arXiv subjects

Arya Fayyazi

Publications and source records attributed to Arya Fayyazi.

14 recordsLinked to original sources

arXiv2026

SafeStart: Mitigating the Overlooked Safety Cost of Active Expert Reduction in MoE LLM Serving

Mixture-of-Experts (MoE) language models send each token through a few of many expert subnetworks. Serving methods make these models faster by running even fewer experts per token, and they are validated on accuracy benchmarks. We show that these methods can weaken the model's refusal of harmful requests, and we introduce SafeStart, which largely restores refusal while keeping most of the speedup. In vLLM on one RTX A6000, keeping about two thirds of the selected experts speeds up offline decoding by 15 to 28% at batch size 8. We find that, at this setting, under a jailbreak that tells the model not to refuse, the share of harmful requests that the model answers rises by 18.4 percentage points on Qwen1.5-MoE, 9.5 on OLMoE, and 6.6 on DeepSeek-V2-Lite, the three of nine models where it rises significantly, while MMLU accuracy drops by less than two points. We discover that most of this increase arises while the model processes the prompt and generates its first few tokens. SafeStart therefore runs all experts for the prompt and the first eight generated tokens, and the reduced set of experts afterwards, through a fused kernel in vLLM. With experts removed, SafeStart keeps at least 85% of the offline decode speedup at batch sizes 8 and 32 and brings the change under this jailbreak to between -2.1 and +1.3 points for three expert-reduction methods on these three models. On 889 held-out requests, in an analysis fixed before generation, truncation raises harmful compliance by 5.8 to 21.4 points, and SafeStart leaves a rise of 0.6 to 2.6 points, non-inferior to the original model on two of the three models. Our results show that faster MoE serving need not cost safety: a serving system can keep most of the gain from running fewer experts by running all of them at the start of each response, and it should test refusal, not only accuracy, before it enables such a method.

cs.AI↗

arXiv2026

C-GRPO: Conformal Group Relative Policy Optimization

Reasoning-oriented language model training increasingly relies on multiple on-policy rollouts, but standard Group Relative Policy Optimization (GRPO) uses a fixed group size, wasting compute on easy prompts and under-exploring hard ones. Existing adaptive-budget methods alleviate this inefficiency, yet their stopping rules are heuristic and do not provide a finite-sample calibration guarantee. We propose CGRPO, which replaces the fixed rollout budget of standard GPRO with calibrated conformal thresholds over task-specific nonconformity scores, combining Adaptive Prediction Set (APS)-style calibration for closed-form reasoning, a first-success score for execution-based code, automatic miscoverage selection, and periodic re-calibration. We prove finite-sample coverage for the resulting stopping rule and show that, under monotone policy improvement, recalibration induces a non-increasing expected training budget. Across six benchmarks spanning mathematics, graduate-level science, and code generation, CGRPO reduces mean training rollouts by 49-70% relative to fixed-budget GRPO while retaining strong fixed-budget accuracy, achieves the strongest pass@1 among evaluated static and dynamic baselines on different datasets, and transfers across model families with up to 82% rollout savings.

cs.LG↗

arXiv2026

The Sharp Directions Are Against You: Curvature Analysis of Activation Steering

Activation steering adds a vector to a language model's internal representations at inference time. It is a lightweight alternative to fine-tuning for behavioral control. The standard construction, contrastive activation addition (CAA), is fragile, as on many inputs it shifts behavior in the wrong direction. To explain this fragility, we first study the geometry of the steering vector. We introduce steering Hessian, the matrix of second order derivative of the model's behavioral loss with respect to the steering vector. It captures the output behavior response sensitivity of the model under small perturbations of the steering vector along each direction. Its top eigenvectors are sharp directions: tiny perturbations along them can cause large, erratic, input-dependent behavioral swings. Bottom eigenvectors are flat directions: behavior is robust along them. Removing the sharp component of a steering vector approximately preserves its direction and magnitude (cosine > 0.974, norm reduction < 3% for typical settings) while reducing the vector's directional curvature by up to 12x (median 4x across 36 conditions). We evaluate on three model families (Qwen2.5-7B, Llama-3.1-8B, Mistral-7B), with four behavioral tasks, and 36 token-level conditions. Steering along sharp directions reverses the intended behavior in every condition tested, with anti-steer rates of 53-100%. Sharp directions account for only 1-5% of the CAA vector's squared norm. However, they drive a disproportionate share of the gap between the steered and unsteered next-token distributions during autoregressive generation. We term this one-line correction as sharp removal; it outperforms standard CAA in open-ended generation across all 25 conditions tested. A jailbreaking experiment further shows that sharp directions act as noise rather than signal: they disrupt coherent behavioral control regardless of steering polarity.

cs.AI↗

arXiv2026

The Agent Knew: Internal States Forecast Unauthorized Actions and Stop Them, and the Lies That Follow

Language-model agents sometimes act beyond their authorization and then misreport what they did. On CAPE-Bench, matched agent tasks in which one fact decides whether an action is authorized, frontier open-weight agents do both, and one behavior marks the lie: retroactive permission laundering, acting and then asking permission as if the action were pending, precedes a false report in 70% of 1081 fresh informed violations, against 8% otherwise (preregistered odds ratio 7.0); frontier agents on tau^2-bench instead fabricate the authorization before violating. A linear readout of the agent's residual stream forecasts the violation before the action (AUROC 0.94) and the false report before the report, and beats chain-of-thought monitors when the agent is pressured to hide its reasoning. Gating every proposed action on this forecast cuts violations over whole episodes in four model families, and sending repeated flags to an approver keeps authorized work (Gemma-4: violations 0.70 -> 0.09, completion 1.00; a model approver that escalates unclear cases keeps this gain with a person seeing 25% of asks), at about 10-8 of the agent's compute and 109 times less than a chain-of-thought monitor. On AgentDojo, an external prompt-injection benchmark, the same readout calibrated on 29 tasks, with the steering unchanged, cuts attack success on unseen tasks in two families (Gemma-4 11.7% -> 5.7%, gpt-oss 28.5% -> 17.5%) while keeping benign utility (Gemma-4 0.637 vs 0.637; gpt-oss 0.667 vs 0.681). Preventing the violation is also the most reliable way to prevent the lie (misreported violations 0.10 -> 0.01 pooled over three families, preregistered), whereas lie-specific fixes after a violation are weak. All tests are preregistered and hash-locked; code, data and the full record, including tests that did not pass, are at https://github.com/AryaFayyazi/Honesty.

cs.AI↗

arXiv2026

Running Fewer Loops Removes Refusal Behaviour in Recurrent-Depth Language Models

A recurrent-depth language model reuses one block of weights several times in a row. The number of repetitions, which we call loops, is chosen when the model runs, not when it is trained, and running fewer is cheaper and faster, so a deployment may choose to do it. Running fewer loops also makes these models less safe, and the loss of safety is larger than the loss of anything else we measured. We study four released Ouro checkpoints and Huginn-0125. At half of its trained depth, Ouro-1.4B-Thinking still solves 0.692 of the GSM8K maths problems, against 0.775 at full depth, keeping 89% of its accuracy. Over the same reduction, the rate at which it helps with harmful requests rises from 0.038 to 0.205, 5.3 times higher. The information is not lost: a linear classifier reads whether a request is harmful from the model's internal state after a single loop, with an ROC AUC of 0.967 against 0.986 at full depth. What the missing loops remove is the refusal that should follow. We measure the effect with four instruments, one generating no text at all, and reproduce it on a second model family. Removing a single direction from Ouro-1.4B's internal state at full depth drops its refusal rate on harmful prompts from 0.950 to 0.130, identifying what the shallow model is missing. Two run-time repairs do not help: the models' early-exit rules do not spend more loops on harmful requests, and choosing a depth per request does not recover the loss. Changing the training does. Standard safety training makes the shallow model safe only by refusing 0.865 of harmless requests, while training at randomly chosen depths reaches a separation of +0.399 between harmful and harmless refusal at one loop, against +0.118. On Huginn-0125 it keeps harmful help at or below 0.040 at every depth from 1 to 32 loops, where standard training reaches 0.194, and it costs no accuracy at full depth.

cs.AI↗

arXiv2026

Sharpen Without Search: On-Policy Distillation of Sequence-Level Power Distribution

A language model can give a correct answer more probability than any single incorrect answer and still usually sample an incorrect one, because the incorrect answers together hold more probability. The power distribution raises each complete answer's probability to a power above one and renormalizes, shifting probability toward answers the model finds most likely (sharpening). Sampling from it improves reasoning without changing parameters, but needs many scored candidates per query. We show that a model can instead be trained to produce such answers in one generation. On-policy power distillation (OPPD) runs a sequential Monte Carlo sampler in which the model being trained generates candidates and a frozen teacher's power distribution weights them; the same probabilities weight each answer in a maximum-likelihood update. Training raises single-generation accuracy by up to 23.0 points on MATH500 and 27.3 on GSM8K over the untrained model at the same temperature, and one generation scores 2.4 and 3.5 points above published power sampling with 64 candidates, recovering 94 percent of the gain that 16 candidates give the untrained model. For context, against GRPO trained with verified rewards from the same checkpoint and budget, OPPD scores 3.8, 4.0 and 5.4 points higher on MATH500, GSM8K and AIME using no reference answers; the two are complementary, and OPPD applied after GRPO adds up to 9.3 points. Trained only on mathematics, OPPD raises HumanEval accuracy by up to 5.3 points. One loss coefficient moves the sharpening exponent the model absorbs between 1.19 and 2.02, against 1.14 for ordinary on-policy distillation, and it rises mostly on the model's own answers. Gains hold across model families and sizes, including a model already trained with verified rewards, where lowering the temperature gives nothing and OPPD adds 4.4 points on MATH500. Code: https://github.com/ArminAzizi98/OPPD.

cs.LG↗

arXiv2026

Fair on the Surface? Benchmarking Hidden-Output Fairness Gaps in LLM Recommenders

Fairness audits for LLM-based recommenders have largely focused on observable outputs, implicitly assuming that stable recommendations reflect stable internal processing. We challenge this assumption with FairGap, the first benchmark to jointly evaluate recommendation fairness at two levels: observable output shift (OBS) and hidden representation shift (IBS), measured through controlled counterfactual identity probes across gender, age, and race. Their relationship is summarized via Representation-Output Alignment (ROA), with quadrant diagnostics for identifying user-level hidden-output mismatch. Applied to six open-weight LLM families across three domains, FairGap reveals pervasive hidden-output decoupling: ROA rarely exceeds 0.22, and a non-negligible user population shows stable outputs despite substantial internal shifts, a mode that output-only audits cannot detect by design. Further, activation steering that reduces IBS by up to 8x simultaneously worsens OBS, demonstrating a fundamental tension between internal and output-level fairness that existing frameworks are unequipped to diagnose.

cs.AI↗

arXiv2026

COFT: Counterfactual-Conformal Decoding for Fair Chain-of-Thought Reasoning in Large Language Models

Large language models (LLMs) can reveal and amplify societal biases during chain-of-thought (CoT) generation. We present COFT (Chain of Fair Thought), a training-free decoding method that applies token-level fairness control at decode time, with distribution-free marginal validity guarantees (under exchangeability) for any frozen causal language model. COFT operates in three stages. First, it creates a masked counterfactual prompt by replacing sensitive spans with neutral tokens. Second, it compares the factual and masked logit distributions through lightweight logit fusion to attenuate attribute-driven biases. Third, it uses dual-branch split-conformal calibration to certify per-step candidate token sets at a user-chosen risk level. We evaluate COFT across six models and multiple bias benchmarks. Our method reduces standard bias metrics by 30-55% (median 38%) while preserving task utility and language quality. Reasoning accuracies remain unchanged within run-to-run noise margins. The computational overhead is modest, equivalent to one additional cached forward pass (<=11%). COFT offers a clear, auditable path to safer CoT generation with significant bias reduction, negligible utility loss, and no requirement for retraining, auxiliary classifiers, or weight access.

cs.CL↗

arXiv2026

Proof-of-Perception: Certified Tool-Using Multimodal Reasoning with Compositional Conformal Guarantees

We present Proof-of-Perception (PoP), a tool-using framework that casts multimodal reasoning as an executable graph with explicit reliability guarantees. Each perception or logic node outputs a conformal set, yielding calibrated, stepwise uncertainty; a lightweight controller uses these certificates to allocate compute under a budget, expanding with extra tool calls only when needed and stopping early otherwise. This grounds answers in verifiable evidence, reduces error compounding and hallucinations, and enables principled accuracy-compute trade-offs. Across document, chart, and multi-image QA benchmarks, PoP improves performance and reliability over strong chain-of-thought, ReAct-style, and program-of-thought baselines while using computation more efficiently.

cs.CV↗

arXiv2025

MARCO: Hardware-Aware Neural Architecture Search for Edge Devices with Multi-Agent Reinforcement Learning and Conformal Prediction Filtering

This paper introduces MARCO (Multi-Agent Reinforcement learning with Conformal Optimization), a novel hardware-aware framework for efficient neural architecture search (NAS) targeting resource-constrained edge devices. By significantly reducing search time and maintaining accuracy under strict hardware constraints, MARCO bridges the gap between automated DNN design and CAD for edge AI deployment. MARCO's core technical contribution lies in its unique combination of multi-agent reinforcement learning (MARL) with Conformal Prediction (CP) to accelerate the hardware/software co-design process for deploying deep neural networks. Unlike conventional once-for-all (OFA) supernet approaches that require extensive pretraining, MARCO decomposes the NAS task into a hardware configuration agent (HCA) and a Quantization Agent (QA). The HCA optimizes high-level design parameters, while the QA determines per-layer bit-widths under strict memory and latency budgets using a shared reward signal within a centralized-critic, decentralized-execution (CTDE) paradigm. A key innovation is the integration of a calibrated CP surrogate model that provides statistical guarantees (with a user-defined miscoverage rate) to prune unpromising candidate architectures before incurring the high costs of partial training or hardware simulation. This early filtering drastically reduces the search space while ensuring that high-quality designs are retained with a high probability. Extensive experiments on MNIST, CIFAR-10, and CIFAR-100 demonstrate that MARCO achieves a 3-4x reduction in total search time compared to an OFA baseline while maintaining near-baseline accuracy (within 0.3%). Furthermore, MARCO also reduces inference latency. Validation on a MAX78000 evaluation board confirms that simulator trends hold in practice, with simulator estimates deviating from measured values by less than 5%.

cs.LG↗

arXiv2025

FAIR-SIGHT: Fairness Assurance in Image Recognition via Simultaneous Conformal Thresholding and Dynamic Output Repair

We introduce FAIR-SIGHT, an innovative post-hoc framework designed to ensure fairness in computer vision systems by combining conformal prediction with a dynamic output repair mechanism. Our approach calculates a fairness-aware non-conformity score that simultaneously assesses prediction errors and fairness violations. Using conformal prediction, we establish an adaptive threshold that provides rigorous finite-sample, distribution-free guarantees. When the non-conformity score for a new image exceeds the calibrated threshold, FAIR-SIGHT implements targeted corrective adjustments, such as logit shifts for classification and confidence recalibration for detection, to reduce both group and individual fairness disparities, all without the need for retraining or having access to internal model parameters. Comprehensive theoretical analysis validates our method's error control and convergence properties. At the same time, extensive empirical evaluations on benchmark datasets show that FAIR-SIGHT significantly reduces fairness disparities while preserving high predictive performance.

cs.CV↗

arXiv2025

Dynamic Co-Optimization Compiler: Leveraging Multi-Agent Reinforcement Learning for Enhanced DNN Accelerator Performance

This paper introduces a novel Dynamic Co-Optimization Compiler (DCOC), which employs an adaptive Multi-Agent Reinforcement Learning (MARL) framework to enhance the efficiency of mapping machine learning (ML) models, particularly Deep Neural Networks (DNNs), onto diverse hardware platforms. DCOC incorporates three specialized actor-critic agents within MARL, each dedicated to different optimization facets: one for hardware and two for software. This cooperative strategy results in an integrated hardware/software co-optimization approach, improving the precision and speed of DNN deployments. By focusing on high-confidence configurations, DCOC effectively reduces the search space, achieving remarkable performance over existing methods. Our results demonstrate that DCOC enhances throughput by up to 37.95% while reducing optimization time by up to 42.2% across various DNN models, outperforming current state-of-the-art frameworks.

cs.LG↗

arXiv2025

FACTER: Fairness-Aware Conformal Thresholding and Prompt Engineering for Enabling Fair LLM-Based Recommender Systems

We propose FACTER, a fairness-aware framework for LLM-based recommendation systems that integrates conformal prediction with dynamic prompt engineering. By introducing an adaptive semantic variance threshold and a violation-triggered mechanism, FACTER automatically tightens fairness constraints whenever biased patterns emerge. We further develop an adversarial prompt generator that leverages historical violations to reduce repeated demographic biases without retraining the LLM. Empirical results on MovieLens and Amazon show that FACTER substantially reduces fairness violations (up to 95.5%) while maintaining strong recommendation accuracy, revealing semantic variance as a potent proxy of bias.

cs.IR↗

arXiv2024

NeuroBlend: Towards Low-Power yet Accurate Neural Network-Based Inference Engine Blending Binary and Fixed-Point Convolutions

This paper introduces NeuroBlend, a novel neural network architecture featuring a unique building block known as the Blend module. This module incorporates binary and fixed-point convolutions in its main and skip paths, respectively. There is a judicious deployment of batch normalizations on both main and skip paths inside the Blend module and in between consecutive Blend modules. Additionally, we present a compiler and hardware architecture designed to map NeuroBlend models onto FPGA devices, aiming to minimize inference latency while maintaining high accuracy. Our NeuroBlend-20 (NeuroBlend-18) model, derived from ResNet-20 (ResNet-18) trained on CIFAR-10 (CIFAR-100), achieves 88.0\% (73.73\%) classification accuracy, outperforming state-of-the-art binary neural networks by 0.8\% (1.33\%), with an inference time of 0.38ms per image, 1.4x faster than previous FPGA implementation for BNNs. Similarly, our BlendMixer model for CIFAR-10 attains 90.6\% accuracy(1.59\% less than full precision MLPMixer), with a 3.5x reduction in model size compared to full precision MLPMixer. Furthermore, leveraging DSP blocks for 48-bit bitwise logic operations enables low-power FPGA implementation, yielding a 2.5x reduction in power consumption.

cs.AR↗