acceptodds
Under review as a conference paper at ICLR 2027

Transfiver: Human-AI Co-Inference through a Shared Editable State

Abstract

For people and AI systems to reason together, people need to inspect and correct the information the system uses in subsequent computation. We introduce the TRANSparent Framework for Interactive, Verifiable, Editable Representation (Transfiver), an architecture for human-AI co-inference through a shared editable state. Its central idea is that interaction-specific information is maintained in a single persistent state that both the model and the human update. Transfiver distinguishes two modes of state evolution. An implicit stream update requires the model to interpret ongoing interaction and determine whether new information revises an existing state item or creates a new one. In an explicit directed edit, a human inspects a record and changes it. Both act on the same underlying state, so a human correction changes the state that subsequent computation reads, rather than adding another instruction or separate record. We evaluate two implementations that test complementary aspects of this design. Record-store experiments show two results. A retracted record stays out of the prompt when the system rebuilds its input for the next question. Its text stays readable for inspection and restoration. A learned model reads editable records while it generates each answer, so retracting the record that holds an answer removes that answer. These results support shared editable state as a basis for human-AI collaboration.

Primary area:
foundation or frontier models, including LLMs
Keywords:
human, human ai, cowork, one state, editable, shared

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_o8ZXOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper discusses a very important topic - human-AI collaboration, in particular, co-inference. It proposes using a shared editable state, something that is already common in systems like OpenClaw or Claude that maintain .md files that can be edited by humans and AI. What this paper adds is not clear - a more direct, interactive way to edit? So the general approach might be known, there is certainly a lot more to be done in that direction and I could easily see many papers in that direction. The current submission fails to provide good motivating examples. The intro talks about problems how agents claimed to accomplish a task but did not. But it is not obvious how your approach really solves this - Does a user have to overwrite the AI? The evaluation essentially confirms that the approach work, but it is not surprising that editing the state has an impact - it is already the reality for many systems. So overall I fail to see the big insight from these evaluations. So while I like the idea and the abstract problem, the evaluation showing its relevance for a meaningful benchmark comparing it against other techniques is absent.

Critical strengths

see above

Critical weaknesses

see above

Questions for authors

see above

Reviewer_FdXpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes one persistent state that a model and a person both update: implicit stream updates (the model decides whether new information revises or creates a record) and explicit directed edits (a person retracts, restores or rewrites a record). It states three requirements: history sufficiency, state restoration, and state interventions. Two prototypes are evaluated: (i) a record store with frozen readers on LongMemEval, and (ii) a learned slot memory with a copy decoder over a frozen Qwen2.5-7B, evaluated on FactConsolidation, held-out MQuAKE-CF and HaluMem.

Reported results: store-level edits persist across questions while prompt-only edits do not (Table 3); retracting the slot that holds an answer removes it in the learned model (203/209 and 489/494; controls lose at most one per seed); a rewritten value is emitted exactly in 95/330 cases; save–reload reproduces all tested answers.

Critical strengths
  1. A reusable target/control/undo intervention protocol with save–reload audits. Retracting the supporting record, retracting four matched controls, and undoing the edit rules out the trivial explanation that any edit changes behavior. This is the paper's most durable contribution, since it can audit other editable-memory systems. It matters for the overall assessment because it is what makes the paper worth reconsidering after revision. It is most informative on models whose read is soft rather than hard-masked, as here.

  2. Candid, internally consistent reporting, including negative results. Rewrite at 95/330, stale values after retracting the current record (Table 17), both-value answers in HaluMem (Table 5), and new-slot-only writes (B.4) are all reported. This matters because it lets a reader locate the claim–evidence gap precisely, and it is the reason I rate soundness as addressable rather than fundamentally flawed.

Critical weaknesses
  1. The headline claims are not evaluated. The title and abstract describe human–AI co-inference through one state updated in two modes. The evidence is oracle editing: benchmark annotations supply both the record address and the replacement value, and no person inspects or corrects anything. The stream mode never revises an existing record (the incoming record goes to a new slot in 198/198 cases, B.4; the conclusion defers this). Adding stream updates to supplied edits moves Table 11 from 67 to 68/78. The two prototypes are disjoint, and the record store violates the paper's own principle that memory, reasoning and output are not divided. Why it is consequential: if a store with a frozen reader counts as Transfiver, the architecture reduces to a store with a soft-delete flag, which is standard. If the learned reader is required, the evidence covers only the learned model, whose scope is narrow (Weakness 2). Either way the paper's central contribution is not established, and the abstract's closing sentence (these results support shared editable state as a basis for human–AI collaboration) does not follow.

  2. The positive results hold largely by construction, are narrow, and are not isolated from simpler baselines. A prompt-only change cannot survive a prompt rebuild. In the learned model, Eq. 3 zeroes the read weight of a retracted slot, about 99% of tokens come from the copy branch, and the old and new values are counterfactual facts the backbone does not know. So retraction removing the answer, restoration recovering it, and controls leaving it alone are implementation checks, not tests of the hypothesis that answers are computed from editable state. The learned model works only on short templated facts and is evaluated near in-distribution (trained on MQuAKE pairs, tested on held-out MQuAKE-CF and FactConsolidation). On conversational turns, its top-1 selection is 7/44 against BGE's 9/44, and exact match is 2/44 even with the correct record supplied. On HaluMem its own choice yields neither value in 94.8% of answers. Baselines do not isolate the state design. A retriever given the same correction ties or beats the edited store (Table 13: 59 vs 59 on Qwen-7B, 60 vs 67 on Mistral). The edited record is force-included in the read (Appendix A.6: removing the rule drops 67 to 57/78). The state-versus-prompt comparison (Table 6) pits a task-trained copier against zero-shot prompting, and no prompt arm removes the withdrawn line, so the claim that a prompt cannot say which record was withdrawn is not shown. Why it is consequential: the results the paper reports cannot distinguish Transfiver from a store with a flag or a retriever with the same correction, and the one setting that would (the learned model) does not work on the conversational records that motivate the paper.

Questions for authors
  1. Which result in the paper distinguishes the learned model from the record store plus a soft-delete flag, or from a retriever given the same correction? If none, what is the claimed contribution beyond the evaluation protocol?
  2. In Table 6, what do the prompt arms score when the withdrawn line is removed from the prompt, which is the prompt-side equivalent of retraction?
  3. What is rewrite success when the record is named, and under a containment criterion rather than exact match? This separates writing failures from selection failures.
  4. If a retracted fact is mentioned again later in the stream, what does the model answer? Does retraction hold?
  5. What experiment could falsify history sufficiency, such as different histories reaching the same state (permuted slots per Eq. 10, edit-then-undo against no edit)? Does the save–reload audit do anything beyond checking serialization determinism?
Reviewer_xbJWOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Transfiver, where a human and an AI system share the same stored state and can both edit it. This state should include all information from their interactions that affects future behavior, and edits should change how the system works later. The authors test two versions: a record store that allows records to be removed and brought back, and a learned slot memory with a copy decoder linked to a frozen language model. Tests on LongMemEval, FactConsolidation, held-out MQuAKE-CF, and HaluMem show that record edits can remain effective over time. However, the system still has problems choosing which records to edit, writing new values, and telling old and current values apart within the same record.

Critical strengths

1> The experiments compare retracting the supporting record with retracting unrelated records and other highly weighted records. Removing the supporting record makes 203/209 previously correct answers on FactConsolidation and 489/494 on held-out MQuAKE-CF incorrect, while the control interventions have much smaller effects. Together with restoration and fresh-process reload checks, these tests provide useful evidence that the editable state affects the answers.

2>The paper reports correction-matched retrieval baselines, poor autonomous selection on conversational records, limited rewriting success, and answers containing both current and old values. These results show that successfully storing an edit does not always lead to a correct answer. They also make it easier to judge what the experiments support.

Critical weaknesses

1> In the record-store experiments, an edit to persistent storage remains after input reconstruction, while a one-turn prompt change does not. This shows that the implementation works as intended, but does not by itself show a clear benefit of the architecture. More importantly, Appendix A.5 already gives a correction-matched comparison: an ordinary retriever matches the edited store on Qwen2.5-7B, with 59/78, and performs better on Mistral-7B, with 67/78 versus 60/78. The paper needs to explain what its shared-state requirement adds beyond persistent editing, availability filtering, and serialization.

2> The abstract highlights understanding incoming information to update an existing item or create a new one. However, Appendix B.4 reports that all 198 tested incoming updates are written to different slots, and the conclusion leaves reliable updates to existing records for future work. Also, benchmark annotations provide the edit addresses, and the strongest learned-model results also provide the answer record. Without a supplied address, answer containment drops from .917 to .528 on FactConsolidation and from .884 to .572 on MQuAKE-CF. On the 29-question HaluMem subset, even when the correct record is supplied, .609 of answers contain both old and current values, while only .149 contain the current value alone. These results give stronger support for controlled access to records than for correctly understanding and using corrections.

Questions for authors

1> Could the authors give a clear example where a versioned, editable retrieval store with availability flags and complete serialization fails to meet a Transfiver requirement? How does the proposed implementation solve this problem?

2>Can the tested learned model reliably update an existing record when new information replaces the old information?

3>Could the authors compare the learned state with a simpler retrieval-and-copy implementation, using the same records and supplied addresses? This would help separate its contribution from oracle addressing and copying, and show whether the gains come from the proposed representation or from selecting and copying the given record.

4>Table 20 reports 95/330 exact replacements, another 79 outputs containing the replacement, and 76/90 successful rewrites among initially exact-correct answers. Could the authors report the results in four groups: replacement-only, original-only, both, and neither, with and without a supplied record address?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.