Tool-using language-model agents accumulate histories of their own mistakes, and a system that compacts, repairs or hands off such a history decides what it says about those mistakes while their effects stay in the world. Prior work cleans the history at a fixed world, where showing a corrected step and saying it is done move together. We ask whether the same correct content helps or hurts depending on how the history states it and what the world holds. We freeze an agent mid-task with its exact world state, edit the history or the world, and let a receiver model continue. On AgentCE-Bench, after the agent commits a wrong value, we hold the world fixed and add the correct value by two natural edits: a truthful note beside the preserved error or the agent's own completed action. The note raises recovery, the share of remaining targets the receiver gets right, from 51% under a matched neutral note to 66%. The completed action lowers it below the unedited history, from 50% to 44%, so the value's presence alone does not predict the effect. On AppWorld, after a database write, we vary whether its receipt reports success or failure and whether its effect is present. A success receipt lowers whole-task success from 45% to 28% while the effect is absent and raises it from 38% to 49% once present. The two effects differ by 27 points pooled over two receivers, a difference positive in six of eight receivers at 97.5% and seven at 95%. A history edit's effect thus depends on the world it describes, and a history made to look clean while the world still holds the error can leave an agent worse off than no edit at all.
Figures & tables
Figure 1. The two designs, record above the line and world below. Left, AgentCE-Bench: a world holding a committed error under two records carrying the same correct value, on recovery. Right, AppWorld: one original success receipt over two worlds, on whole-task success. Each mark is an effect pooled over gpt-5.5 and claude-sonnet-5 , with its 95% interval against the comparator it names; blue is a gain and orange a loss, in every figure. In each design the effect’s sign changes with the half varied. A schematic in two panels under a strip showing the procedure: freeze, edit the record, restore, continue, score. Each panel is split by a horizontal line into a record half above and a world half below. The left panel, AgentCE-Bench, has two record boxes above the line and one shared world box below it, in which the wrong value was committed and still stands. The blue box is a truthful note beside the error, marked helps, plus 15.20 points against a neutral note; the orange box is the agent's own completed action, marked hurts, minus 6.43 points against the unedited record. The right panel, AppWorld, has one shared box above the line holding the original success receipt of a completed database write, with the failure receipt named as the comparator. Below the line, with the step's effect absent the receipt is false and hurts, minus 16.92 points; with it present the receipt is true and helps, plus 10.38 points. Each value is drawn as a dot with an interval on a small zero line.
Contrast (look-ahead endpoint)
pp [95%]
T − N, all 319 states
+15.20 [12.46, 18.00]
part at the slot the note names
+11.57 [9.85, 13.43]
part at the other look-ahead slots
+3.63 [1.67, 5.64]
unmentioned slots, renormalised
+5.75 [2.56, 8.93]
F − O, all 319 states
− 6.43 [ − 8.88, − 3.97]
part at the slot the edit names
− 2.30 [ − 3.68, − 0.97]
Table 1. Where the gain and the loss fall on AgentCE-Bench, on the three-slot look-ahead endpoint of §4.1; unadjusted 95% task-cluster bootstrap intervals. Each 319-state total, T − N then F − O, is split between the slot the edit names and the other look-ahead slots; “renormalised” is T − N as a rate over only the slots the note does not name. The last four rows restrict to the 47 states whose committed error lies outside the look-ahead targets. All but the two totals are descriptive.
Figure 2. The sign reversal at a fixed world, on AgentCE-Bench, pooled over gpt-5.5 and claude-sonnet-5 (top) and per receiver. T − N (blue): the truthful note against a matched neutral note. F − O (orange): the same value as the agent’s own completed action, against the unedited record. The grey band spans the two marks; the black mark is their gap, D = (T − N) − (F − O). The T − O and F − N rows compare each edit against the other edit’s comparator; the bottom row is the control N − O. Hollow marks an interval that includes zero. Recovery on the look-ahead targets; 95% task-cluster bootstrap intervals. A dot-and-interval plot in three groups: pooled over both receivers, gpt-5.5 and claude-sonnet-5. Each group shows two effects on the share of remaining targets correct, joined by a grey band: on the left, in orange, F minus O, the same value as the agent's own completed action against the unedited record, which is negative; on the right, in blue, T minus N, the same value as a truthful note against the neutral note, which is positive. Below each band a larger black dot with its interval is D, printed with its value. Below the groups, pooled T minus O is positive and pooled F minus N negative; the bottom control row, the neutral note against the unedited history, is a hollow marker whose interval crosses zero, labelled unresolved.
Figure 3. A success receipt’s effect flips sign with the world on AppWorld. Each point is the success receipt minus the failure receipt on whole-task success, with the database step’s effect independently absent or present; the arrow is their interaction, read at 97.5%. 95% task-family bootstrap intervals on the two effects; 1,040 continuations over 65 task families, pooled over gpt-5.5 and claude-sonnet-5 . Per receiver: Figure 4; cell means: Figure S1. A single panel with two positions on the x axis, step's effect absent and step's effect present. Each shows the effect of a success receipt on task success, success minus failure receipt, as a dot with an interval. Where the effect is absent the dot is negative, minus 16.92 points, labelled hurts in orange; where it is present it is positive, plus 10.38 points, labelled helps in blue. A grey arrow joins the two, and the interaction, plus 27.31 points with its 97.5 percent interval, is printed beside it.
Figure 4. Figure 3’s interaction for each of eight receivers, grouped by provider. Left: unadjusted 95% (thick) and 97.5% (thin) task-family bootstrap intervals; † marks the six of eight excluding zero at 97.5%; seven exclude it at 95%. Black: the reported frame, 65 task families for §5.4’s two receivers and the 63 all six added receivers completed for the rest. Grey: an added receiver’s own 65-family frame (none for grok-4.6 ). Right: each receiver’s two record effects on whole-task success, effect absent (squares) and present (circles), at 95%; orange a loss, blue a gain; both resolve with opposite signs in five receivers. Hollow: the 95% interval includes zero. A forest plot with one row per receiver model, grouped by provider: Anthropic, Google, \mbox{OpenAI} and xAI. Each row shows the record by world interaction in percentage points as a dot with a thick 95 percent interval and a thin 97.5 percent interval, and a dashed vertical line marks zero. Seven of the eight thick intervals lie entirely to the right of zero; the gpt-4o-mini interval crosses zero, and its marker is hollow and labelled unresolved. For some receivers a second, grey interval shows the same receiver on its own sixty-five-family frame. The estimate and interval are printed beside each row, with a dagger where the 97.5 percent interval also excludes zero. A right-hand panel gives each receiver's two record effects, negative with the step's effect absent and mostly positive with it present, and the last column gives the number of task families in the frame.
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Design
Benchmark
Receivers
Four-arm design
AgentCE-Bench
gpt-5.5 , claude-sonnet-5
Targeted control
AgentCE-Bench
gpt-5.5 , claude-sonnet-5
Channel decomposition
AgentCE-Bench
gpt-5.5 , claude-sonnet-5
Record × world factorial
AgentCE-Bench
gpt-5.5 , claude-sonnet-5 , gpt-5.6-sol
Record × world crossing
AppWorld
gpt-5.5 , claude-sonnet-5
Receiver expansion
AppWorld
adds gemini-3.1-pro-preview , grok-4.6 , gpt-5 , gpt-5-mini , gpt-4o-mini , gpt-4o : eight receivers in four families
Appendix
Table 6
Design
Delivered / planned
Blocks
Frame
Four-arm design
638 complete blocks over 319 states, four continuations each
the agent’s own history, unedited; the world carries the error’s consequence
F
history_healed__world_consequence__matched
the history rewritten so the error never appears, against the unchanged wrong world; the world still carries the consequence
T
e29-truthful__self
the history preserved byte-for-byte, plus one trailing truthful retrospective correction
N
e29-neutral__self
the history preserved byte-for-byte, plus one trailing content-neutral note, length-matched to T
Appendix
Table 8
Contrast
Scope
Estimate
95% interval
Resolution
T − N
pooled
+15.20 pp
[+12.46, +18.00]
positive
gpt-5.5
+15.94 pp
[+12.75, +19.17]
positive
claude-sonnet-5
+14.47 pp
[+10.82, +18.18]
positive
F − O
pooled
−6.43 pp
[−8.88, −3.97]
negative
gpt-5.5
−8.10 pp
[−11.44, −4.65]
negative
claude-sonnet-5
−4.75 pp
[−8.05, −1.52]
negative
Appendix
Table 9
Contrast
Estimate
95% interval
Resolution
N − O
+0.44 pp
[−1.36, +2.32]
unresolved
T − O
+15.65 pp
[+13.06, +18.31]
positive
Appendix
Table 10
Cell
Record the agent sees
World it acts in
S_pre
receipt says the mutation succeeded
effect absent (database before the step)
S_post
receipt says the mutation succeeded
effect present (database after the step)
F_pre
receipt says the mutation failed
effect absent (database before the step)
F_post
receipt says the mutation failed
effect present (database after the step)
Appendix
Table 11
Scope
S_pre
S_post
F_pre
F_post
pooled
28.46
48.85
45.38
38.46
gpt-5.5
40.77
69.23
66.92
56.15
claude-sonnet-5
16.15
28.46
23.85
20.77
Appendix
Table 12
Contrast
Scope
Estimate
Interval
Level
Resolution
interaction
pooled
+27.31 pp
[+14.23, +41.54]
97.5%
positive
gpt-5.5
+39.23 pp
[+23.08, +56.15]
95%
positive
claude-sonnet-5
+15.38 pp
[+3.85, +27.69]
95%
positive
record effect, world pre
pooled
−16.92 pp
[−24.62, −9.62]
95%
negative
gpt-5.5
−26.15 pp
[−38.46, −13.85]
95%
negative
claude-sonnet-5
−7.69 pp
[−14.62, −1.54]
95%
negative
Appendix
Table 13
Figure S1. Cell means of the record × world crossing on AppWorld, per receiver: whole-task success under the success and the failure receipt, with the focal step’s effect absent or present (1,040 continuations, 65 task families). The receivers sit at different operating points, so the panels are not a comparison of susceptibility (Appendix C.2 ).
Re-issued the focal request
Frame
S_pre
S_post
F_pre
F_post
gpt-5.5
65 fam.
50.8 [39.2, 62.3]
16.2 [9.2, 23.8]
46.2 [36.2, 56.2]
16.2 [9.2, 23.8]
claude-sonnet-5
65 fam.
54.6 [43.8, 64.6]
4.6 [0.8, 9.2]
90.8 [84.6, 96.2]
53.1 [41.5, 64.6]
gemini-3.1-pro-preview
shared 63
30.2 [19.8, 40.5]
9.5 [4.0, 15.9]
61.1 [50.8, 70.6]
27.0 [17.5, 37.3]
grok-4.6
shared 63
65.9 [54.8, 77.0]
6.3 [2.4, 10.3]
86.5 [78.6, 93.7]
53.2 [42.1, 64.3]
gpt-5
shared 63
65.9 [55.6, 76.2]
31.0 [19.8, 42.1]
58.7 [47.6, 69.0]
26.2 [16.7, 35.7]
gpt-5-mini
shared 63
72.2 [61.9, 81.7]
43.7 [32.5, 54.8]
75.4 [66.7, 83.3]
42.1 [31.7, 52.4]
Appendix
Table 15
Wrote an exact copy of the focal record
S_pre
S_post
F_pre
F_post
gpt-5.5
53.8 [38.6, 68.8]
11.2 [3.8, 20.2]
70.0 [56.4, 82.4]
18.8 [8.6, 30.4]
claude-sonnet-5
55.0 [40.0, 69.5]
3.8 [0.0, 10.0]
91.2 [82.6, 97.8]
51.2 [35.9, 66.7]
Appendix
Table 16
Receiver
Family
Frame
Interaction
95% interval
Verdict
gpt-5.5
OpenAI
65 families
+39.23 pp
[23.08, 56.15]
positive
claude-sonnet-5
Anthropic
65 families
+15.38 pp
[3.85, 27.69]
positive
gemini-3.1-pro-preview
Google
63 shared
+61.90 pp
[46.03, 77.78]
positive
grok-4.6
xAI
63 shared
+43.65 pp
[28.57, 59.52]
positive
gpt-5-mini
OpenAI
63 shared
+23.81 pp
[14.29, 34.13]
positive
gpt-5
OpenAI
63 shared
+16.67 pp
[0.79, 32.54]
positive
Appendix
Table 17
Receiver
Units
Interaction
Record effect, pre
Record effect, post
World effect, success record
World effect, failure record
gemini-3.1-pro-preview
520, 65 fam.
+60.00 [+44.62, +76.15]
−43.85 [−55.38, −32.31]
+16.15 [+7.69, +25.38]
+50.00 [+37.69, +62.31]
−10.00 [−18.46, −1.54]
grok-4.6
504, 63 fam.
+43.65 [+28.57, +59.52]
−28.57 [−38.89, −19.05]
+15.08 [+5.56, +24.60]
+26.98 [+16.67, +38.10]
−16.67 [−25.40, −7.94]
gpt-5-mini
520, 65 fam.
+23.85 [+14.62, +33.85]
−14.62 [−22.31, −7.69]
+9.23 [+3.08, +15.38]
−0.77 [−10.77, +9.23] unresolved
−24.62 [−34.62, −15.38]
gpt-5
520, 65 fam.
+15.38 [+0.77, +30.77]
−10.00 [−20.00, −0.77]
+5.38 [−4.62, +16.15] unresolved
+8.46 [+0.00, +16.92]
−6.92 [−17.69, +3.85] unresolved
gpt-4o
520, 65 fam.
+60.00 [+41.54, +79.23]
−41.54 [−53.08, −30.77]
+18.46 [+8.46, +29.23]
+46.15 [+33.85, +57.69]
−13.85 [−24.62, −3.08]
gpt-4o-mini
520, 65 fam.
+1.54 [−6.92, +10.77] unresolved
+1.54 [−6.15, +9.23] unresolved
+3.08 [−2.31, +9.23] unresolved
−10.00 [−20.77, +0.00] unresolved
−11.54 [−20.77, −2.31]
Appendix
Table 18
Figure S2. The targeted control on AgentCE-Bench: specificity, the gain on slots no arm named and no world repaired when the false claim matches the repaired slot, pooled at 97.5% and split by focal slot at 95% (paired frame, 223 checkpoints, 4,014 units; Appendix E.3 ).
Quantity
Value
Planned
4,320 units in 480 blocks, 240 checkpoints
Analysed
4,068 units in 452 blocks
Never collected
189 units in 21 blocks
Permanently held
63 units in 7 blocks
Appendix
Table 20
Contrast
Scope
Estimate
Interval
Level
Resolution
specificity (primary)
pooled
+4.628 pp
[+2.989, +6.317]
97.5%
positive
gpt-5.5
+5.800 pp
[+3.814, +7.865]
95%
positive
claude-sonnet-5
+3.456 pp
[+1.312, +5.477]
95%
positive
corresponding attenuation
pooled
+3.282 pp
[+0.938, +5.684]
95%
positive
gpt-5.5
+3.039 pp
[+0.336, +5.675]
95%
positive
claude-sonnet-5
+3.525 pp
[+0.028, +7.092]
95%
positive
Appendix
Table 21
Measure
Estimate
Interval
Level
Resolution
steps used
−0.297 steps
[−0.502, −0.108]
97.5%
negative
hit horizon
−0.026 proportion
[−0.046, −0.008]
97.5%
negative
touches the common cell
+1.667 steps
[+1.223, +2.127]
97.5%
positive
Appendix
Table 22
Contrast
Scope
Estimate
Interval
Resolution
flag − neutral ( primary , Bonferroni two-comparison)
pooled
+1.62 pp
[−1.04, +4.36] at 97.5%
unresolved
gpt-5.5
+3.50 pp
[+0.47, +6.53]
positive
claude-sonnet-5
−0.26 pp
[−3.50, +2.82]
unresolved
full − answer
pooled
+4.75 pp
[+2.72, +6.79]
positive
gpt-5.5
−0.21 pp
[−2.40, +1.99]
unresolved
claude-sonnet-5
+9.72 pp
[+6.58, +12.96]
positive
Appendix
Table 23
Contrast
pooled
gpt-5.5
claude-sonnet-5
gpt-5.6-sol
rewrite − original, world unrepaired
−5.54 [−8.08, −3.15]
−6.02 [−9.51, −2.55]
−5.06 [−8.71, −1.48]
−11.58 [−15.55, −7.65]
rewrite − original, world repaired
−1.64 [−4.25, +0.97] unres.
−5.47 [−9.39, −1.76]
+2.20 [−1.44, +5.87] unres.
−7.46 [−11.00, −3.84]
world repair under original history
+3.08 [+0.77, +5.39]
+2.41 [−0.61, +5.47] unres.
+3.74 [+0.22, +7.32]
+2.45 [−0.90, +5.84] unres.
world repair under rewritten history
+6.98 [+4.31, +9.70]
+2.96 [−0.70, +6.65] unres.
+11.00 [+6.87, +15.13]
+6.56 [+2.64, +10.58]
Appendix
Table 24
Collection
Units re-derived
Agreements
Label split (success / failure)
Record × world crossing
1,040
1,040
419 / 621
Expansion: gemini-3.1-pro-preview
520
520
326 / 194
Expansion: grok-4.6
512
512
413 / 99
Expansion: gpt-5
520
520
228 / 292
Expansion: gpt-5-mini
520
520
225 / 295
Expansion: gpt-4o
520
520
210 / 310
Appendix
Table 25
Dimension
Count
a: the record and the world treated as two separable objects
91 of 360
b: what the agent is shown varied while the underlying world state is held fixed
53 of 360
a ∧ b: both
23 of 360
c: the matched append control (five attributes)
0 of 360
d: what a record reveals held apart from what it asserts
3 of 360, one only partially ( Gumaan, 2026 ; Panickssery et al., 2024 ; Hu and Ramachandran, 2026 )
a ∧ b ∧ (c ∨ d): the conjunction the positioning argument rests on
Large language model (LLM) agent reliability concerns dependable, safe, and verifiable behavior across tools, multi-step tasks, and changing environments.
Does repairing an episode make its experience a better memory for the next task? We transfer the same failed source before and after accepted repair to a fixed target, alongside independent execution. Our 3,300 runs cover 100 ThinkingBox pairs and the same 100 APEX pairs with and without source-state inheritance, under eleven conditions. ThinkingBox's Full/Skill/Hybrid correction gains are 44/29/32 percentage points, with corrected performance 25/22/18 points above independence; inference weakens at the task-family level. Yet 12 of Full's 15-point larger correction gap over Skill come from worse uncorrected performance, not better corrected memory. Moreover, 22 of Full's 46 upward transitions restore observed baseline success. Neither APEX regime establishes comparable aggregate correction benefits. Action evidence connects workflow gains with reusable obligations and convention conflicts with source-local choices. Text APEX's accepted execution reaches 52% versus its summary's 40%, without robust global/group-level superiority or an estab- lished advantage over independence. Smaller handoffs reduce input but increase calls. The value of repairing experience is therefore distinct from the value of reusing it: memory updates require both a previous-version reference and a fresh-start reference.
Agent revisions expose a fundamental correctness--efficiency trade-off during concurrent execution. Discarding ongoing work preserves latest-version correctness but wastes progress that may remain valid, whereas reusing prior work preserves efficiency but risks propagating stale state into outputs and tool effects. Existing recovery strategies resolve this trade-off in an imbalanced way with coarse-grained policies: they either favor efficiency by allowing potentially stale work to continue, or favor correctness by restarting the workflow or recomputing a linear suffix from the earliest conflict, thereby discarding unaffected progress. We present Revise, a validity-guided runtime for fine-grained recovery in structured agent workflows. When a revision arrives, Revise first intersects its delta with recorded data and control dependencies and propagates the resulting impact through the partially executed DAG to identify affected work. It then stops invalid work, preserves validity-established progress beyond the earliest conflict, and recomputes only the affected region. Incomplete provenance conservatively expands recovery, while reused results are revalidated before commit. Analysis of real coding-agent traces show online recovery opportunities: 118 sessions retain observable work before a queued later message is delivered; across 167 overlapping assistant responses, enqueue-to-completion overlap reaches 56.55~s at p95. Across 300 challenging revision/commit executions, Revise matches a latest-version oracle with no stale outputs or effects. On unmodified LangGraph and LLMCompiler applications using Qwen3-14B, it reduces model calls by 40.6--56.0% relative to full restart and by 31.3--43.6% relative to suffix recomputation. Under serving pressure, it further reduces revision-to-correct-completion tokens by 13.26% and improves SLO goodput by 3.07--5.43%.
Ruoling Qi, Xuaner Wu, Penghang Liu +2
Institute of Artificial Intelligence, China Telecom (TeleAI) · Shanghai Jiao Tong University
Recent advances in large language models (LLMs) have enabled agents to tackle long-horizon tasks across diverse environments. To further improve agent performance, existing language world models typically predict environment observations, yet reconstructing high-entropy, execution-dependent tool responses offers limited value when real feedback is available. Meanwhile, agents suffer from task-state contamination, where unsupported assumptions and outdated plans persist in history and distort subsequent decisions. We propose the Agent-Editing World Model (AEWM), which models how reasoning and actions shape future task progress rather than simulating tool responses. AEWM combines Action Judge to distinguish Critical, Exploratory, and Noisy decisions with State Revision to edit noisy reasoning--action continuations from the same observed history. EditAct integrates these capabilities with real execution, directly changing the state underlying subsequent decisions rather than merely providing critiques. We train AEWM across Search, Terminal, and Software Engineering through mid-training and supervised fine-tuning. AEWM achieves 70.5% macro-F1 on our Action Judge benchmark, exceeding the strongest frontier baseline by 10.6 points. Across six benchmarks and three agent backbones, EditAct improves average scores by 3.2--6.7 points over the strongest baseline. Furthermore, rejection sampling fine-tuning on verified EditAct trajectories, termed AEWM-RFT, improves over Self-RFT by 2.2--2.6 points across three domains without online AEWM guidance.
Shuang Sun, Guoxin Chen, Fanzhe Meng +6
Gaoling School of Artificial Intelligence, Renmin University of China