Publications
Research
Hill Research publishes at top-tier venues in AI, machine learning, and medical informatics. Our research is the foundation that TriClick is built on — spanning LLM reasoning, AI agents, retrieval-augmented generation, knowledge graphs, and translational biology.
13 conference and journal papers in 2026, including ACL, EMNLP, KDD, SIGMETRICS, SoCC, MLSys, JAMIA Open, and npj Digital Medicine, plus three AAAI 2026 workshop presentations.
ReserveGate: Reserve-Aware Evidence Scheduling for Long-Horizon Language Agents
Hill Research · 2026
An evidence scheduler for long-horizon language agents working under shared token and tool-time budgets. Repeated retrieval and verification can exhaust resources early and leave too little capacity to resolve critical constraints later. ReserveGate estimates the value of each evidence action, enforces separate resource limits, and combines adaptive resource pricing with a progressively released reserve that protects late-stage capacity.
- Tight-budget task success up 2.6 to 3.8 percentage points over the strongest matched controller, across three controlled agent environments
- Gains of 4.6 to 6.2 points in the longest episode quartile
- With full-system costs included, 5.5 to 6.6% lower end-to-end token consumption and 5.9 to 9.2% lower P99 tool time
TraceWeaver: Language-Conditioned Action Selection for Web Agents via Coupled Rollouts and Episodic Memory
Hill Research · 2026
Helps web agents choose actions when the information needed to verify the constraints in an instruction only appears on later pages or after later tool calls. Constraint-carrying latent rollouts combine predicted successor states, the active instruction constraints, and evidence retrieved from an immutable episodic-memory snapshot. Coupling efficient latent dynamics with constrained action proposals and critic-guided search keeps task requirements intact through multi-step planning.
- Task success up 2.7 to 4.7 percentage points over the strongest outcome baseline under matched evaluation, across WebShop, Mind2Web, WebArena, and held-out WorkArena
- On WebShop, planning uses 406 tokens per step instead of 4,396, and 768 ms of model-only latency instead of 4,210 ms, compared with text-level LLM rollouts
SocialCritic: Grounding Multi-Agent Dialogue via Advantage-Based Endorsement Suppression
Hill Research · 2026
Addresses hallucination propagation in multi-agent dialogue, where an unsupported claim from one agent is accepted by another and then shapes downstream decisions. A learned receding-horizon controller ranks and repairs candidate action sequences using separately trained utility, norm, and epistemic advantage critics.
- 692 test episodes across LIGHT, DealOrNoDeal, SOTOPIA, and Persuasion-Conflict, over 10 generation seeds
- Goal completion up 6.9 percentage points and actionable accuracy up 7.9 points over SAFE-Dialogue
- Unsupported endorsement down 44.9% and norm violations down 42.7%
- Critics transferred to unseen benchmarks keep actionable-accuracy gains of at least 5.2 points
- At the default operating point, 1.7 external round trips and 748 ms per turn
An audit-ready EHR foundation model for order-time medication-safety auditing across health systems
Hill Research · 2026
An EHR foundation model for auditing high-risk medication orders at the moment they are placed. A Clinical Trajectory Language encodes time, units, provenance, and observability, so the model can represent evidence that is missing, externally sourced, or stale rather than treating absence as normal. Evaluated retrospectively across seven health systems against a reference standard frozen before final scoring.
- 847,000 eligible high-risk medication orders across seven health systems
- Failure to alert, with hard constraints: 3.2 per 1,000 orders, against 9.4 for rules only and 4.8 for the strongest LLM comparator
- False-alert rate: 33.9, against 38.7 for rules only
- Scope limit stated by the authors: an offline process metric that did not measure clinical actions, adverse drug events, or patient harm, and does not establish clinical effectiveness
Tessera: Contract-Driven Cost-SLO Optimization for Multi-tenant GPU Clouds
Hill Research · 2026
A contract-driven control plane for GPU fleets that serve latency-critical inference, throughput-oriented training, and bursty analytics at the same time. Workloads declare verifiable elasticity envelopes and per-stage bottleneck hints, which lets the scheduler minimize dollar cost while holding SLOs instead of relying on coarse resource requests.
- 28 to 45% lower weighted GPU-hour cost than static peak provisioning
- 1.32 to 1.58x aggregate goodput
- 10 to 18% cheaper than the heterogeneity-aware Gavel baseline, with tail constraints preserved
- Up to 2.1x smaller peak P99 latency spikes under mixed-tenant interference
- Implemented on Kubernetes, evaluated on 256 nodes and 512 GPUs
Cadence: Taming Coupled Contention in Disaggregated Agent Clouds
Hill Research · 2026
Addresses coupled contention in multi-tenant LLM agent serving on disaggregated GPU pools and remote KV-cache tiers, where optimizing GPU utilization alone still produces cache hot spots, cross-rack fan-out, and admission overload. Cadence jointly allocates GPU compute, KV-cache residency, and network bandwidth for agent graphs under per-tenant SLO and fairness constraints.
- 1.4 to 1.8x lower P99 end-to-end latency than GPU-only scheduling
- 6.0 to 7.6x fewer SLO violations
- Up to 54% higher cluster goodput at full load
- Evaluated on a 32-GPU, 400 Gbps leaf and spine testbed with eight concurrent tenants
MosaicKV: SLO-Aware KV-Cache Virtualization for Multi-Tenant LLM Serving in the Cloud
Hill Research · 2026
A KV-cache virtualization layer for multi-tenant LLM inference on shared GPU servers. Because KV caches consume tens of gigabytes per replica, providers otherwise over-provision GPU memory or reject long-context requests, and naive offloading causes tail-latency spikes. MosaicKV treats the KV cache as a schedulable, migratable resource across GPU-memory partitioning, HBM versus host-DRAM placement, prefetching, and micro-batching.
- 1.93 to 2.31x higher goodput under SLO
- 38% smaller peak HBM KV footprint
- Interactive-tenant P99 time to first token held within the 400 ms SLO during flash-crowd bursts, against 1,055 ms for Unified Memory
- Evaluated on an 8-GPU H100 server with a 64-tenant mixed workload across Llama-3.1-70B, Qwen2.5-32B, and Mixtral-8x7B
From Tool Traces to Strategy Banks: Mining Failure Signatures for Budgeted Repository-Level Code Repair
Shuang Cao, Rui Li · 2026
ESDA mines the execution traces an agent leaves while repairing a failing repository into structured failure signatures, and builds a bank of reusable debugging strategies stored in typed slots so they can be recombined as new tasks arrive. A cost-aware learning-to-rank stage decides what to try first, so the strategies most likely to succeed are tried within the first few attempts and fewer expensive evaluator calls are needed.
- RepoBench success rate of 84.7%, 38.7 percentage points above the strongest baseline
- 4.0x fewer evaluator calls
- Strategy retriever ranking quality of 53.4 MRR, with a rank correlation of 0.89
- 71 to 74% of the gains retained when transferred to a frozen setting
- Integrity audit for reward hacking: 0.7%
Real-Time Clinical Analytics at Scale: A Platform Built on Large Language Models-Powered Knowledge Graphs
Hill Research · 2026
Describes ClinicalMind, the knowledge graph layer underneath TriClick. Initializes from 300 curated authoritative sources, reducing LLM invocation costs by 70%. Processes 110,000 clinical documents and 60,000 EMRs with 1.5M core concepts and 3M relationships.
- 110,000 clinical documents + 60,000 EMRs processed
- 1.5M core concepts, 3M primary relationships
- Average query delay: 1.7 seconds
- BLEU: 0.85, ROUGE: 0.92
From Trajectories to Graphs: Contract-Checked Editing for Verifier-Guided LLM Reasoning
Dr. Jack Li et al. · 2026
Introduces contract-checked graph editing for LLM reasoning. Represents LLM outputs as typed reasoning graphs and runs a deterministic structural gate before the expensive verifier, filtering structurally invalid candidates immediately.
- Verifier-runnable recombination: 41.2% → 92.8%
- Accuracy: +6.1 on MATH, +9.1 on MATH Level 5
- 42% fewer verifier calls
HyperWorld: Hybrid World Models for Grounded Language Agents
Dr. Jack Li et al. · 2026
A hybrid world model combining SSM-based dynamics with entity-centric episodic memory and critic-guided rollout planning. Enables AI agents to simulate the impact of their decisions before committing to action.
- Outperformed GPT-4+ReAct by 11-14 points on ALFWorld, WebShop, SciWorld
- 53% reduction in constraint violations
EviDex: Provenance-Weighted Evidence-Path Indexing for Fresh and Auditable Retrieval under Continuous Updates
Rui Li, Shuang Cao, Ruihua Liu, Alexandre Duprey · 2026
Replaces periodic RAG refresh with log-structured online compaction over intent-partitioned evidence-path buckets. Every retrieved path carries its provenance for regulatory audit trails. Now available in the ACM Digital Library.
- Evidence-set violation at 15 min: 1.3% (vs 2.4% baseline)
- Cost: $0.68 per 1k queries — 42% cheaper than adaptive TTL
- 10M docs / 16 nodes: 1,856 queries/sec, p99 latency 2.14s
- Clinical correctness: 0.884 on 800-question physician-rated test
Ontology-Guided Long-Term Memory for Conversational RAG
Dr. Jack Li · 2026
Extracts durable user facts into a lightweight ontology memory graph and routes between graph-first and dense-first retrieval with a budget-aware learnable router. Solves the problem of dense retrieval failing in long multi-session conversations.
- Recall@10: 0.70 (vs 0.58 for dense-only)
- 47% reduction in cross-modality disagreement
- 81% cost reduction vs long-context methods
Med-ICE: Multi-Agent Consensus Framework for Trustworthy Medical AI
Zhiyuan Chen et al. · 2026
A multi-agent LLM framework for high-stakes medical tasks. Multiple agents generate diverse reasoning chains, a semantic consensus module aligns reasoning patterns, and iterative refinement continues until convergence — like a panel of specialists debating a diagnosis.
- 5-8% improvement in factual accuracy
- Fewer unsupported claims across agents
- Transparent disagreement surfacing for clinicians
CSLAN: Cross-Species Latent Alignment Network
Dr. Rui Wu et al. · 2026
A transfer learning framework that bridges mouse and human single-cell datasets. Uses species-invariant genomic features to identify human trauma-related immune cells with minimal human samples.
- 96.67% accuracy with only 240 human samples
- Bridges cross-species data scarcity
- Scalable to disease modeling and drug development
Dynamic Consistency Index for Gene Expression Dynamics
Dr. Rui Wu · 2026
A new metric for measuring how gene expression evolves across immune cell types over time. Enables modeling of gene expression dynamics with high accuracy while estimating uncertainty.
- High-accuracy temporal gene expression modeling
- Built-in uncertainty estimation
- Supports robust biological insights
Consensus-Based Framework for Reducing LLM Hallucinations in Clinical AI
Zhiyuan Chen · 2026
Models from different families challenge each other in a consensus-based framework, improving accuracy and advancing toward transparent, audit-ready clinical AI systems.
- Improved accuracy over single-judge models
- Cross-family model consensus validation
- Designed for audit-ready clinical settings
Interested in Our Research?
We're always looking for collaborators in clinical AI, NLP, and biostatistics. Get in touch.