Results

Hill Pepper Inc

11/19/2026 | Press release | Distributed by Public on 11/19/2026 10:00

Three Hill Research Papers at EMNLP 2026 in Budapest

  1. Home
  2. / News
  3. / Research
  4. / Three Hill Research Papers at EMNLP 2026 in Budapest
Research

Three Hill Research Papers at EMNLP 2026 in Budapest

Hill Research | 19 November 2026

Three papers from Hill Research appear at the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026), held October 24 to 29 in Budapest, Hungary. SocialCritic is in the main conference proceedings. ReserveGate and TraceWeaver are in Findings of the Association for Computational Linguistics: EMNLP 2026.

Shuang Cao and Hill Research CTO Dr. Jack (Rui) Li are equal-contribution first authors on all three. CEO Dr. Louise Liu and Chief Product Officer Dr. Alexandre Duprey are co-authors on ReserveGate and TraceWeaver.

About EMNLP 2026

EMNLP is one of the leading conferences in natural language processing. Main conference papers appear in the EMNLP proceedings and Findings papers in the companion Findings of ACL volume; both are published in the ACL Anthology.

What the Three Papers Share

Each one takes on a failure that shows up late in a long agent task: an unsupported claim that one agent accepts from another, a budget spent before the steps that need it, and a constraint whose consequences only appear several pages later.

SocialCritic: Grounding Multi-Agent Dialogue via Advantage-Based Endorsement Suppression

Shuang Cao and Rui Li. EMNLP 2026, main conference.

In multi-agent dialogue, a hallucination becomes an operational problem when another agent accepts it and later decisions inherit the error. The paper names this cross-speaker step unsupported endorsement and introduces SocialCritic, a learned receding-horizon controller that ranks and repairs candidate action sequences using separately trained utility, norm, and epistemic advantage critics.

Results (692 test episodes across LIGHT, DealOrNoDeal, SOTOPIA, and a held-out Persuasion-Conflict set, over 10 generation seeds):

  • Compared with SAFE-Dialogue, goal completion up 6.9 points and actionable accuracy up 7.9 points, while endorsement falls 44.9% and norm violations 42.7%
  • Critics trained with one benchmark left out keep an actionable-accuracy advantage on it, by at least 5.2 points
  • At the default operating point, 1.7 external round trips and 748 ms per turn

ReserveGate: Reserve-Aware Evidence Scheduling for Long-Horizon Language Agents

Shuang Cao, Rui Li, Louise Ruihua Liu, and Alexandre Duprey. Findings of EMNLP 2026.

A long-horizon agent has to decide not only which evidence to use but when to pay for it. Retrieval and verification can settle the uncertainty in front of it while using up the tokens and tool time a later constraint will need. ReserveGate is an episode-level token and latency scheduler: it keeps both resources within their limits, raises the price of evidence when spending runs ahead, and releases a protected reserve as the episode progresses.

Results (three controlled agent environments):

  • Tight-budget task success up 2.6 to 3.8 percentage points over the strongest matched controller
  • Up 4.6 to 6.2 points in the longest episode quartile
  • With full-system costs charged, end-to-end tokens down 5.5 to 6.6% and p99 tool time down 5.9 to 9.2% relative to that controller

TraceWeaver: Language-Conditioned Action Selection for Web Agents via Coupled Rollouts and Episodic Memory

Shuang Cao, Rui Li, Louise Ruihua Liu, and Alexandre Duprey. Findings of EMNLP 2026.

Web agents often have to act before later pages or tool responses reveal whether an instruction's constraints can still be met. TraceWeaver plans with constraint-carrying latent rollouts: each imagined step combines the predicted next state, the active instruction constraints, and evidence from an immutable episodic-memory snapshot, and a critic guides the search.

Results:

  • Task success up 2.7 to 4.7 points over the strongest outcome baseline on WebShop (500 episodes), Mind2Web (1,000 tasks), WebArena (812 tasks), and held-out WorkArena (165 tasks); every paired-bootstrap 95% confidence interval excludes zero
  • 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

Why It Matters

Statistical programming for a clinical trial has the same shape as these tasks: a long chain of steps, requirements set early that only bite late, and checks that each cost time. Keeping an agent grounded and within budget across the whole chain, not only at the step in front of it, is the problem TriClick works on.

Learn More

Hill Pepper Inc published this content on November 19, 2026, and is solely responsible for the information contained herein. Distributed via Public Technologies (PUBT), unedited and unaltered, on November 19, 2026 at 16:00 UTC. If you believe the information included in the content is inaccurate or outdated and requires editing or removal, please contact us at [email protected]