysplasia grading, using 969 H\&E-stained whole slide images from four institutions. Patch-level classifiers trained on expert annotations identified diagnostically relevant regions, followed by attribute-aware multiple instance learning for slide-level classification as non-dysplastic Barrett's esophagus (NDBE), low-grade dysplasia, or high-grade dysplasia. On a held-out test set (n = 130), macro-averaged AUC was 0.96 (95\% CI 0.93--0.98). For binary dysplasia detection, sensitivity was 98\% (80/82) and specificity 81\% (39/48), with a negative likelihood ratio of 0.030 and quadratic-weighted \kap{} of 0.83 against expert consensus. No high-grade dysplasia case was misclassified as NDBE. Performance was consistent across institutions without stain normalization, indicating that stain-robust foundation-model representations may remove the site-specific preprocessing that has constrained prior systems to institution-specific deployment.
When multimodal decision models determine actions, moderation outcomes, or quality judgments, their reliability depends on preserving semantic decisions across presentation changes, ignoring irrelevant additions, and distinguishing information consistent with task evidence from information that conflicts with it. Existing evaluations study decision-interface biases and multimodal model behavior in separate settings, leaving no common protocol to compare multimodal decision models with their backbones on practical tasks. We introduce \textbf{MM-SystemOneBias}, which evaluates 13 models on 5,323 original instances from nine public datasets and 11 task scenarios using six paired interventions: option order, option label, external judgment, image--text description, irrelevant information, and rating scale. Across seven matched decision-model--backbone pairs, decision models are more stable under option-label, irrelevant-information, and rating-scale changes in all seven pairs; six pairs also show lower option-order flip rates. They incur smaller accuracy losses under incorrect external judgments in all seven pairs and conflicting descriptions in six. In six pairs, they gain less accuracy from correct external judgments and show less favorable accuracy changes under consistent descriptions. Among pairs with unchanged semantic decisions, mean absolute changes in answer probabilities range from 4.56 to 11.55 percentage points, revealing input effects that accuracy alone misses.
Agents based on large language models (LLMs) can now delete files and send payments on behalf of their users. News media reports routinely describe failures of these agents (e.g., Replit's coding agent deleting a production database), but a report often only says what went wrong and rarely (in)validates the incident or finds its root cause. Incident catalogs collect these reports, and benchmarks record whether agents fail, but neither replicates an incident nor tests the cause its report might identify. In this paper, we present ROGUE-AI (Replication of Agentic Security Events: Audit and Investigation), a multi-agent framework that closes the above gap and rebuilds each reported incident offline together with a matched control that lacks only the blamed cause, and runs agents on both. Our results suggest that ROGUE-AI is able to reproduce the reported harm in 89% of its 108 attempted replications. Agents built on five current GPT, Claude, Gemini and DeepSeek models made the reported mistake in 18% to 33% of runs, against at most 1% without the blamed cause. Of the 72 incidents with real-world harm, 62 had no underlying attacker, and in these, agents did what their task asked for but missed an ordinary hazard, such as a database shared with the live site. Replications without an attacker caused more harm than replications of prompt injection for all five models. Defenses against prompt injection look for injected instructions, but these replications contain none. Further, a warning after each tool output that names the hazard and its consequence lowered harm for all five models, whereas a note of the same length that does not mention the hazard left most of the harm in place. Built from real incidents, the ROGUE-AI replications form a reusable audit framework that vendors can rerun before each new release.
Linear probes are the cheapest LLM lie detectors but often fail to transfer across kinds of lie. Label-free directions from honest and deceptive role-play are untested at 70B scale and on lies about knowledge, where a lie and an honest denial can be identical. We widen a published persona bank with roles that deny knowledge they have, plus honest counterparts. At each model's middle layer the resulting persona direction ranks lies across Liars' Bench better than single-dataset probes on three of its four models; it fails on Qwen, even when Qwen stays in role far more often. On Llama-3.3-70B-Instruct it beats directions from a random persona bank, but its gain over the published direction is within the range of random widenings. On lies about knowledge, persona directions mainly track a reply's wording: where lies and honest denials share their words, our three main contrasts stay inside the random-direction band on Llama and Qwen and reach only its edge on Gemma, though on Mistral one clears it. A knowledge direction contrasts questions the model can answer with ones nobody can, each followed by the same scripted denial. It separates lies from identical honest denials on two models and narrowly on a third, but not on the fourth; its other results partly depend on the benchmark's wording. On the Gender-Secret fine-tune it clears the band by a few hundredths on Llama and fails on Mistral. The sum of the two directions ranks lies better than a supervised probe at the same depth on all four models (on one, both near chance) but is level with the persona direction on Llama's average and behind that probe on held-out lies. More varied role-play questions do not help the persona direction on Liars' Bench.
Current audio-driven video diffusion models can synthesize realistic talking-head videos but generally require dozens of diffusion denoising steps over multi-second video chunks, making them unsuitable for streaming applications such as voice assistants and virtual avatars, where low-latency and online causal generation are required properties. AURA generates video chunk by chunk with a rolling KV cache, requiring only the audio of the current chunk to start generating, rather than a multi-second audio look-ahead like bidirectional models. It comes in two sizes: AURA 1.3B streams 576p video in real time on a single GPU (20.7 fps for a 16 Hz playback rate), making it suited to interactive avatar applications, and AURA 5B streams 720p video at 9.7 fps. Our two-stage distillation pipeline first aligns the causal student with the teacher, by training it to match the teacher's trajectories with fewer diffusion steps and a chunk-wise sliding window causal attention mask, then applies Distribution Matching Distillation (DMD) with backward simulation to close the remaining quality gap. On standard talking-head benchmarks, AURA 1.3B runs 23x faster than its own bidirectional teacher and reaches a quality comparable to TalkVerse at a lower resolution, and AURA 5B matches its TalkVerse teacher at 720p while running 17x faster. To support reproducibility and further research on streaming avatar generation, we release the weights of both models along with a dataset of 30K synthetic videos and their precomputed teacher trajectories.
Intersectional bias arises when a model treats people who hold several marginalised identities at once differently from what their separate identities would predict. Auditing such intersectional subgroups remains a challenge: their number grows quadratically, and there is no agreed baseline for what independent attributes would predict. Without such a baseline, many existing audits flag a subgroup when it exceeds each attribute alone, a condition that attributes acting independently already satisfy. We test this standard criterion on 150 synthetic datasets built with no interaction. With the shuffle null, it flags 66.1 false subgroups per dataset on average, and at least one in every dataset. When we plant interactions of known size into these datasets, a repaired null lets the criterion recover only 32% of them. We introduce ADDIS (Additive Intersectional Screen), which fits an additive model over all attributes, scores each subgroup by its Additive Residual Distance (ARD) from that model's prediction, and tests it against a Freedman-Lane null. On our synthetic benchmark, ADDIS flags 0.09 false subgroups per dataset and recovers 84% of the same planted interactions. We then build IndiPersona, 10,000 synthetic Indian personas with six identity attributes, and audit 13 open language models. The standard criterion flags 7.1% of the 738 tested subgroups on average; ADDIS flags 3.0%. Gender identity crossed with sexual orientation is the largest pairing in every model, and counterfactual personas confirm it and additionally reveal a caste x class interaction in all but one model that the observational test lacked the data to detect. Much of what audits report as intersectional bias can thus arise from independent single-attribute effects alone, yet a small core of interactions survives.
Large Web platforms operate large fleets of ranking models, each selecting features from a shared pool of tens of thousands. We present an end-to-end agentic system that optimizes features across the fleet under resource constraints shared across models and those set per model, such as feature pipeline cost and serving throughput (QPS), respectively. In this system, LLM agents propose hypotheses, and deterministic tools execute experiments and verify them. For each model, an LLM generates and shortlists cross features from a combinatorial candidate space, and all features are then ranked by several importance methods. Feature blending mixes these rankings into a single selection, using Pareto fronts across tasks for multi-task models. An LLM Researcher agent then tunes the feature selection in 10--25 asynchronous trials under resource caps, and a trimming step keeps the pipeline cost within budget. Across production models, the system improves normalized entropy by 0.31% on average within each model's budget, and the trimming step cuts pipeline cost by 2.3% with no regression beyond measurement noise. The optimized models were launched, and in online A/B tests they improved a primary online metric by 0.20%.
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.
Agent applications are increasingly deployed, yet agent-specific attacks pose real threats to them. Upon a security incident, organizations backtrack to identify the CEP (compromise entry point), remediate the damage, and prevent its recurrence. The main artifacts available for backtracking are trajectory logs, which record agent functionalities, and source code; app(lication)-specific function call logs are often fairly limited. However, trajectory logs are at the agent functionality abstraction layer, while the CEP candidates are function calls at the app code execution layer. Due to this semantic gap, the analyst needs to manually inspect the relevant code for each CEP candidate to infer its corresponding agent functionality. In this work, we present Origin, a backtracking tool for the analyst to submit a small number of queries to identify CEPs that existing tools cannot provide. Our approach poses two challenges: (1) define the intermediate abstraction that captures the mapping relationships between the two semantic layers; (2) determine the region-to-variable mapping for constructing the cross-layer graph defined in the intermediate abstraction. Each region is a specific group of agent functionalities. Origin addresses the challenges by (1) the cross-layer dependency graph (CLDG) that connects region-inferred functions at the semantic level to the variables in the low-level taint propagation subgraph, plus the associated query language, and (2) the region inference algorithm. Case studies on five popular agent applications across eight attack incidents show that an analyst using Origin correctly identifies all the CEPs.
Retrieving government schemes requires matching citizens' information needs to formal administrative descriptions, often across languages and terminology. We present a bilingual search system for Maharashtra government schemes and public services that combines BM25 retrieval with sentence-embedding similarity through equal-weight score fusion. The system indexes aligned English-Marathi scheme descriptions and supports multilingual, Marathi-focused, and translation-mediated semantic configurations. Evaluation on manually curated bilingual queries compares BM25, dense retrieval, and hybrid retrieval using IndicSBERT and MahaSBERT. BM25 provides a strong baseline with a mean reciprocal rank (MRR) of 0.864. Hybrid retrieval increases MRR to 0.872 with IndicSBERT and 0.879 with MahaSBERT, whereas the corresponding dense-only systems achieve 0.663 and 0.655. MahaSBERT-based fusion also increases precision from 0.781 to 0.808 while matching BM25 recall at 0.946. These results support the complementary use of lexical and semantic signals in this collection, while showing that improvements over a strong lexical baseline are modest and metric-dependent. The study provides an application-focused analysis of hybrid retrieval for bilingual public-service information access.
Task success alone does not specify which behaviors post-training should reinforce. Even successful teacher trajectories can contain unsupported assumptions, unproductive repetition, or superficial verification. We propose Skill-FT (Skill-Guided Fine-Tuning), which uses natural-language skills to suppress undesirable habits and elicit underrepresented effective behaviors during teacher rollouts, then fine-tunes a target model on the resulting trajectories. We introduce AutoMaster, an autonomous post-training system in which an LLM agent develops these interventions through trajectory analysis, skill design, data selection, and feedback from rollouts and target-model evaluation. Across Terminal, Search, and software engineering tasks with 4B and 9B target models, AutoMaster improves mean performance over Vanilla SFT on successful unguided teacher trajectories by 1.42-8.99 percentage points, with the teacher, task pool, training set size, and SFT recipe held fixed. Together with the ablations and trajectory analyses, these results suggest that behavioral shaping can make a fixed model a more effective teacher. We further find that retaining skill instructions during SFT yields higher mean performance than removing them across all six settings, despite their absence at evaluation and identical supervised outputs. These findings suggest that preserving skill context as part of a training demonstration can support better behavioral internalization.
Decision models enable fast, confidence-aware decisions by scoring candidate answers in a single forward pass, without generating text or intermediate reasoning. Yet reasoning can substantially improve decision quality, raising the question of whether its benefits can be retained without the cost of generating it. We investigate this question with \textbf{GLiDE}, a family of decision models spanning 0.4B to 27B parameters. Building on Qwen hybrid language models, GLiDE enables bidirectional interactions between the context, question, and candidate answers by removing causal masking from full-attention layers. To transfer reasoning capabilities into single-pass inference, we combine soft-target distillation from multiple teachers, including a reasoning model, with progressive post-training on diverse decision tasks. We further explore parameter-free recurrent depth to increase inference-time computation without generating tokens. Across five model sizes, GLiDE lies on the performance--parameter Pareto frontier of the Hugging Face \DI{}, with each model outperforming evaluated alternatives of equal or smaller size. \dmtwo{} (27B) achieves a hidden-set score of 71.52, with an expected calibration error of 2.0\% and a median latency of 92.9\,ms on a single GPU. Further checkpoint averaging and parameter-free recurrent depth yield GLiDE~27B, which reaches 73.37 with an expected calibration error of 2.1\%.
Commercial video-generation APIs bill by model and tier, while intermediaries could substitute cheaper models, lower tiers, or cached clips. What can a delivered file establish about its origin? We combine measurements of passive attribution and cryptographic provenance into a graded audit. We test attribution probes and zero-shot detectors under re-encoding, downscaling, and frame-rate changes, and survey C2PA and GB 45438 metadata across 19 commercial endpoints. Only authenticated provenance supports accusations: tier contradictions or duplicate signed instances under fresh-generation contracts, excluding retries and idempotent returns. Unverifiable signed-endpoint deliveries trigger original-file demands under pass-through clauses; archive-reuse and delivery-parameter screens support review only. Six-model container and bitstream accuracy falls from 83.9% on raw files to 45.3% after uniform re-encoding, using different feature sets. Subsampling real references from 24 to 3 fps raises pooled D3 AUROC from 0.845 to 0.988. Only BytePlus authenticates tier claims under our pinned policy, including 187/187 audited deliveries across three Seedance tiers; tested remuxing and re-encoding leave processed Seedance files unverifiable. The audit identifies authenticated contradictions in 100/100 raw 720p signed tier-swap cases, with zero observed false flags on the tested honest pools.
Rad-Linter checks chest X-ray reports against image-derived evidence to flag potential errors for review. A pretrained image classifier supplies scores for nine pathologies, and a 1.5B-parameter language model, fine-tuned with LoRA on 1,816 examples, checks their consistency with the report. We compare it with a keyword-and-negation rule baseline given the same reports and scores. Evaluation uses controlled report-score pairings from OpenI, with consistency labels derived from automatic report labels and score thresholds. Each counterfactual quad crosses two real reports with two real score vectors from other patients, creating two consistent and two inconsistent pairs. Every report and score vector occurs in both conditions, so either input alone permits only 50% accuracy. On 131 quads, Rad-Linter achieves 98.5% pair accuracy versus 80.5% for the rule baseline, and classifies all four pairs correctly in 94.4% of quads versus 56.5%. The checkers perform similarly on omission items, including unmentioned and negated findings; larger differences occur on the remaining item types, which require interpreting how findings are described. The same language model without fine-tuning remains near 50% accuracy. On a second set of 181 quads that also includes reports on which the automatic labelers disagree, pair accuracy is 93.4% versus 75.3%. These controlled results show that, given the same image-derived scores, a trained language model detects report-score inconsistencies more accurately than the tested rule baseline, with the advantage concentrated in interpreting how findings are worded.
Real-time task guidance is inherently flexible: when an agent guides a user through a task, multiple responses and timings can be appropriate, and sometimes the best action is to remain silent. Existing benchmarks often miss this flexibility by treating synthesized dialogues as fixed targets for what to say and when to say it. We introduce FLUID (Flexible, Evidence-Grounded Situated Task Guidance), a unified framework for modeling, evaluating, and enabling flexible interaction in multimodal task guidance. Using naturally occurring human-human guidance from WTaG and HoloAssist, FLUID grounds interventions in their supporting evidence, necessity, and timing flexibility, and uses this representation both to evaluate flexible guidance and to build a real-time agent for task guidance. Our agent achieves the strongest overall balance of response correctness, intervention precision, timely guidance, and response latency among evaluated systems, while still leaving room for improvement in avoiding speech overlap. Together, FLUID moves situated task guidance beyond rigid reference matching toward more natural and adaptive human-AI interaction. Project page: https://fluid-guidance.github.io.
02 — My publication
Published research.
Book Chapter · 2024
Enhancing AI Malware Detection Using Neural Network
with Binary Data Analysis
A peer-reviewed book chapter applying feedforward neural networks to raw binary executable data for malware classification — bypassing traditional signature-based detection methods. The approach demonstrated competitive detection accuracy without manual feature engineering, with the neural network outperforming baseline classifiers on precision, recall, and F1-score.
BibTeX citation
@inbook{sufi2024malware,
title = {Enhancing AI Malware Detection Using Neural Network
with Binary Data Analysis},
booktitle = {Proceedings of Atlantis Press},
year = {2024},
doi = {10.2991/978-94-6463-589-8_7},
url = {https://doi.org/10.2991/978-94-6463-589-8_7}
}