arXiv subjects

Kshitij Sah

Publications and source records attributed to Kshitij Sah.

4 recordsLinked to original sources

arXiv2026

When Bigger Becomes Brittle: Why Compression Hurts More as Language Models Scale

Quantization and pruning reduce the cost of deploying language models, but their effects are usually reported as benchmark scores. Such scores record only the loss that remains at the output. Inside the model, compression damages some components more than others, and the rest of the network can partly compensate. The observed loss therefore reflects both the damage and this compensation, which benchmark scores cannot separate. How the two combine, and whether their balance changes with model size, remains unclear. We study this compensation for in-context retrieval, a capability carried by a small, identifiable set of attention heads. To separate the two, the expected loss from ablating these heads is decomposed exactly into the part restored by the rest of the network, attributed to the components responsible, and the part that reaches the output. This decomposition is then compared across model sizes. For damage of comparable size, smaller models restore most of the loss, whereas larger models retain most of it, because the compensating responses offset less of it as models grow. Compressing the heads produces the same pattern. Compression that is nearly harmless in small models can therefore be costly in large ones.

cs.AI↗

arXiv2026

Better and Worse with Scale: How Contextual Entrainment Diverges with Model Size

Larger language models become simultaneously better and worse at handling contextual information -- better at ignoring false claims, worse at ignoring irrelevant tokens. We formalize this apparent paradox through the first scaling laws for contextual entrainment, the tendency of models to favor tokens that appeared in context regardless of relevance. Analyzing the Cerebras-GPT (111M-13B) and Pythia (410M-12B) model families, we find entrainment follows predictable power-law scaling, but with opposite trends depending on context type: semantic contexts show decreasing entrainment with scale, while non-semantic contexts show increasing entrainment. Concretely, the largest models are four times more resistant to counterfactual misinformation than the smallest, yet simultaneously twice as prone to copying arbitrary tokens. These diverging trends, which replicate across model families, suggest that semantic filtering and mechanical copying are functionally distinct behaviors that scale in opposition -- scaling alone does not resolve context sensitivity, it reshapes it.

cs.CL↗

arXiv2026

DISSECT: Diagnosing Where Vision Ends and Language Priors Begin in Scientific VLMs

When asked to describe a molecular diagram, a Vision-Language Model correctly identifies ``a benzene ring with an -OH group.'' When asked to reason about the same image, it answers incorrectly. The model can see but it cannot think about what it sees. We term this the perception-integration gap: a failure where visual information is successfully extracted but lost during downstream reasoning, invisible to single-configuration benchmarks that conflate perception with integration under one accuracy number. To systematically expose such failures, we introduce DISSECT, a 12,000-question diagnostic benchmark spanning Chemistry (7,000) and Biology (5,000). Every question is evaluated under five input modes -- Vision+Text, Text-Only, Vision-Only, Human Oracle, and a novel Model Oracle in which the VLM first verbalizes the image and then reasons from its own description -- yielding diagnostic gaps that decompose performance into language-prior exploitation, visual extraction, perception fidelity, and integration effectiveness. Evaluating 18~VLMs, we find that: (1) Chemistry exhibits substantially lower language-prior exploitability than Biology, confirming molecular visual content as a harder test of genuine visual reasoning; (2) Open-source models consistently score higher when reasoning from their own verbalized descriptions than from raw images, exposing a systematic integration bottleneck; and (3) Closed-source models show no such gap, indicating that bridging perception and integration is the frontier separating open-source from closed-source multimodal capability. The Model Oracle protocol is both model and benchmark agnostic, applicable post-hoc to any VLM evaluation to diagnose integration failures.

cs.CV↗

arXiv2026

CAGE: Bridging the Accuracy-Aesthetics Gap in Educational Diagrams via Code-Anchored Generative Enhancement

Educational diagrams -- labeled illustrations of biological processes, chemical structures, physical systems, and mathematical concepts -- are essential cognitive tools in K-12 instruction. Yet no existing method can generate them both accurately and engagingly. Open-source diffusion models produce visually rich images but catastrophically garble text labels. Code-based generation via LLMs guarantees label correctness but yields visually flat outputs. Closed-source APIs partially bridge this gap but remain unreliable and prohibitively expensive at educational scale. We quantify this accuracy-aesthetics dilemma across all three paradigms on 400 K-12 diagram prompts, measuring both label fidelity and visual quality through complementary automated and human evaluation protocols. To resolve it, we propose CAGE (Code-Anchored Generative Enhancement): an LLM synthesizes executable code producing a structurally correct diagram, then a diffusion model conditioned on the programmatic output via ControlNet refines it into a visually polished graphic while preserving label fidelity. We also introduce EduDiagram-2K, a collection of 2,000 paired programmatic-stylized diagrams enabling this pipeline, and present proof-of-concept results and a research agenda for the multimedia community.

cs.CV↗