Reliable, Efficient, and Safe (Vision) Large Language Models for Deployment
We propose a full-day workshop on deploying trustworthy (vision-)large language models under real-world resource constraints. In production environments, reliability, efficiency, and safety are not independent objectives — they are jointly bounded by explicit token, compute, memory, latency, energy, and monetary budgets. This workshop introduces a “trust under budget” framework organized around three pillars:
By unifying these challenges under a single budget-aware lens, the workshop bridges the AI research and deployment communities, fostering new methods, benchmarks, and evaluation protocols for (v)LLMs that are not only capable but certifiably trustworthy within industrial constraints.
The workshop connects directly to search, recommendation, and advertising under latency and cost constraints; Web agents requiring reliable tool use and prompt-injection defenses; and content generation and moderation at platform scale. It brings together The Web Conference's academic and industry communities in information retrieval, machine learning, and scalable systems. Contributions will link model-level methods to service outcomes such as answer quality, agent reliability, moderation robustness, and serving cost, grounding trustworthy AI research in practical Web deployment.
Specific speaker confirmations to be finalized. (Listed in no particular order.)
Large language models increasingly support search, recommendation, content generation, and autonomous Web agents. Deploying these systems requires more than benchmark accuracy: answers must remain faithful, agents must resist manipulation, and safety checks must fit within serving budgets. As adoption grows, token limits, GPU memory, latency, energy, and cost make reliability, efficiency, and safety inseparable deployment concerns.
This workshop studies these interactions through a “trust under budget” framework. Truncated reasoning can reduce correctness, compression can weaken refusal behavior, and runtime guardrails consume latency. These trade-offs motivate joint evaluation under explicit resource constraints. We welcome theoretical methods, benchmarks, and industrial case studies across three themes:
Maintain correctness, faithfulness, and stability with limited tokens and compute:
Meet latency, memory, throughput, and cost requirements:
Preserve alignment and robustness after acceleration:
We invite regular papers (up to 8 pages) and short / position papers (up to 4 pages), excluding references, formatted in the ACM style. All submissions receive double-blind peer review via OpenReview. All submissions must be original and not under review elsewhere. Accepted papers will receive poster slots; six will also be selected for contributed talks.
| Workshop Papers Submission | 04 January 2027 |
| Workshop Papers Notification | 01 February 2027 |
| Final Camera-Ready Papers Deadline | 16 February 2027 |
| Workshop Date | 10–14 May 2027 (TBD within conference dates) |
All deadlines are 11:59 PM Anywhere on Earth (AoE). Dates will be finalized and announced soon.
To recognize the strongest contributions, the workshop will present a Best Paper Award and a Best Student Paper Award. Awards are selected by the Program Committee based on technical quality, novelty, and relevance to trustworthy, efficient, and safe (v)LLM deployment under real-world resource budgets. Winners will be invited to present their work in a dedicated in-person oral session.
Awarded to the submission that makes the most significant and well-executed contribution across the workshop’s three themes.
Awarded to the strongest submission whose first author is a student at the time of submission.
Winners will be announced at the workshop. In-person presentation is required to receive the award. Recipients will be listed here after the event.
The Trust Under Budget Language Models (TUBLMs) Rising Star Award recognizes early-career researchers whose work advances reliable, efficient, and safe deployment of (vision-)large language models under explicit resource constraints. Awardees will receive a certificate and be invited to deliver an in-person spotlight talk at the workshop.
Modern Web applications increasingly rely on large models that must operate within strict token, compute, memory, latency, energy, and cost budgets. This award highlights early-career researchers who make such systems more reliable, efficient, and safe — and who connect principled methods to real deployment outcomes such as answer quality, agent reliability, moderation robustness, and serving cost. We warmly encourage applications from researchers of underrepresented groups in the AI, ML, Web, and systems communities.
| Application Materials Due | TBD |
| Recommendation Letters Due | TBD |
| Awards Announced | TBD (at the workshop) |
All deadlines are 11:59 PM Anywhere on Earth (AoE). The application portal and dates will be announced soon.
The full-day workshop features six invited talks (30 min each, including Q&A), six contributed talks (15 min each), and a poster session — split across a morning and an afternoon session.
| Time | Programme |
|---|---|
| 09:00–09:05 | Welcome remarks (Ivor W. Tsang) |
| Morning Session | |
| 09:05–09:35 | Invited Talk 1 |
| 09:35–09:50 | Contributed Talk 1 |
| 09:50–10:20 | Invited Talk 2 |
| 10:20–10:35 | Contributed Talk 2 |
| 10:35–11:05 | Invited Talk 3 |
| 11:05–11:20 | Contributed Talk 3 |
| 11:20–12:20 | Lunch |
| 12:20–13:30 | Poster Session |
| Afternoon Session | |
| 13:30–14:00 | Invited Talk 4 |
| 14:00–14:15 | Contributed Talk 4 |
| 14:15–14:45 | Invited Talk 5 |
| 14:45–15:00 | Contributed Talk 5 |
| 15:00–15:30 | Coffee Break |
| 15:30–16:00 | Invited Talk 6 |
| 16:00–16:15 | Contributed Talk 6 |
| 16:15–16:20 | Closing remarks |
For any questions about the workshop, submissions, or awards, please contact the Organizing Committee:
Email: www_tub_committe@163.com