Searcharxiv⌕ Search

arXiv subjects

Mingjian Zhao

Publications and source records attributed to Mingjian Zhao.

3 recordsLinked to original sources

Testing Conversation Preservation Across LLM APIs: A Rule-Based Oracle for Detecting Silent Information Loss

LLM applications increasingly rely on translators and compatibility layers to move conversations across APIs. Yet a translated request can be accepted by the target API while silently losing information about instruction roles, tool interactions, error states, or generation controls. We introduce isloss, a deterministic oracle for testing conversation-information preservation across LLM API translations. It maps source and translated requests and replies to a shared intermediate representation, ConvIR, and applies explicit rules to distinguish legitimate representation changes from information loss. Unlike schema validation or reference-based differential testing, isloss requires neither a reference translation nor additional model calls. It also distinguishes avoidable loss from degradation caused by target API limitations. We evaluate isloss using specification-based fault injection, human review, and experiments on translators, gateways, and client libraries. On held-out tests, it detects 95/125 planted faults (76.0%) with 2/100 false alarms (2.0%). Across a synthetic request suite spanning six configurations of five translator codebases, 127/685 completed translations (18.5%) pass static target checks while silently changing prompt, role, or tool information in ways the target API could have preserved. In translated replies, 203/558 streams (36.4%) cause exceptions or silent corruption in at least one tested official OpenAI SDK. Targeted experiments across six models show that translation-induced changes can alter model behaviour when tasks depend on the affected information. These results establish conversation preservation as a distinct correctness requirement for reliable LLM interoperability, agent execution, and model evaluation.

cs.SE↗

Near-Zero Token Embeddings Silently Stall Training in Gated Sequence Models

Among 291 surveyed open-weight language model repositories (280 distinct embedding tables), 52.6% of untied gated and hybrid checkpoints (20/38) contain reachable input embeddings with norms below $10^{-6}$, compared with 0% of weight-tied models. Such near-zero rows can arise during pretraining when unseen rare or template tokens receive zero input gradients while AdamW weight decay shrinks untied embeddings. In tested Transformer configurations, non-initial positions or a BOS token prevent this failure mode. In gated sequence models (Mamba, Gated DeltaNet, xLSTM), a near-zero token can amplify input-embedding gradients by more than 30 orders of magnitude (up to 35). At zero input, current-token readout factors (SiLU gates or linear queries) vanish, suppressing historical-state readout. Residual RMS values can then fall below the square root of the RMSNorm floor across consecutive layers, causing Pre-RMSNorm to amplify backward gradients by up to $10^3$ per layer. When a gradient entry exceeds $1.84 \times 10^{19}$, FP32 global-norm clipping in PyTorch, Hugging Face Trainer, and FSDP2 can overflow to infinity and zero the current batch's gradients. In Telugu continual pretraining of Nemotron-3-Nano-4B, this discards updates on approximately 20% of steps and leaves triggered rows frozen. We propose two remedies. For released checkpoints, GroupClip decouples input-embedding clipping using a 64-bit or rescaled norm; its rescaled implementation eliminates update loss and heals triggered rows in one step with negligible measured end-to-end overhead. For pretraining, GateFloor ($\mathrm{SiLU}(z) + δ$, $δ\in [0.01, 0.1]$) keeps the gate positive at zero input, reducing near-zero gradients by 28 orders of magnitude in experiments up to 370M parameters, with perplexity comparable to the baseline.

cs.LG↗

Wizundry: A Cooperative Wizard of Oz Platform for Simulating Future Speech-based Interfaces with Multiple Wizards

Wizard of Oz (WoZ) as a prototyping method has been used to simulate intelligent user interfaces, particularly for speech-based systems. However, as our societies' expectations on artificial intelligence (AI) grows, the question remains whether a single Wizard is sufficient for it to simulate smarter systems and more complex interactions. Optimistic visions of 'what artificial intelligence (AI) can do' places demands on WoZ platforms to simulate smarter systems and more complex interactions. This raises the question of whether the typical approach of employing a single Wizard is sufficient. Moreover, while existing work has employed multiple Wizards in WoZ studies, a multi-Wizard approach has not been systematically studied in terms of feasibility, effectiveness, and challenges. We offer Wizundry, a real-time, web-based WoZ platform that allows multiple Wizards to collaboratively operate a speech-to-text based system remotely. We outline the design and technical specifications of our open-source platform, which we iterated over two design phases. We report on two studies in which participant-Wizards were tasked with negotiating how to cooperatively simulate an interface that can handle natural speech for dictation and text editing as well as other intelligent text processing tasks. We offer qualitative findings on the Multi-Wizard experience for Dyads and Triads of Wizards. Our findings reveal the promises and challenges of the multi-Wizard approach and open up new research questions.

cs.HC↗