arXiv subjects

Deep Gandhi

Publications and source records attributed to Deep Gandhi.

15 recordsLinked to original sources

arXiv2026

How Modular Is a Frontier Mixture-of-Experts? A Pre-registered Causal Test in Which Apparent Expert Modularity Mostly Dissolves

Sparse Mixture-of-Experts (MoE) models route each token to a few of many experts, inviting the hypothesis that experts form functional modules tied to capabilities or languages. We test this causally on Command A+, a frontier open-weights MoE (218B total / 25B active; 128 experts, 8 active, +1 shared). We build a routing-mass atlas, pre-register six family-to-axis hypotheses before any intervention, and ablate each family at inference time against a size-matched random-expert null, measuring whether it selectively breaks its own axis (worst off-target effect at most one third of on-target). Crucially, we test the same families under four metrics and a held-out, independent-corpus run with bootstrap confidence intervals. Our finding is cautionary: robust functional modularity is rare and measurement-dependent. Of six pre-registered families, only one, the Arabic-language family, is a clean selective module that survives an independent corpus and a conservative statistical bar (1/6; a more permissive pre-registered point rule admits 3/6, but that count is threshold-sensitive). Every other family has a real causal effect yet fails selectivity, and its apparent modularity flips with the measurement: with the corpus, the metric, and the statistical bar. A positive control on Qwen3-30B-A3B recovers its published disjoint structure, confirming the method detects modularity when present. The verdict reproduces on the un-quantized BF16 model, ruling out a 4-bit quantization artifact. We conclude that ablation-based modularity verdicts are not safe unless the corpus, metric, and statistical bar are controlled. We release the atlas and ablation data.

cs.LG↗

arXiv2026

Can a Model Catch Its Own Hallucinations for Free?: Label-Free Doubt Signals Hold Their Own Against a Labelled Dataset for Abstention

Large language models state false facts as fluently as true ones, yet a model often "knows" internally when it is on shaky ground: the probability it assigns to its own answer tends to dip on the facts it gets wrong. The usual way to act on this, teaching a model to abstain rather than guess, requires a labelled dataset of right and wrong answers. We ask whether the model's own confidence, which is free and needs no labels, can do that job instead. We fine-tune each model (with LoRA) to answer when its frozen confidence is high and to say "I'm not sure" when it is low, using the signal alone and no correctness labels. Across six open-weights models (1B-8B, two families) on short-form factual question answering, with correctness adjudicated by an independent judge model, this label-free recipe holds its own against label-supervised abstention-tuning: at matched coverage we find no statistically detectable difference between the two. A control that drills hard examples instead of abstaining does not help, indicating the gain comes from calibration, not rote memorization. The signal's one blind spot is confidently wrong facts, which it cannot flag. A model's own doubt is thus a near-free substitute for a labelled dataset when teaching it when to abstain. Code and artifacts are available on request.

cs.CL↗

arXiv2026

Train, Retrieve, or Both? A Four-Arm Head-to-Head for Correct Statutory Citation on the Ontario Residential Tenancies Act

Self-represented tenants, landlords, and help-desk staff need to be pointed at the provision of law that actually governs a question, with a correct statutory citation. We study this task on the Ontario Residential Tenancies Act, 2006 (RTA) and its core regulation, asking the operator's question empirically: is fine-tuning enough, or is hybrid retrieval needed? We run a four-arm head-to-head on Qwen2.5-7B-Instruct (base zero-shot, LoRA SFT-only, RAG-only, and an SFT+RAG hybrid), scored on citation exact-match (section+subsection) over a small, human-verification-pending real eval set. The base model cannot cite the RTA and SFT-only mis-recalls sections; retrieval is essential and drives hallucination to zero by construction; and the SFT+RAG hybrid scores highest at 0.481 exact-match with zero hallucinated citations. Its edge comes from SFT making provision selection more robust to the higher-recall candidate sets that hurt zero-shot RAG. Notably, this cheap bge-small hybrid matches or beats a pipeline built on bigger, specialized retrieval models (a larger embedder and a cross-encoder reranker), and a larger/improved training set does not help either: strong statutory-citation performance here does not require specialized retrieval models or more data. The artifact zeroes hallucination and clears the lift-over-base bar but does not reach the aspirational 0.70 exact-match target. All results are on a small, human-verification-pending real eval set and are reported as preliminary.

cs.LG↗

arXiv2026

Judging to Improve: A De-biased VLM-as-3D-Judge Protocol for Single-Image 3D Generation

A companion study established a de-biased, cross-model VLM-as-3D-judge that reliably ranks single-image-to-3D mesh quality where cheap geometry and CLIP proxies fall short. This paper asks: can that judge's preferences specialize a strong open generator, TRELLIS, on one asset class (furniture), cheaply and without human labels? Taking the judge from ranking to optimization is where the work lives. Pushing a VLM judge into the training and evaluation loop exposes failure modes ranking never triggered, so our contribution is an optimization-grade hardening of the judge: a training judge (Qwen2.5-VL-7B) held distinct from an evaluation judge (InternVL3-8B) to break circularity; position-bias correction; and fixes for three failure modes (image overload, geometry-hiding splat renders, and reference-free judging that rewards clean-but-wrong outputs), with calibration evidence (clear-gap win-rate 0.83-1.0; base-vs-base ~0.5). Using this protocol as an independent evaluator, and working only from public models and data with lightweight parameter-efficient adaptation, we find our methods match the strong base rather than exceed it. Independent base samples carry essentially no learnable preference (0.94 order-flip rate), so signal must be engineered by quality-contrastive construction. Across six adaptation methods, two input regimes, and a severity sweep, the most targeted - conditioner repair under severe degradation - reaches parity (0.50) with the base, while no method clears the >=65% win-rate target. The result is mechanistic: clean inputs saturate the judge, flow-DIT fine-tuning washes out through the sampler, and conditioning repair is the locus that moves geometry. Win-rates are directional at n=8 objects. Matching a strong public-data base with cheap adaptation is itself informative: exceeding it needs more than lightweight PEFT on public data, and the judge protocol is reusable.

cs.LG↗

arXiv2026

Reliable Neural-Codec Text-to-Speech by ASR Self-Verification and Distillation: Near-Zero Catastrophic Failures Across Models and Codecs

Open autoregressive neural-codec text-to-speech (TTS) models sound excellent on typical inputs yet suffer stochastic catastrophic failures: on a meaningful fraction of utterances they emit silence, terminate early, or collapse into repetitive or hallucinated content. We show this failure mode is cheap to remove. Under a single format-robust metric (a catastrophic-failure rate via an ASR round-trip), best-of-N ASR self-verification drives failures to near-zero: no observed failures remain by N=2 on a standard corpus (LibriSpeech) and by N=4 on a hard prompt set. This is not an artifact of one model: the reduction replicates across four open codec-TTS systems and three neural codecs (XCodec2, SNAC, Mimi), reaching the near-zero floor by N=2 on three of the four. We then make the fix free at inference time by distilling the self-verified behaviour into the model, which recovers much of the robustness in single-shot decoding, closing ~52-58% of the failure mass on hard inputs at no test-time cost. The distillation gain concentrates where it is needed (hard inputs); on already-reliable prose there is no headroom and no detectable change. A controlled comparison adds a clean negative: offline direct preference optimization (DPO/IPO) does not beat plain supervised distillation, and an online iterative variant is promising but not statistically separable at our evaluation size. We report honestly the one model that resists (a larger Llasa where scale did not obviously help) and a rare-word capability ceiling that no self-distillation method overcomes

cs.SD↗

arXiv2026

A Cross-Model VLM-Judge Protocol for Single-Image 3D Mesh Quality (and Why Cheap Proxies Fall Short)

Single-image-to-3D generators are improving quickly, but there is no agreed, human-free way to tell whether one generated mesh is better than another. Practitioners commonly rely on cheap automatic proxies (render-space CLIP similarity and mesh geometry-validity statistics), yet how well these track perceived quality is unestablished. We make two contributions. First, we propose and validate a reproducible VLM-judge evaluation protocol: a fixed 24-view headless render rig, two independent vision-language judge families, and a mandatory position-bias correction that queries both presentation orders and keeps only order-consistent verdicts. The two judge families agree substantially with each other (Cohen's kappa = 0.66), well above the chance-agreement floor. Second, using this protocol as the reference, we show the cheap proxies do not substitute for it. Geometry validity is only a weak signal on average (because, as we show, it is bimodal) and stays below our pre-registered target, while render-CLIP is at chance. A learned Bradley-Terry head collapses onto a single manifoldness statistic (giving render-CLIP a negative weight) and matches geometry-only exactly, so learning the feature weights buys nothing. The proxy is also bimodal: it is significantly above chance on contrasts with visible geometric defects but at chance on ambiguous contrasts, consistent with geometry validity tracking the judge only when the defect is visually salient. We therefore recommend the VLM-judge protocol as a reliable, reproducible evaluator under the conditions tested (two feed-forward generators on Google Scanned Objects, with a face-drop degradation regime) and advise against geometry/CLIP proxies as optimization targets.

cs.LG↗

arXiv2026

Realizing Native INT8 Compute for Diffusion Transformers on Consumer GPUs: A Fused INT8 GEMM Kernel for Ideogram 4.0

Post-training INT8 (W8A8) quantization of diffusion transformers is widely deployed as a speed optimization, yet on consumer Ampere GPUs it is frequently slower than the FP8 and NF4 alternatives it is meant to beat. We trace this to a software artifact: the production "INT8" forward quantizes weights and activations only to immediately dequantize them back to bf16 and run a bf16 matrix multiply, never engaging the GPU's INT8 tensor cores, so the hardware's compute advantage is left entirely unrealized. We close this gap with a single fused Triton INT8 GEMM (int8xint8->int32 on Ampere tensor cores, with per-token x per-channel dequantization and bias folded into the epilogue, autotuned per GEMM shape) dropped into the Ideogram 4.0 diffusion transformer's linear layers in place of the dequantize-to-bf16 path. In the kernel, the int8xint8->int32 accumulation is bit-exact against torch._int_mm and the dequantized output matches the reference at cosine similarity 1.0 with no NaNs, running 2.8-4.2x faster than bf16 per GEMM. End to end it delivers a ~1.1x (~9-10%) speedup at 768px, and at 1024px it generates an image in 156.5 s on a single RTX 3090, faster than the single-card NF4 (164.5 s) and FP8 (172.9 s) baselines, at no measurable quality cost on these point estimates (PickScore/CLIPScore). INT8 thus goes from the slowest variant to the fastest, and 1024px becomes single-GPU feasible. The primary speed criterion (beat FP8, by ~9.5%) is comfortably met; the NF4 margin (~4.9%, single-run n=4) is within run-to-run variance we did not quantify and is best read as consistent with meeting the stretch target. We close with an honest deployment map: the win is specific to consumer Ampere, and on A100 and B200 the same kernel loses to those cards' fast native bf16/FP8 paths.

cs.LG↗

arXiv2026

Neither Parallel Nor Sequential: How DiffusionGemma Actually Commits Tokens

Open diffusion language models are marketed as parallel, non-autoregressive decoders, yet the order in which a shipped checkpoint actually commits its tokens is almost never measured. We instrument DiffusionGemma 26B, a masked discrete-diffusion mixture-of-experts model built on Gemma 4, hooking its sampler's accept step to record which canvas positions commit, when, and at what confidence. Across a 686-prompt, six-regime probe suite we find that its decoding is neither parallel nor block-autoregressive: it follows a partial left-to-right commit bias whose apparent strength depends almost entirely on the granularity at which you look. Order is weak token by token and strengthens smoothly as the analysis is coarsened, so the model's "block size" turns out to be an artifact of the measuring ruler rather than the architecture. The model commits in large simultaneous batches, leaving much of the within-batch order genuinely undefined rather than merely unobserved. The behaviour is regime-dependent: structured JSON is committed in essentially arbitrary order, and a position's commit confidence tracks correctness on mathematical reasoning but carries no signal on factual recall. Commitment is aggressive, finishing in a short late burst well inside the step budget, while task accuracy matches the model's autoregressive Gemma-4 sibling. Beyond these findings, our central contribution is methodological: measuring decoding order honestly demands handling trailing-EOS padding, within-regime confounding, commit non-monotonicity, block-size sensitivity, and large commit-batch ties, each of which can otherwise manufacture a decoding-order result that is not really there.

cs.LG↗

arXiv2026

Holding the FP8 Quality Ceiling at 8-Bit Weights and Activations: INT8 and GGUF Post-Training Quantization of Ideogram 4.0 for Consumer GPUs

We study post-training quantization (PTQ) of Ideogram 4.0, a 9.3B flow-matching diffusion transformer (DiT) that realizes classifier-free guidance with two separate-weight copies of a single-stream backbone and is conditioned by a Qwen3-VL text encoder, targeting Ampere RTX~3090 GPUs, which lack FP8 tensor cores. Because Ideogram~4.0 is trained on structured JSON captions, we evaluate every variant under schema-valid JSON prompts produced by an LLM expander built to Ideogram's published caption specification, and score them with a battery spanning human-preference (HPSv2), CLIP, and PickScore for standalone quality; PP-OCR exact-match and edit distance for text; and PSNR/SSIM/LPIPS for fidelity to the FP8 reference (the highest-precision public checkpoint) output. On a 300-prompt benchmark with paired bootstrap confidence intervals, an INT8 W8A8 recipe (per-channel weights, per-token dynamic activations, SmoothQuant, and bf16 protection of a small high-fragility layer set) is statistically indistinguishable from FP8 on CLIP and PickScore (paired CIs include zero) and within ~0.004 HPSv2, and, at its 8-bit size, is the most faithful reproduction of the FP8 output (LPIPS 0.243 vs 0.277/0.306 for the half-size 4-bit baselines; the INT8-Q4_K gap excludes zero). A GGUF Q4_K quantization reaches the same standalone quality as the published NF4 baseline at the same on-disk size, making it the Pareto choice on the quality-memory frontier. We further show that under JSON prompts all four variants reach parity on standalone quality, the variants separate on fidelity and text rendering, not on aggregate image-quality scores, and that text legibility, near-zero when the model is prompted with raw strings, reaches 55% OCR exact-match under the JSON captions it expects. We release the INT8 W8A8 and GGUF Q4_K quantized weights on Hugging Face under a gated, non-commercial license.

cs.LG↗

arXiv2026

RobustDebias: Debiasing Language Models using Distributionally Robust Optimization

Pretrained language models have been shown to exhibit biases and social stereotypes. Prior work on debiasing these models has largely focused on modifying embedding spaces during pretraining, which is not scalable for large models. Fine-tuning pretrained models on task-specific datasets can both degrade model performance and amplify biases present in the fine-tuning data. We address bias amplification during fine-tuning rather than costly pretraining, focusing on BERT models due to their widespread use in language understanding tasks. While Empirical Risk Minimization effectively optimizes downstream performance, it often amplifies social biases during fine-tuning. To counter this, we propose \textit{RobustDebias}, a novel mechanism which adapts Distributionally Robust Optimization (DRO) to debias language models during fine-tuning. Our approach debiases models across multiple demographics during MLM fine-tuning and generalizes to any dataset or task. Extensive experiments on various language models show significant bias mitigation with minimal performance impact.

cs.AI↗

arXiv2024

BraTS-PEDs: Results of the Multi-Consortium International Pediatric Brain Tumor Segmentation Challenge 2023

Pediatric central nervous system tumors are the leading cause of cancer-related deaths in children. The five-year survival rate for high-grade glioma in children is less than 20%. The development of new treatments is dependent upon multi-institutional collaborative clinical trials requiring reproducible and accurate centralized response assessment. We present the results of the BraTS-PEDs 2023 challenge, the first Brain Tumor Segmentation (BraTS) challenge focused on pediatric brain tumors. This challenge utilized data acquired from multiple international consortia dedicated to pediatric neuro-oncology and clinical trials. BraTS-PEDs 2023 aimed to evaluate volumetric segmentation algorithms for pediatric brain gliomas from magnetic resonance imaging using standardized quantitative performance evaluation metrics employed across the BraTS 2023 challenges. The top-performing AI approaches for pediatric tumor analysis included ensembles of nnU-Net and Swin UNETR, Auto3DSeg, or nnU-Net with a self-supervised framework. The BraTSPEDs 2023 challenge fostered collaboration between clinicians (neuro-oncologists, neuroradiologists) and AI/imaging scientists, promoting faster data sharing and the development of automated volumetric analysis techniques. These advancements could significantly benefit clinical trials and improve the care of children with brain tumors.

eess.IV↗

arXiv2024

The Brain Tumor Segmentation in Pediatrics (BraTS-PEDs) Challenge: Focus on Pediatrics (CBTN-CONNECT-DIPGR-ASNR-MICCAI BraTS-PEDs)

Pediatric tumors of the central nervous system are the most common cause of cancer-related death in children. The five-year survival rate for high-grade gliomas in children is less than 20%. Due to their rarity, the diagnosis of these entities is often delayed, their treatment is mainly based on historic treatment concepts, and clinical trials require multi-institutional collaborations. Here we present the CBTN-CONNECT-DIPGR-ASNR-MICCAI BraTS-PEDs challenge, focused on pediatric brain tumors with data acquired across multiple international consortia dedicated to pediatric neuro-oncology and clinical trials. The CBTN-CONNECT-DIPGR-ASNR-MICCAI BraTS-PEDs challenge brings together clinicians and AI/imaging scientists to lead to faster development of automated segmentation techniques that could benefit clinical trials, and ultimately the care of children with brain tumors.

cs.CV↗

arXiv2024

Training and Comparison of nnU-Net and DeepMedic Methods for Autosegmentation of Pediatric Brain Tumors

Brain tumors are the most common solid tumors and the leading cause of cancer-related death among children. Tumor segmentation is essential in surgical and treatment planning, and response assessment and monitoring. However, manual segmentation is time-consuming and has high inter-operator variability, underscoring the need for more efficient methods. We compared two deep learning-based 3D segmentation models, DeepMedic and nnU-Net, after training with pediatric-specific multi-institutional brain tumor data using based on multi-parametric MRI scans.Multi-parametric preoperative MRI scans of 339 pediatric patients (n=293 internal and n=46 external cohorts) with a variety of tumor subtypes, were preprocessed and manually segmented into four tumor subregions, i.e., enhancing tumor (ET), non-enhancing tumor (NET), cystic components (CC), and peritumoral edema (ED). After training, performance of the two models on internal and external test sets was evaluated using Dice scores, sensitivity, and Hausdorff distance with reference to ground truth manual segmentations. Dice score for nnU-Net internal test sets was (mean +/- SD (median)) 0.9+/-0.07 (0.94) for WT, 0.77+/-0.29 for ET, 0.66+/-0.32 for NET, 0.71+/-0.33 for CC, and 0.71+/-0.40 for ED, respectively. For DeepMedic the Dice scores were 0.82+/-0.16 for WT, 0.66+/-0.32 for ET, 0.48+/-0.27, for NET, 0.48+/-0.36 for CC, and 0.19+/-0.33 for ED, respectively. Dice scores were significantly higher for nnU-Net (p<=0.01). External validation of the trained nnU-Net model on the multi-institutional BraTS-PEDs 2023 dataset revealed high generalization capability in segmentation of whole tumor and tumor core with Dice scores of 0.87+/-0.13 (0.91) and 0.83+/-0.18 (0.89), respectively. Pediatric-specific data trained nnU-Net model is superior to DeepMedic for whole tumor and subregion segmentation of pediatric brain tumors.

eess.IV↗

arXiv2023

A Federated Approach for Hate Speech Detection

Hate speech detection has been the subject of high research attention, due to the scale of content created on social media. In spite of the attention and the sensitive nature of the task, privacy preservation in hate speech detection has remained under-studied. The majority of research has focused on centralised machine learning infrastructures which risk leaking data. In this paper, we show that using federated machine learning can help address privacy the concerns that are inherent to hate speech detection while obtaining up to 6.81% improvement in terms of F1-score.

cs.LG↗

arXiv2022

A Federated Approach to Predicting Emojis in Hindi Tweets

The use of emojis affords a visual modality to, often private, textual communication. The task of predicting emojis however provides a challenge for machine learning as emoji use tends to cluster into the frequently used and the rarely used emojis. Much of the machine learning research on emoji use has focused on high resource languages and has conceptualised the task of predicting emojis around traditional server-side machine learning approaches. However, traditional machine learning approaches for private communication can introduce privacy concerns, as these approaches require all data to be transmitted to a central storage. In this paper, we seek to address the dual concerns of emphasising high resource languages for emoji prediction and risking the privacy of people's data. We introduce a new dataset of $118$k tweets (augmented from $25$k unique tweets) for emoji prediction in Hindi, and propose a modification to the federated learning algorithm, CausalFedGSD, which aims to strike a balance between model performance and user privacy. We show that our approach obtains comparative scores with more complex centralised models while reducing the amount of data required to optimise the models and minimising risks to user privacy.

cs.LG↗