LLMs
3024
265
The Right Information Extraction Pipeline Depends on the Document: Accuracy-Energy Trade-offs for Small, Local Models
Whether an information extraction pipeline should process page images or parsed text depends on the document, and the answer flips across the layout spectrum. We study this trade-off under a constraint that rules out (closed) cloud services: privacy-sensitive documents processed on-premise by small (8B parameter) text-only and vision--language models, evaluated on both accuracy and energy over a design space spanning input representation, model family, and inference configuration. Benchmarking on the near-plain-text Kleister-NDA contracts and the layout-rich VRDU forms, we find that batching is the dominant energy lever, cutting energy per page by 38-85% at no cost in accuracy, while FP8 quantization saves 27-32% when requests are served one at a time but less than 1mWh per page (9-19%) once batching is applied. Preprocessing dominates what remains: neural OCR costs 17more energy per page than classical OCR and never reaches the Pareto frontier. Which representation wins flips with the type of document: vision--language models on layout-rich documents and small text-only models with a cheap parser on near-plain text, where they are both more accurate and cheaper than any vision--language configuration. Our work yields concrete guidelines for energy-efficient, privacy-compliant local information extraction.
AgentXploit: Autonomous Repository-to-Runtime Red-Teaming for AI Agents
AI agents combine language models with external data and tools that can modify files, call APIs, or execute code. Security failures can arise when adversarial content changes an agent's tool use or when the surrounding software contains vulnerabilities such as path traversal or command injection. We study authorized white-box pre-deployment auditing, where the auditor has access to the target repository and a controlled runtime, but successful attacks must still act through the task-defined attacker interface and be confirmed by an external verifier. We present AgentXploit, a two-role auditing system that separates repository-level attack-path discovery from runtime exploitation. The Analyzer Agent traces attacker-controlled inputs to sensitive operations and records code-supported candidate attack paths; the Exploiter Agent turns these paths into concrete attacks and revises them using runtime feedback. We also introduce AgentXploit-Bench, containing 72 reproducible vulnerabilities across 12 open-source AI-agent systems and frameworks. Across three runs, AgentXploit reaches 59.3% end-to-end success, compared with 38.4% for Codex. Under a token-budget-matched comparison, Codex reaches 46.3%. On AgentDojo, where injection points are provided, the Exploiter Agent reaches 79.2% attack success versus 52.7% for AgentVigil. These results highlight repository discovery and runtime exploitation as distinct challenges in end-to-end agent security auditing.
Benchmarking Attention for Tabular Foundation Models
Tabular in-context learners such as TabPFN, Mitra, or ConTextTab rely on alternating row and column attention over 2D sequences of latent embeddings. These attention patterns differ markedly from the one-dimensional case in language models: row attention involves longer sequences while column attention operates on much shorter ones, and the strided memory layout of tabular data makes producing contiguous tensors costly. Moreover, the hidden dimensions used in current models are small compared to recent language models. Yet efficient attention has been studied mostly for one-dimensional sequences, leaving the two-dimensional tabular setting unexplored. To this end, we create a reproducible benchmarking setup and study the unique characteristics of tabular attention across several backends -- Torch SDPA (efficient and cuDNN), FlashAttention-2/3/4, and the inference-only backends vLLM and SageAttention -- measuring forward and backward throughput across realistic tabular shapes on three GPU generations (A100, H100, B200). We find that the optimal backend choice differs between column and row attention and varies across hardware as well as model specifics: While the FlashAttention implementations tailored for each GPU generation perform overall best, they are at times outperformed by CuDNN in the case of column attention at longer sequences with cross-over points depending on the head dimension. Among inference-only backends, SageAttention performs well for row attention and large sequences beyond 16\,k rows. Our reproducible benchmark lays the foundation for future improvements to table-native attention. The self-contained benchmarking and evaluation code is openly available at: https://github.com/SAP-samples/tabular-attention-benchmark
Beyond Approved Actions: Runtime Validation of Persistent Outcomes in Agent Workflows
Large language model agents increasingly act on software systems, no longer merely generating text but also changing databases and online services. However, an approved database update may succeed yet leave an unapproved notification because execution can produce persistent effects beyond the requested change. Current safeguards can approve an action or record its aftermath, but without checking the persistent result before continuation, an unapproved outcome can be accepted as success and propagated to later steps. We present EffectMatch, a runtime that collects persistent changes within a controlled execution boundary and compares them with what the application approved for the current state and execution. The comparison governs commit and dependent execution. In comparative evaluation on 206 public business tasks, EffectMatch preserved all clean executions and prevented all tested incorrect commits. Six 20-run ablations exposed the failure caused by each removed mechanism, while 80 task-topology cases preserved truthful handoffs and blocked invalid continuation. Together, these results show that EffectMatch blocks the silent acceptance and downstream propagation of persistent outcomes inconsistent with application approval.
Softmax Reparameterization for Output-Head Quantization
Large vocabularies make output heads a substantial inference cost in small language models. We propose softmax reparameterization, a post-training method that selects a functionally equivalent output head before quantization. The method subtracts a scalar multiple of the vocabulary-row mean from every output row and selects the coefficient by validation KL separately for RTN, activation-weighted MSE, and full-Hessian GPTQ. This one-dimensional search includes the original head and fixed mean-centering, preserves the full-precision softmax distribution, and leaves the trained decoder unchanged; a rank-one correction handles nonlinear logit paths such as soft-capping. Across seven heads, W4 gains concentrate where baseline quantization substantially distorts predictions: on Phi-4-mini, AW-MSE KL falls from 0.936 to 0.256. The gains survive stronger GPTQ calibration and remain complementary to exact per-channel scaling and affine quantization. Across four heads and three W4 quantizers, frozen WikiText-selected coefficients also transfer to C4 and OpenWebMath, outperforming mean-centering in all 18 comparisons where the frozen coefficient differs from 1 and matching it in the remaining six. At W2, used as a compression stress test, benefits broaden across nearly the full model--quantizer matrix. Matched residual analysis shows that improved fidelity can accompany greater logit reconstruction error while reducing the residual's Fisher-weighted cost. For shift-compatible heads, reparameterization adds no inference operation and preserves packed W4 execution: with the decoder held in BF16, quantizing the Phi output head reduces batch-one generation latency by 10.8% relative to the BF16-head baseline.
Resource-Optimized and Energy-Aware Agentic AI Framework Anchored on Blockchain for Secure Software Supply Chains
This paper proposes a blockchain-backed agentic security framework designed to safeguard the complete software development lifecycle (SDLC) while also securing the agentic AI components responsible for monitoring it. The framework coordinates a set of specialised security agents, covering source integrity, dependency and SBOM analysis, CI configura tion auditing, artifact verification, and runtime policy evaluation, each supported by a large language model (LLM) that interprets artefacts, reasons over tool outputs, and produces structured security reports. To ensure agent trustworthiness, every agent generates a cryptographically signed attestation that is recorded in a permissioned blockchain via smart contracts, including an agent registry, an immutable attestation log, and an enforceable release-policy module. Communication among agents and with blockchain nodes is secured using a consortium-operated certificate authority, ensuring authenticated and tamper-resistant interactions. A detailed use-case and sequence flow demonstrate how a source code security agent performs analysis, anchors its attestation on-chain, and triggers a verifiable allow/block deployment decision. The proposed framework of fers decentralised integrity transparent provenance, uninterrupted security assurance and a generalisable architecture to incorporate the agentic AI into the modern software supply chain security.
MoSAR: Mixture of Semantic Attention Regimes for Learning Adaptive and Approximable Attention Geometries
The quadratic complexity of dense self-attention remains a central bottleneck for long-context language modeling. Many efficient alternatives address this cost by deciding in advance where attention should be sparse or local. We argue that attention approximation should instead be approached as a geometric problem, with the relevant interaction geometry learned from data: natural-language dependencies are input-dependent and difficult to prescribe in advance, so the model should learn where positional relevance can decay and where broader interactions must be preserved. We introduce Mixture of Semantic Attention Regimes (MoSAR), which learns such an adaptive, controlled-decay geometry over query--key interactions. Input-conditioned query and key routers, applied after positional encoding, select mixtures over short, medium, and global regimes, inducing a continuous distance-dependent attention field rather than a fixed sparsity pattern. This geometry is learned during training and can subsequently be discretized through top-1 routing. In controlled pre-training experiments with matched 500M-parameter models, MoSAR learns a substantially lower-reach attention geometry without degrading language-modeling quality, improving perplexity over dense RoPE at the training context length. Under length extrapolation, MoSAR achieves the best perplexity among all evaluated variants, including strong baselines such as ALiBi. Moreover, the learned geometry remains stable under deterministic top-1 discretization, suggesting that it is not only adaptive, but also amenable to low-cost approximation at inference time.
RupeeBias: Auditing Demographic Bias in Indian Economic Guidance from Large Language Models
Individuals turn to large language models (LLMs) for guidance across a wide range of economic tasks, from comparing loan options and planning savings to deciding what raise to ask for or how much to charge for their services. LLMs are known to reproduce social biases, and biased economic guidance may influence what users believe they are worth, what they ask for, and what they ultimately accept. This risk is especially salient in India, where economic outcomes are shaped by demographic categories such as caste and urban-rural location. Existing LLM bias benchmarks, however, are largely designed around Western demographic categories and therefore miss key axes of economic disparity in the Indian context. We introduce RupeeBias, a benchmark for auditing demographic bias in LLM-generated economic guidance across Indian economic settings. RupeeBias consists of 39,150 prompts spanning four use cases: salary estimation, salary increment estimation, counter-offer recommendation, and service pricing recommendation. The benchmark follows a single-attribute counterfactual design, holding the description of the user's qualifications, experience, or service offering fixed while varying one demographic identifier at a time. RupeeBias covers 87 India-specific demographic identifiers across six axes: caste, religion, regional identity, gender, disability, and urban-rural location, with all prompts constructed in both English and Hinglish. We evaluate nine LLMs on RupeeBias and find systematic demographic disparities across all six axes. For otherwise identical prompts that differ only in demographic identifier, LLM-generated economic outputs differ by 20.2% on average. We publicly release RupeeBias to support future research on demographic bias in LLM-generated economic guidance across India-specific demographic and economic contexts.
Acoustic-to-Text KV Compression for Full-Duplex Speech Models
Full-duplex speech language models continuously accumulate acoustic key-value (KV) states, making long-running interactions memory-intensive. During listening, the model can finish processing an audio unit before the next arrives; we term the remaining interval listening-time slack. We propose acoustic-to-text KV compression, which introduces a transcription side channel to convert incoming speech into compact textual memory within this interval. When the cache exceeds a target budget during inference, older acoustic states are evicted while transcripts and recent acoustic context remain. We train the side channel with LoRA using cross-entropy on transcription segments. To preserve listening and speaking behavior, we apply knowledge distillation to the original model's token-level output distributions at native prediction positions. On ten-minute LongSpeech sessions, our MiniCPM-o 4.5 implementation reduces peak streaming KV-cache size by 64.6% compared with the same model without eviction. The proposed method also improves transcription, temporal question answering, and summarization over the baseline. Full-Duplex-Bench evaluations further show comparable pause-handling, turn-taking, and interruption performance.
DIAL: Position-Debiased LLM Judges with Adaptive Human Preference Calibration
Large language models (LLMs) as a judge enable scalable evaluation, but their judgments can be sensitive to response order and, even after removing such position effects, can still diverge systematically from human preferences.We introduce DIAL, a unified framework that combines abundant LLM comparisons with limited human comparisons to separate judge-specific position effects, learn shared structure in position-debiased LLM preferences, and adaptively calibrate that structure toward the human preference target. Theoretically, we study three aspects of DIAL: (i) identification of latent LLM preferences, position effects, and human calibration; (ii) adaptive estimation that balances LLM anchoring against limited human evidence; and (iii) fixed-weight uncertainty quantification for the calibrated human preference. Empirically, we evaluate position debiasing and human alignment separately in controlled simulations and on three human-preference benchmarks, showing that DIAL remains robust to unbalanced response order, achieves strong human-aligned rankings with limited labels, and adapts toward human evidence when LLM information is imperfect. Our real-data study collects over 410K judgments from 21 LLM judges in both display orders, providing a resource for future studies of LLM-judge bias, heterogeneity, and human alignment.
Which Influence Are We Estimating? The Role of Counterfactual Specifications in Data Attribution
Estimating the influence of training examples on model behavior is essential for data debugging, valuation, and attribution. Existing influence estimators often produce incompatible rankings, which are commonly ascribed to approximation error. We argue that a more fundamental source of disagreement is specification mismatch: influence depends on the behavior being attributed, the intervention applied to each training example, and the counterfactual training process that maps the intervention to a model response. These choices are especially important when the target behavior requires a tractable surrogate, such as query loss, a logit, or a margin. We formalize influence as a counterfactual estimand, distinguish specification mismatch across estimands from approximation error in estimating a fixed estimand, and organize representative estimators by their implied specifications. We further derive a local decomposition that exposes how behavior signals, training signals, and counterfactual parameter responses interact. Controlled experiments show that exact estimands under different specifications can induce different rankings, whereas approximation error grows as perturbations move farther from their linearization points. Experiments on noisy label detection and LLM attribution show that specification choices significantly affect attribution quality, especially for the choice of behavior surrogate. Behavior-aligned specifications can identify target-specific training examples obscured by default loss-based or similarity-based specifications. These results establish specification analysis as a necessary first step for interpreting and comparing data influence estimators.
Accounting for Bias Enables Sustainable LLM Evaluation
LLM-as-a-judge has become the de facto standard for scalable, subjective evaluation, yet current leaderboards compensate for systematic measurement bias by running ever more comparisons, an approach that is both statistically unsound and computationally wasteful. The root cause is an incomplete measurement model, treating LLM judges as neutral, interchangeable instruments ignores documented biases like position bias, verbosity bias, judge severity, and self-enhancement, that no volume of additional data can eliminate. We propose a unified latent variable framework that jointly models pairwise and ordinal data while explicitly correcting for these confounders, recovering reliable rankings from substantially fewer comparisons. Because fitting this model costs negligible compute relative to a single round of LLM inference, bias correction is not only more statistically rigorous but also a more sustainable approach to trustworthy evaluation.
SPO: Discovering Adaptive Large Neighborhood Search Operators via Stackelberg Program Optimization
Large neighborhood search (LNS) relies critically on destroy and repair operators, whose effectiveness depends on both adaptation to the evolving LNS state and interaction between the two roles. We introduce Stackelberg Program Optimization (SPO), an LLM-based framework for discovering adaptive executable destroy-repair programs. SPO conditions operator decisions on a compact LNS state, allowing state-dependent behavior to emerge through program discovery, and organizes destroy-repair discovery as a Stackelberg interaction over program space that reflects their asymmetric dependency. Role-specific credits evaluate destroy programs as leaders and repair programs as conditional follower responses, guiding a coupled optimization process that combines LLM generator learning with population-based evolutionary search over programs. Experiments on the traveling salesperson problem and capacitated vehicle routing problem show that SPO outperforms strong baselines across a broad range of settings and generalizes beyond the discovery scale to larger instances and benchmark sets. Behavioral analyses further demonstrate state-dependent operator behavior and coupled destroy-repair improvement during discovery.
Semantic Navigation for Issue Localization in Code Repository
Repository-level issue localization aims to identify and rank the files and functions relevant to resolving a reported issue. LLM agents approach this task iteratively: they identify a set of potentially relevant locations, inspect the corresponding code, and revise their judgments about these candidates as new evidence is acquired. Existing environments, however, provide limited support for this loop: agents must search for unresolved relation targets, reconstruct entity semantics from raw source code, and revise candidates without evidential basis. To address these limitations, we present SemNav, a framework that leverages deterministic retrieval to seed a broad candidate set and an LLM agent to continually refine that set, thereby combining initial coverage with evidence-guided revision. SemNav supports this process through three key components. A Semantic Navigation Graph resolves program relations on demand through a language server, enabling direct navigation to related entities across files. Issue-conditioned Semantic Cards provide compact, source-grounded interpretations of each entity's role and relevance to the issue. A persistent Candidate Workspace records each candidate together with its evidential basis, enabling grounded verification, revision, and ranking. Across SWE-bench Lite and PLocBench, SemNav outperforms existing baselines, improving File Hit@10 from 68.33\% to 82.67\% with Gemma 4B. Component ablations and trajectory analysis support the complementary roles of all three components, while Semantic Cards reduce working-context load by 48.2\% relative to full-source reading. SemNav further ranks first on all seven evidence-quality metrics on SWE-Explore and improves downstream issue resolution from 44.00\% to 52.33\%.
Improving Visual Sensitivity of LLMs on Multimodal Machine Translation with Metric-based Loss Weighting
Multimodal Machine Translation aims to incorporate additional signal from non-textual modalities to improve translations by resolving ambiguities. While models, through multimodal fusion, are able to accept images related to the source text, they can ignore this information. Therefore, increasing their visual sensitivity remains an active research area. In this work, we introduce a training method, Metric-based Loss Weighting, that improves visual grounding of translations by increasing the loss function for tokens that benefit from the accompanying image. We identify these tokens using the Point-wise Cross-mutual Information (PCXMI) metric, which compares the model's output probabilities with and without visual context. We introduce a Congruency-based PCXMI metric and experimentally show that both metrics working in combination yield the best results. We evaluate our method by fine-tuning three pretrained Multimodal Large Language Models on the task of Image-guided Machine Translation for three language directions. Metric-based Loss Weighting outperforms other tested methods on the CoMMuTE contrastive dataset, improving accuracy by up to more than 7 percentage points compared to standard fine-tuning, while maintaining strong general translation performance.
LLM Agents Can Easily Tamper With Their Own Traces
Asynchronous monitoring, incident investigations, and compliance audits primarily rely on agent traces to reconstruct what happened. These analyses assume that LLM agents cannot tamper with their own execution traces. We show that local LLM agents such as Claude Code, Codex, Antigravity, Open Code and Grok Build fail to enforce this boundary. All tested harnesses, except Muse Code, allowed agents to delete their traces when asked, without triggering monitor guardrails. We also validate that external attackers can exploit this gap to induce trace deletion. Finally, we show that trace tampering behavior emerges naturally in frontier models, when agents try to improve their rewards. We advise practitioners to ensure trace logging happens through an independent interception mechanism outside of the agent's control, preserving trace integrity even in cases of full host compromise. Overall, our findings identify a concrete failure of trace integrity in agent infrastructure which can be used to conceal misaligned behaviors like scheming or sabotage.
SemMSA: Latent Semantic-Aided Robust Multimodal Sentiment Analysis with Incomplete Data
Recent research on Multimodal Sentiment Analysis (MSA) has focused on learning from language, visual, and acoustic modalities with incomplete data to infer human sentiment. Most studies typically compensate for missing information by reconstructing modality features or designing complicated fusion mechanisms. However, these methods still suffer from spurious generation and noisy guidance due to the lack of high-level semantic grounding in partially observed multimodal evidence. To address these issues, we propose SemMSA, a latent semantic-aided framework that constructs rich sentiment-relevant semantics with LLMs, fully integrating with all modalities via anchor-free spectral alignment. It mainly consists of Cross-modal Semantic Refinement (CSR) and Cross-modal Spectral Alignment (CSA). Specifically, CSR first adaptively extracts visual and acoustic representations by corresponding adapters to form a unified multimodal prefix with language in the frozen LLM embedding space. It then iteratively produces continuous discriminative semantic states through a token-efficient latent refinement process without decoding explicit text. Next, CSA simultaneously aligns the refined semantics with all modalities by enhancing the dominant spectral component of their kernel Gram matrix. This captures global nonlinear dependencies among all representations without relying on a predefined anchor modality. In addition, an instance-level spectral separation constraint preserves cross-sample discriminability and mitigates representation collapse. Extensive experiments on SIMS, MOSI, and MOSEI benchmarks demonstrate that SemMSA achieves state-of-the-art performance.
Coding Agents for Generalized Task and Motion Planning Problems
Task and motion planning (TAMP) problems remain difficult even with full observability and object-centric states because discrete decisions are tightly coupled to geometric, kinematic, and dynamic constraints. Generalized TAMP addresses this difficulty by exploiting regularities across problem instances to reduce planning effort on new instances. However, existing methods require substantial TAMP-specific engineering. We investigate whether coding agents can automate this process by synthesizing programs that generalize across instances. Given a task description and simulator access, each agent chooses how to interact with the environment while developing a program within a fixed synthesis budget. The program is then frozen and evaluated on unseen instances. We evaluate Claude Code (Opus 5) and Codex (GPT-5.6 Sol and GPT-6 Astra) on 28 simulated environments from KinDER and PDDLStream, with object counts beyond those evaluated in the original benchmark. Across all program synthesis methods, we evaluate 980 generated programs on 100 held-out instances each, 98,000 evaluation episodes in total. Overall, we find that coding agents are surprisingly effective at generalized TAMP: all three agent configurations outperform hand-engineered planners, one-shot generation, and an LLM-based generalized planning baseline in mean success (56% to 95% versus 47% for the planners, on the 16 environments where a planner is available). As object counts grow, the agents' programs maintain higher success than the planner, using an order of magnitude less computation per instance on average. Logs show agents using interaction to calibrate physical models, test edge cases, and refine strategies. We release all code, including the full prompts given to the agents. These findings suggest that coding agents are a strong baseline for generalized TAMP.