arxiv arXiv cs.CL · 3h ago

Riazi-8B: An Urdu Large Language Model for Mathematical Reasoning

Recent large language models demonstrate strong mathematical reasoning, but these gains rely heavily on English-centric resources, leaving low-resource languages like Urdu with limited capabilities. To address this gap, researchers developed Riazi-8B, an Urdu model designed specifically for multi-step mathematical problem solving. The model was created through a two-step adaptation process involving continued pre-training on Urdu Wikipedia and supervised fine-tuning on Urdu Chain-of-Thought data derived from GSM8K. Evaluation of Riazi-8B was conducted on the MGSM-Urdu benchmark against existing Urdu instruction-tuned models. The results showed consistent improvements in answer correctness, reasoning quality, response completeness, and Urdu generation compared to baselines. These findings demonstrate that combining Urdu language adaptation with reasoning-focused fine-tuning effectively extends mathematical reasoning capabilities to low-resource languages.

arxiv arXiv cs.CL · 3h ago

Constraint Tax in Open-Weight LLMs: Tool Calling Suppression Under Structured Output Constraints

This study identifies a phenomenon called Tool Suppression, where open-weight language models cease invoking tools when JSON Schema constraints are simultaneously enabled. The authors observed this behavior in a production Agent system and reproduced it through controlled experiments across multiple model families. While tool execution and schema compliance function correctly when evaluated independently, they fail under joint deployment conditions. Analysis reveals that JSON Schema constraints are compiled into grammar-based token masks, rendering tool-call tokens unreachable during decoding. To interpret these findings, the paper proposes the Constraint Priority Inversion hypothesis, suggesting schema satisfaction dominates action selection under simultaneous constraints. The authors mitigate this issue by introducing Transparent Two-Pass Execution, an inference-time strategy that decouples tool execution from response generation. This approach restores tool invocation while preserving structured output guarantees without requiring model retraining. The research highlights that evaluating capabilities separately may overlook critical reliability issues in production systems.

arxiv arXiv cs.CL · 3h ago

REVERIEMEM: Perspective-Bounded Memory for Book-Based Role-Playing Agents

Recent large language model role-playing systems often fail in long-narrative contexts due to factual overreach and stylistic monotony. Factual overreach occurs when characters access information outside their narrative perspective, while stylistic monotony flattens character voices through static profile descriptions. To address these issues, the authors propose REVERIEMEM, a three-layer memory architecture designed for book-based character agents. This system utilizes an episodic layer for first-person scene memories, a semantic layer for visibility-tagged facts, and a personality layer for situation-dependent behavioral patterns. The researchers also introduce KBF-QA, a benchmark consisting of 4,386 questions across eight novels to test knowledge boundaries. Experimental results show that REVERIEMEM improves Knowledge Boundary Fidelity by 34.6 percentage points compared to prior methods. Additionally, the model achieves approximately a 79% win rate on BOOKWORLD's five-dimension pairwise narrative protocol. These findings suggest that perspective-bounded memory effectively enhances both factual accuracy and character-grounded narrative generation.

arxiv arXiv cs.CL · 4h ago

MedGuards: Multi-Agent System for Reliable Medical Error Detection and Correction

The authors propose MedGuards, a medical safety guardrail framework designed to detect and correct errors in text generated by Large Language Models. This system treats error handling as a multi-agent in-context learning task where specialized agents separately perform detection, localization, and correction. A confidence-guided arbitration mechanism resolves disagreements among agents using reasoning traces and confidence scores without requiring additional model training. The study introduces the Keyword-Prioritized Correction Score (KPCS), a new metric that evaluates the accuracy of critical keywords within reference text. Experiments conducted across four multilingual medical datasets of clinical notes demonstrate significant improvements in performance metrics. These results highlight enhanced interpretability, robustness, and adaptability for safer LLM deployment in healthcare. The code for the MedErrBench benchmark is publicly available on GitHub.

github llama.cpp · 4h ago

llama.cpp b9786 Release Adds OpenCL Non-Contiguous Row Support

The llama.cpp project has released version b9786, introducing support for non-contiguous rows in normalization via OpenCL. This update is part of the ongoing development by the ggml-org team to enhance hardware compatibility and performance across various platforms. The release provides binaries for macOS Apple Silicon, Intel Macs, and iOS XCFrameworks. Linux users can access builds for Ubuntu x64, arm64, and s390x architectures using CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL backends. Android support is available for arm64 CPU devices, while Windows offers extensive options including CPU, CUDA 12 and 13, Vulkan, OpenVINO, SYCL, and HIP. The release also lists disabled builds for KleidiAI on macOS and openEuler platforms.

arxiv arXiv cs.CL · 4h ago

Framework Evaluates When GraphRAG and Agentic RAG Are Needed

The authors introduce a framework for evaluating and comparing regular, GraphRAG, Modular, and Agentic Retrieval-Augmented Generation (RAG) on semi-structured knowledge bases. They implement nine standardized scenarios spanning simple document retrieval to complex hybrid text-graph integration and agentic multi-step planning. A novel context engineering method is presented to address memory overflow issues in advanced RAG variants through new representations and agentic loop design. This optimization achieves a 19% to 53% reduction in token usage while efficiently managing retrievals. Further analysis reveals a retrieval-generation gap where expanded retrieval does not proportionally improve generation quality. The study suggests that current retrieval-oriented metrics may overstate the benefits of advanced retrieval techniques. These data-driven insights aim to guide the development of production-ready intelligent RAG systems.

arxiv arXiv cs.CL · 4h ago

BITEMBED: Extreme Low-Bit Framework for LLM-Based Text Embeddings

The paper introduces BITEMBED, an extreme low-bit framework designed to address the high deployment costs of LLM-based text embedders by targeting both encoding efficiency and vector storage. The method converts pretrained LLM backbones into BitNet-style encoders featuring ternary weights, quantized activations, and lightweight normalization refinement. To adapt these models for representation learning, BITEMBED employs continual contrastive pre-training followed by supervised contrastive fine-tuning. This fine-tuning process utilizes similarity-distribution distillation and attention-relation distillation from a full-precision teacher model. Beyond backbone quantization, the framework trains output embeddings to support multiple storage precisions, allowing for flexible trade-offs between performance and storage costs. Experiments on the MMTEB benchmark using Qwen3-0.6B and Gemma3-270M demonstrate that BITEMBED performs largely comparably to full-precision teacher embedders.

arxiv arXiv cs.CL · 4h ago

TRACE: Lightweight Detection of Corpus Poisoning in RAG via Token Influence Attribution

Retrieval-Augmented Generation systems face significant risks from corpus poisoning attacks that manipulate outputs through malicious documents. Existing detection methods often require auxiliary classifiers or additional LLM verification, which introduces substantial computational overhead. To address this, researchers introduced TRACE, a lightweight framework that identifies poisoning by tracing answer-related tokens via influence attribution. The system first discovers recurrent high-influence keywords across retrieved documents to flag potential threats. It then performs secondary verification to confirm the specific influence of these tokens on model predictions. Experiments conducted on three QA benchmarks and six LLMs demonstrate strong detection performance for the framework. Additionally, TRACE successfully uncovers attacker-specified target answers during the verification process.

arxiv arXiv cs.CL · 4h ago

RAS: Measuring LLM Safety Through Refusal Alignment

The authors propose SafeVec, a white-box evaluation procedure that measures LLM safety using internal representations instead of generated outputs. This method extracts layer-wise refusal directions from a safety-aligned reference model to identify stable layers where safe and unsafe behaviors are separable. It then scores target models by checking if their hidden states align with these refusal directions during unsafe prompts. The resulting metric, RAS (Refusal Alignment Score), maps this alignment to a calibrated 0-100 safety score. Experiments across Llama, Gemma, and Qwen families show RAS effectively separates aligned models from uncensored variants. Additionally, the metric tracks output-level attack success rates while being substantially faster than judge-based evaluations. These findings suggest refusal alignment offers a compact and efficient signal for white-box safety assessment.

arxiv arXiv cs.CL · 4h ago

OPERA: Aligning Open-Ended Reasoning via Objective Perplexity-based Reinforcement Learning

The OPERA framework addresses the instability of applying reinforcement learning to open-ended tasks by replacing external judge models with intrinsic rewards derived from perplexity dynamics. This approach quantifies uncertainty reduction at critical reflective states, eliminating stylistic biases and positional inconsistencies common in LLM-as-a-judge systems. During the cold-start phase, the method utilizes guiding words to synthesize diverse reasoning traces and employs perplexity-prioritized rollouts to identify logically consistent branches. This pipeline generates a large-scale dataset of 20,000 high-quality reasoning trajectories for training. Implementing OPERA on the Qwen3-8B model establishes a new state-of-the-art among open-source models. The system achieves parity with or surpasses proprietary models like Gemini2.5 and MiniMax-M2.5 in specific open-ended tasks. Empirical evaluations confirm the scalability and efficacy of this objective perplexity-based alignment strategy.

arxiv arXiv cs.CL · 4h ago

Do Encoders Suffice? A Systematic Comparison of Encoder and Decoder Safety Judges for LLM Adversarial Evaluation

This study evaluates whether fine-tuned ModernBERT encoder classifiers can serve as cost-effective alternatives to LLM-based judges for safety evaluation. The researchers benchmarked ModernBERT and Ettin against rule-based prefix matching, fine-tuned LLM classifiers, and various LLM judge methodologies. These LLM judges included strategies from StrongReject, ShieldGemma, JailbreakBench, AILuminate, SorryBench, Claude-as-a-judge, and models like LlamaGuard 3 and 4. The encoder classifiers were trained on judge-labeled data using a majority-voting label strategy and tested on a gold-standard holdout dataset. Performance was measured using F1 score, false negative rate, and precision-recall metrics across open-source adversarial datasets. Results were further analyzed by attack technique, including single-turn prompting, decomposition, escalation, and context manipulation. The findings provide guidance on when encoder classifiers can reliably replace LLM-based judges without substantial performance loss.

media Hugging Face Forums · 4h ago

Niodoo: A Local Runtime for Hidden State Steering of Frozen LLMs

Jason Van Pham has released Niodoo, a local runtime designed to steer frozen large language models through their hidden states. The project aims to correct last-step errors by injecting noise or "physics forces" during inference to break token loops. This approach allows smaller models to improve performance without fine-tuning, targeting specific failure cases like the Llama strawberry prompt benchmark. The system generates its own telemetry tags and utilizes TDA analysis to monitor internal model states for looping behavior. Van Pham developed this tool organically through months of self-directed research and red-teaming, emphasizing reproducible results from pinned hashes. The code is available on GitHub under the repository Ruffian-L/niodoo-hidden-state-steering.

media Hugging Face Forums · 4h ago

User Reports Tool and MCP Server Unavailability for Step 3.7 Flash on HuggingChat

A user on the Hugging Face forums reported that the Step 3.7 Flash model lost the ability to use tools and connect to MCP servers starting that morning. The poster expressed strong satisfaction with the model's performance, noting its high quality relative to its low resource consumption and cost. They emphasized a desire to continue using this specific model rather than switching to alternatives due to its efficiency. The user explicitly asked whether this loss of functionality is permanent and if there are any steps they can take to restore access. The post highlights community concern regarding the sudden disruption of tooling capabilities for a popular, cost-effective model.

media Hugging Face Forums · 5h ago

Prompt Format Inquiry for Training Unsloth/Phi-3.5-mini-instruct

A user seeks advice on the optimal prompt formatting strategy for training the Phi-3.5-mini-instruct model using Unsloth. The inquiry contrasts maintaining a custom text format against utilizing a standard chat template for dataset preparation. The current implementation employs a function that structures data into '### Input:' and '### Output:' sections, appending an end-of-text token. This approach processes JSON-encoded input and output fields derived from a Hugging Face Dataset object. The provided example illustrates a complex structure involving financial insights, merchant names, dates, and transaction totals. The user intends to deploy the trained model via a custom API and requests guidance on whether to retain this format or switch to a chat template.

github llama.cpp · 5h ago

llama.cpp b9785 Release with Hardened Caps Check and Multi-Platform Binaries

The llama.cpp project has released version b9785, featuring a code change to harden caps checks as detailed in pull request #24973. This update provides pre-built binaries for macOS Apple Silicon, Intel Macs, and iOS via an XCFramework, with KleidiAI support disabled on Apple Silicon. Linux distributions including Ubuntu are supported for CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL backends across x64, arm64, and s390x architectures. Android users can access arm64 CPU binaries, while Windows offers extensive options covering CPU, OpenCL Adreno, CUDA 12 and 13, Vulkan, OpenVINO, SYCL, and HIP. The release also includes builds for openEuler targeting x86 and aarch64 processors with ACL Graph support. A standalone UI package is available alongside the platform-specific releases to facilitate local model inference.

arxiv arXiv cs.CL · 5h ago

Argus Benchmark Evaluates Uncertainty Quantification Stability Across Vision-Language Models and GUI Grounding Datasets

The authors introduce Argus, a benchmark designed to evaluate post-hoc uncertainty quantification for computer-use agents that translate vision-language model predictions into executable GUI actions. The study assesses 28 open-weight methods across four VLM agents and four datasets, alongside eight closed-source methods from three vendors where internal model states are inaccessible. Key findings reveal selective transfer stability, where uncertainty rankings remain consistent across different datasets for a fixed model but degrade significantly when moving between different model classes or observable interfaces. Among open-weight options, hidden-state and density estimation techniques demonstrated the highest stability, while specific regimes favored sampling-based scores or verbalized self-assessment. Within-model ranking transfer proved strong with Spearman rho values up to 0.969, whereas cross-tier transfer to closed-source vendors averaged only +0.08. The research further indicates that conformal click regions shrink radii by 40-60 percent upon calibration but suffer coverage degradation under interface mismatch. To support regime-aware selection, the authors release per-item records, calibration splits, UQ scores, and analysis scripts.

arxiv arXiv cs.CL · 5h ago

Space-Efficient Language Generation in the Limit

This study initiates a resource-aware theory of language generation in the limit under space efficiency constraints. A learner observes an adversarial positive stream from a target language K and must output a hallucination-free hypothesis L while omitting at most Δ strings. The research focuses on DFAs with s states over an alphabet of size k as the hypothesis class for memory-bounded learners. In the exponential-space regime, the authors prove that a learner can exactly identify the target language K. Under stricter memory budgets, they present a streaming algorithm using poly(s,k) space that converges to a hypothesis with a generation gap of Δ= O(k^{2s-2}). This learned hypothesis captures every string in K of length at least 2s-1. The results are complemented by a near-matching lower bound derived from communication complexity, showing that achieving Δ≤ k^{(1-ε)s} requires k^{Ω(εs)} memory. These findings reveal a sharp transition between polynomial-space generation and exponential-space exact identification.

arxiv arXiv cs.CL · 5h ago

How Large Language Models Source Brand Reputation Across Languages and Markets

This study analyzes the citation sources used by large language models when answering questions about brands, focusing on the underlying web references rather than just the generated text. The researchers merged three Rankfor.AI datasets to examine 167,551 URL-grounded citations across 128 brands in 12 home markets and 13 languages. The analysis reveals that AI grounds brand answers overwhelmingly in third-party sources, with 85.7% of citations pointing to sites the brand does not own compared to only 14.3% for owned domains. The source base is highly concentrated and follows a Zipf law, where 80% of citations originate from approximately 18% of domains. Wikipedia emerges as the dominant reference site, being the most-cited domain in 11 of the 12 languages studied. The only exception is Lithuanian, where the business daily vz.lt slightly edges out Wikipedia with a 4.38% share. Additionally, the source mix shows market-specific variations, such as YouTube being the top cited domain for Polish national brands and HR portals supplying more citations than Polish Wikipedia.

arxiv arXiv cs.CL · 5h ago

ToolBench-X: Benchmarking Tool-Using Agents Under Unreliable Environments

The authors introduce ToolBench-X, a new benchmark designed to evaluate large language model agents under recoverable tool-environment unreliability. Unlike existing benchmarks that assume clean and stable environments, this framework injects five structured hazard types: Specification Drift, Invocation Error, Execution Failure, Output Drift, and Cross-source Conflict. The dataset contains executable multi-step tasks across diverse domains with deterministic tools and canonical final answers for automatic evaluation. Crucially, every injected instance remains solvable through valid recovery paths such as retrying, fallback, or verification. Experiments reveal a substantial reliability gap where agents performing well with reliable tools often fail under these hazards. Further analysis indicates that failures stem from limited hazard diagnosis and ineffective recovery rather than tool-use volume or inference budget. Targeted recovery hints successfully recover many failed tasks, whereas test-time scaling yields more limited gains. These findings suggest that evaluation must shift focus from function-call accuracy to task completion in unreliable environments.

arxiv arXiv cs.CL · 5h ago

SARA: Unlocking Multilingual Knowledge in Mixture-of-Experts via Semantically Anchored Routing Alignment

Sparse Mixture-of-Experts (MoE) architectures often struggle with low-resource languages due to cross-lingual routing divergence that limits expert sharing. To address this, researchers propose SARA, a framework that transfers specialized capabilities from high-resource anchor languages to low-resource ones. SARA aligns the internal routing distributions of MoE layers using a symmetric Jensen-Shannon divergence constraint rather than operating on output logits. This approach encourages mechanistic consistency in expert selection across different languages. The authors evaluated the method on two large language models across five low-resource languages and three benchmarks. Results show SARA outperforms standard instruction tuning, achieving gains of +0.8% on Qwen3-30B-A3B and +1.2% on Phi-3.5-MoE-instruct for Global-MMLU. These findings demonstrate that SARA effectively addresses performance bottlenecks in low-resource contexts.